Research ArticleHuman Activity Recognition in AALEnvironments Using Random Projections
Robertas DamaševiIius1 Mindaugas Vasiljevas1 Justas ŠalkeviIius1 and Marcin Wofniak2
1Department of Software Engineering Kaunas University of Technology LT-51368 Kaunas Lithuania2Institute of Mathematics Faculty of Applied Mathematics Silesian University of Technology 44-100 Gliwice Poland
Correspondence should be addressed to Robertas Damasevicius robertasdamaseviciusktult
Received 8 February 2016 Revised 29 April 2016 Accepted 19 May 2016
Academic Editor Ezequiel Lopez-Rubio
Copyright copy 2016 Robertas Damasevicius et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited
Automatic human activity recognition systems aim to capture the state of the user and its environment by exploiting heterogeneoussensors attached to the subjectrsquos body and permit continuous monitoring of numerous physiological signals reflecting the stateof human actions Successful identification of human activities can be immensely useful in healthcare applications for AmbientAssisted Living (AAL) for automatic and intelligent activity monitoring systems developed for elderly and disabled peopleIn this paper we propose the method for activity recognition and subject identification based on random projections fromhigh-dimensional feature space to low-dimensional projection space where the classes are separated using the Jaccard distancebetween probability density functions of projected data Two HAR domain tasks are considered activity identification and subjectidentification The experimental results using the proposed method with Human Activity Dataset (HAD) data are presented
1 Introduction
The societies in the developed countries are rapidly agingIn 2006 almost 500 million people worldwide were 65 yearsof age or older By 2030 that total number of aged people isprojected to increase to 1 billion The most rapid increase ofaging population occurs in the developing countries whichwill see a jump of 140 by 2030 [1] Moreover the worldrsquospopulation is expected to reach 93 billion by 2050 [2] andpeople who are above 60 years old will make up 28 ofthe population Dealing with this situation will require hugefinancial resources to support the ever-increasing living costwhere human life expectancy is expected to reach 81 years by2100
As older people may have disorders of body functions orsuffer from age-related diseases the need for smart healthassistance systems increases each year A common methodof monitoring geriatric patients is a physical observationwhich is costly requires a lot of human staff and is increas-ingly infeasible in view of massive population aging inthe following years Many Ambient Assisted Living (AAL)
applications such as care-providing robots video surveillancesystems and assistive human-computer interaction technolo-gies require human activity recognition While the primaryusers of the AAL systems are of course the senior (elderly)people the concept also applies to mentally and physicallyimpaired people as well as people suffering from diabetesand obesity who may need assistance at home and people ofany age interested in personal fitness monitoring As a resultthe sensor-based real-time monitoring system to supportindependent living at home has been a subject of manyrecent research studies in human activity recognition (HAR)domain [3ndash10]
Activity recognition can be defined as the process ofhow to interpret sensor data to classify a set of humanactivities [11] HAR is a rapidly growing area of research thatcan provide valuable information on health wellbeing andfitness of monitored persons outside a hospital setting Dailyactivity recognition using wearable technology plays a centralrole in the field of pervasive healthcare [12] HAR has gainedincreased attention in the last decade due to the arrival ofaffordable and minimally invasive mobile sensing platforms
Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2016 Article ID 4073584 17 pageshttpdxdoiorg10115520164073584
2 Computational and Mathematical Methods in Medicine
such as smartphones Smartphones are innovative platformsfor HAR because of the availability of different wirelessinterfaces unobtrusiveness ease of use high computingpower and storage and the availability of sensors suchas accelerometer compass and gyroscope which meet thetechnical and practical hardware requirements for HAR tasks[13ndash15] Moreover technological development possibilities ofother applications are still arising including virtual realitysystemsTherefore these machines present a great possibilityfor the development of innovative technology dedicated forthe AAL systems
One of the keymotivating factors for usingmobile phone-based human activity recognition in the AAL systems is therelationship and correlation between the level of physicalactivity and the level of wellbeing of a person Recording andanalysing precise information on the personrsquos activities arebeneficial to keeping the progress and status of the disease (ormental condition) and can potentially improve the treatmentof personrsquos conditions and diseases as well as decreasingthe cost of care Recognizing indoor and outdoor activitiessuch as walking running or cycling can be useful to providefeedback to the caregiver about the patientrsquos behaviourWhenfollowing the daily habits and routines of users one caneasily identify deviations from routines which can assistthe doctors in diagnosing conditions that would not beobserved during routine medical examination Another keyenabler of the HAR technology is the possibility of providingindependent living for the elderly as well as for patientswith dementia and other mental pathologies which could bemonitored to prevent undesirable consequences of abnormalactivities Furthermore by using persuasive techniques andgamification HAR systems can be designed to interact withusers to change their behaviour and lifestyles towards moreactive and healthier ones [16]
Recently various intelligent systems based on mobiletechnologies have been constructedHARusing smartphonesor other types of portable or wearable sensor platformshas been used for assessing movement quality after stroke[17] such as upper extremity motion [18] for assessinggait characteristics of human locomotion for rehabilitationand diagnosis of medical conditions [19] for postoperativemobilization [20] for detecting Parkinsonrsquos disease backpain and hemiparesis [21] for cardiac rehabilitation [22] forphysical therapy for example if a user is correctly doing theexercises recommended by a physician [23 24] for detectingabnormal activities arising due to memory loss for dementiacare [25 26] for dealing with Alzheimerrsquos [27] and neu-rodegenerative diseases such as epilepsy [28] for assessmentof physical activity for children and adolescents sufferingfrom hyperlipidaemia hypertension cardiovascular diseaseand type 2 diabetes [29] for detecting falls [30 31] foraddressing physical inactivity when dealing with obesity [32]for analysing sleeping patterns [33] for estimating energyexpenditures of a person to assess hisher healthy dailylifestyle [34] and for recognizing the userrsquos intent in thedomain of rehabilitation engineering such as smart walkingsupport systems to assist motor-impaired persons and theelderly [35]
In this paper we propose a new method for offlinerecognition of daily human activities based on feature dimen-sionality reduction using random projections [36] to lowdimensionality feature space and using the Jaccard distancebetween kernel density probabilities as a decision function forclassification of human activities
The structure of the remaining parts of the paper isas follows Section 2 presents the overview of related workin the smartphone-based HAR domain with a particularemphasis on the features extracted from the sensor dataSection 3 describes the proposedmethod Section 4 evaluatesand discusses the results Finally Section 5 presents theconclusions and discusses future work
2 Overview of HAR Featuresand Related Work
All tasks of the HAR domain require correct identification ofhuman activities from sensor data which in turn requiresthat features derived from sensor data must be properlycategorized and described Next we present an overview offeatures used in the HAR domain
21 Features While numerous features can be extracted fromphysical activity signals increasing the number of featuresdoes not necessarily increase classification accuracy since thefeatures may be redundant or may not be class-specific
(i) Time domain features (such as mean median vari-ance standard deviation minimum maximum androot mean square applied to the amplitude andtime dimensions of a signal) are typically used inmany practical HAR systems because of being lesscomputationally intensive thus they can be easilyextracted in real time
(ii) Frequency-domain features require higher computa-tional cost to distinguish between different humanactivitiesThus they may not be suitable for real-timeAAL applications
(iii) Physical features are derived from a fundamentalunderstanding of how a certain human movementwould produce a specific sensor signal Physicalfeatures are usually extracted from multiple sensoraxes based on the physical parameters of humanmovements
Based on the extensive analysis of the literature andfeatures used by other authors (esp by Capela et al [17]Mathie et al [37] and Zhang and Sawchuk [38]) we haveextracted 99 features of data which are detailed in Table 1
22 Feature Selection Feature selection is the process ofselecting a subset of relevant features for use in constructionof the classification model Successful selection of featuresallows for simplification of models to make them easier tointerpret to decrease model training times and to betterunderstand difference between classes Using feature selec-tion allows removing redundant or irrelevant features with-out having an adverse effect on the classification accuracy
Computational and Mathematical Methods in Medicine 3
Table 1 Catalogue of features
Feature number Description Equation (notation)4ndash6 Acceleration (119909- 119910- and 119911-axes) 119886
119909 119886119910 119886119911
7ndash9 Gyroscope (119909- 119910- and 119911-axes) 119892119909 119892119910 119892119911
10ndash15 Moving variance of 100 samples ofacceleration and gyroscope data
var = 1
119873 (119873 minus 1)
(119873
119873
sum
119894=1
1199092
119894minus (
119873
sum
119894=1
119909119894)
2
) here 119909 = 119886119909 119886119910 119886119911 119892119909 119892119910 119892119911
16-17 Movement intensity of acceleration andgyroscope data
MI119886= radic119886
2
119909+ 1198862
119910+ 1198862
119911
MI119892= radic119892
2
119909+ 1198922
119910+ 1198922
119911
18 Movement intensity of differencebetween acceleration and gyroscope data MI
119892119886= radic(119892
119909minus 119886119909)2
+ (119892119910minus 119886119910)
2
+ (119892119910minus 119886119910)
2
19ndash21 Moving variance of 100 samples ofmovement intensity data
var = 1
119873 (119873 minus 1)
(119873
119873
sum
119894=1
1199092
119894minus (
119873
sum
119894=1
119909119894)
2
) here 119909 = MI119886MI119892MI119892119886
22ndash24 Polar coordinates of acceleration data120593119886= arctan (119886
119910 119886119909)
119903119886= radic119886
2
119909+ 1198862
119910
119911119886= 119886119911
25ndash27 Polar coordinates of gyroscope data120593119892= arctan (119892
119910 119892119909)
119903119892= radic119892
2
119909+ 1198922
119910
119911119892= 119892119911
28ndash30 Polar coordinates of difference betweenacceleration and gyroscope data
82 Simple moving average of cumulativesums of acceleration data
SMAMCS119886=
1
119873
119873
sum
119894=1
MCS119886119894
83 Simple moving average of cumulativesums of gyroscope data
SMAMCS119892=
1
119873
119873
sum
119894=1
MCS119892119894
Computational and Mathematical Methods in Medicine 5
Table 1 Continued
Feature number Description Equation (notation)
84Simple moving average of cumulativesums of difference between accelerometerand gyroscope data
SMAMCS119886119892=
1
119873
119873
sum
119894=1
(MCS119886119894minusMCS
119892119894)
85ndash90 Moving 2nd-order cumulative sum ofacceleration and gyroscope data
MCS1015840 =119873
sum
119895=1
119895
sum
119894=1
MCS119894
91ndash93Moving 2nd-order cumulative sum ofdifferences between cumulative sums ofacceleration and gyroscope data
MCS1015840119886119892=
119873
sum
119895=1
119895
sum
119894=1
(MCS119886119894minusMCS119892
119894)
94ndash96 Polar coordinates of moving cumulativesum of acceleration data
1205931015840
119886= arctan (MCS119886
119910MCS119886
119909)
1199031015840
119886= radic(MCS119886
119909)2
+ (MCS119886119910)
2
1199111015840
119886= MCS119886
119911
97ndash99 Polar coordinates of moving cumulativesum of gyroscope data
1205931015840
119892= arctan (MCS119892
119910MCS119892
119909)
1199031015840
119892= radic(MCS119892
119909)2
+ (MCS119892119910)
2
1199111015840
119892= MCS119892
119911
100ndash102Polar coordinates of moving cumulativesum of differences between accelerationand gyroscope data
1205931015840
119886119892= arctan (MCS119886
119910minusMCS119892
119910MCS119886
119909minusMCS119892
119909)
1199031015840
119886119892= radic(MCS119886
119909minusMCS119892
119909)2
+ (MCS119886119910minusMCS119892
119910)
2
1199111015840
119886119892= MCS119886
119911minusMCS119892
119911
There are four basic steps in a typical feature selectionmethod[58] generation of candidate feature subset an evaluationfunction for feature candidate subset a generation stoppingcriterion and a validation procedure
Further we analyse several feature selection methodsused in the HAR domain
ReliefF [59] is a commonly used filter method that ranksfeatures by weighting them based on their relevance Featurerelevance is based on how well data instances are separatedFor each data instance the algorithm finds the nearest datapoint from the same class (hit) and nearest data points fromdifferent classes (misses)
Matlabrsquos Rankfeatures ranks features by a given classseparability criterion Class separability measures include theabsolute value of a statistic of a two-sample 119905-test Kullback-Leibler distance minimum attainable classification errorarea between the empirical ReceiverOperatingCharacteristic(ROC) curve and the randomclassifier slope and the absolutevalue of the statistic of a two-sample unpairedWilcoxon testMeasures are based on distributional characteristics of classes(eg mean variance) for a feature
Principal component analysis (PCA) is the simplestmethod to reduce data dimensionality This reduced dimen-sional data can be used directly as features for classificationGiven a set of 119873 features a PCA analysis will produce newdata variables (PCA components) as linear combinationsof the features with the highest variance in the subspaceorthogonal to the preceding PCA component As variabilityof the data can be captured by a relatively small number ofPCs PCA can achieve high level of dimensionality reductionSeveral extensions of the PCA method are known such askernel PCA sparse PCA and multilinear PCA
Correlation-based Feature Selection (CFS) [60] is a filteralgorithm that ranks subsets of features by a correlation-based heuristic evaluation function A feature is consideredto be a good one if it is relevant to the target concept but isnot redundant to any of the other relevant features Goodnessof measure is expressed by a correlation between featuresand CFS chooses the subset of features which has the highestmeasure The chosen subset holds the property that featuresinside this subset have high correlation with the class and areunrelated to each other
Table 2 summarizes the feature selectiondimensionalityreduction methods in HAR
A comprehensive review of feature selection algorithmsin general as well as in the HAR domain can be found in [5861ndash63]
23 Summary Related work in the HAR domain is summa-rized in Table 3 For each paper the activities analysed typesof sensor data used features extracted classification methodapplied and accuracy achieved (as given by the referencedpapers) are given
3 Method
31 General Scheme The typical steps for activity recognitionare preprocessing segmentation feature extraction dimen-sionality reduction (feature selection) and classification [24]Themain steps of activity recognition include (a) preprocess-ing of sensor data (eg denoising) (b) feature extraction(c) dimension reduction and (d) classification The prepro-cessing step includes noise removal and representation ofraw data The feature extraction step is used to reduce large
6 Computational and Mathematical Methods in Medicine
Table 2 Summary of feature selectiondimensionality reduction methods in HAR
Method Advantages Disadvantages Complexity
PCA
High dimensionalityreduction reduction of noiselack of redundancy of data dueto orthogonality ofcomponents
The covariance matrix is difficult to beevaluated accurately even the simplestinvariance could not be captured by thePCA unless the training data explicitlyprovides for it
119874(1199012
119899 + 1199013
) where 119899 are datapoints each represented with119901 features
ReliefF Low computationalcomplexity
Unstable due to random selection ofinstances 119874(119901 sdot 119899 sdot log 119899)
RankfeaturesFeatures highly correlatedwith already selected featuresare less likely to be included
It assumes that data classes are normallydistributed
It depends upon classseparability criterion
CFS It evaluates a subset of featuresrather than individual features
It fails to select locally predictive featureswhen they are overshadowed by strongglobally predictive features
119874(119899(
1199012
minus 119901
2
))
input sensor data to a smaller set of features (feature vector)which preserves information contained in the original dataThe dimensionality reduction step can be applied to removethe irrelevant (or less relevant) features and reduce thecomputational complexity and increase the performance ofthe activity recognition processThe classification step is usedto map the feature set to a set of activities
In this paper we do not focus on data preprocessing andfeature extraction but rather on dimensionality reduction andclassification steps since these two are crucial for furtherefficiency of AAL systems The proposed method for humanactivity recognition is based on feature dimensionality reduc-tion using random projections [36] and classification usingkernel density function estimate as a decision function(Figure 1)
32 Description of the Method During random projec-tion the original 119889-dimensional data is projected to a 119896-dimensional (119896 ≪ 119889) subspace using a random 119896 times 119889 matrix119877 The projection of the data onto a lower 119896-dimensionalsubspace is 119883119877119875
119896times119873= 119877119896times119889
119883119889times119873
where 119883119889times119873
is the originalset of 119873 119889-dimensional observations In the derived pro-jection the distances between the points are approximatelypreserved if points in a vector space are projected onto arandomly selected subspace of suitably high dimension (seethe Johnson-Lindenstrauss lemma [64]) The randommatrix119877 is selected as proposed by Achlioptas [36] as follows
119903119894119895=
+1 probability 16
0 probability 23
minus1 probability 16
(1)
Given the low dimensionality of the target space we cantreat the projection of low-dimensional observations ontoeach dimension as a set of random variables for which theprobability density function (PDF) can be estimated usingkernel density estimation (KDE) (or Parzenwindow)method[65]
If 1199091 1199092 119909
119873is a sample of a random variable then
the kernel density approximation of its probability densityfunction is
119891ℎ(119909) =
1
119873ℎ
119870(
119909 minus 119909119894
ℎ
) (2)
where 119870 is some kernel and ℎ is the bandwidth (smoothingparameter) 119870 is taken to be a standard Gaussian functionwith mean zero and variance 1 of the examined data features
119870 (119909) =
1
radic2120587
119890minus(12)119909
2
(3)
For a two-dimensional case the bivariate probabilitydensity function is calculated as a product of univariateprobability functions as follows
119891 (119909 119910) =
119891 (119909) sdot
119891 (119910) (4)
Here 119909 and 119910 are data in each dimension respectivelyHowever each random projection produces a different
mapping of the original data points which reveals only a partof the data manifold in higher-dimensional space In caseof the binary classification problem we are interested in amapping that separates data points belonging to two differentclasses best
As a criterion for estimating the mapping we use theJaccard distance metric between two probability density esti-mates of data points representing each class The advantageof the Jaccard distance metric as compared to other metricsof distance such as Kullback-Leibler (KL) divergence andHellinger distance is its adaptability to multidimensionalspaces where compared points show relations to differentsubsets Therefore it is well adapted to the developed modelof human activity features where according to description inthe previous section we have divided them into some sets ofactions Furthermore the computational complexity of theHellinger distance is very high while KL divergence mightbe unbounded
The Jaccard distance which measures dissimilaritybetween sample sets is obtained by subtracting the Jaccardcoefficient from 1 or equivalently by dividing the difference
Computational and Mathematical Methods in Medicine 7
Table 3 Summary of related works in the HAR domain
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Atallah et al[39]
Lying down preparingfood eating anddrinking socialisingreading dressingwalking treadmillwalking vacuumingwiping tables runningtreadmill runningcycling sittingdowngetting up andlying downgetting up
Accelerationsensors
Averaged entropy over 3axes main FFT frequency(averaged) over 3 axesenergy of the 02Hzwindow centred aroundmain frequency over totalFFT energy (3-axisaverage) and averagedmean of cross covariancebetween every 2 axes
ReliefF Simbaand MRMR
kNN Bayesianclassifier 90
Bayat et al [40]
Running slow walk fastwalk aerobic dancingstairs up and stairsdown
Triaxialaccelerometer
Mean along 119911-axisMinMax STD and RMSfor Am APF along 119909-axis119910-axis and 119911-axis VarAPFSTD along 119909-axis 119910-axisand 119911-axis RMS along119909-axis 119910-axis and 119911-axiscorrelation between 119911-axisand 119910-axis and MinMaxalong 119909-axis 119910-axis and119911-axis
Energy entropy meanvariance mean trendwindowed mean differencevariance trend windowedvariance differencedetrended fluctuationanalysis coefficients119883-119885-energy and maxdifference acceleration
ReliefF SFFS kNN NaiveBayes 98
Henpraserttaeet al [42]
Sitting lying standingand walking Accelerometer Mean and standard
Kose et al [45]Walking runningbiking sitting andstanding
AccelerometerMin max averagevariance FFT coefficientsand autocorrelation
None Clustered kNN 952ndash975
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
2 Computational and Mathematical Methods in Medicine
such as smartphones Smartphones are innovative platformsfor HAR because of the availability of different wirelessinterfaces unobtrusiveness ease of use high computingpower and storage and the availability of sensors suchas accelerometer compass and gyroscope which meet thetechnical and practical hardware requirements for HAR tasks[13ndash15] Moreover technological development possibilities ofother applications are still arising including virtual realitysystemsTherefore these machines present a great possibilityfor the development of innovative technology dedicated forthe AAL systems
One of the keymotivating factors for usingmobile phone-based human activity recognition in the AAL systems is therelationship and correlation between the level of physicalactivity and the level of wellbeing of a person Recording andanalysing precise information on the personrsquos activities arebeneficial to keeping the progress and status of the disease (ormental condition) and can potentially improve the treatmentof personrsquos conditions and diseases as well as decreasingthe cost of care Recognizing indoor and outdoor activitiessuch as walking running or cycling can be useful to providefeedback to the caregiver about the patientrsquos behaviourWhenfollowing the daily habits and routines of users one caneasily identify deviations from routines which can assistthe doctors in diagnosing conditions that would not beobserved during routine medical examination Another keyenabler of the HAR technology is the possibility of providingindependent living for the elderly as well as for patientswith dementia and other mental pathologies which could bemonitored to prevent undesirable consequences of abnormalactivities Furthermore by using persuasive techniques andgamification HAR systems can be designed to interact withusers to change their behaviour and lifestyles towards moreactive and healthier ones [16]
Recently various intelligent systems based on mobiletechnologies have been constructedHARusing smartphonesor other types of portable or wearable sensor platformshas been used for assessing movement quality after stroke[17] such as upper extremity motion [18] for assessinggait characteristics of human locomotion for rehabilitationand diagnosis of medical conditions [19] for postoperativemobilization [20] for detecting Parkinsonrsquos disease backpain and hemiparesis [21] for cardiac rehabilitation [22] forphysical therapy for example if a user is correctly doing theexercises recommended by a physician [23 24] for detectingabnormal activities arising due to memory loss for dementiacare [25 26] for dealing with Alzheimerrsquos [27] and neu-rodegenerative diseases such as epilepsy [28] for assessmentof physical activity for children and adolescents sufferingfrom hyperlipidaemia hypertension cardiovascular diseaseand type 2 diabetes [29] for detecting falls [30 31] foraddressing physical inactivity when dealing with obesity [32]for analysing sleeping patterns [33] for estimating energyexpenditures of a person to assess hisher healthy dailylifestyle [34] and for recognizing the userrsquos intent in thedomain of rehabilitation engineering such as smart walkingsupport systems to assist motor-impaired persons and theelderly [35]
In this paper we propose a new method for offlinerecognition of daily human activities based on feature dimen-sionality reduction using random projections [36] to lowdimensionality feature space and using the Jaccard distancebetween kernel density probabilities as a decision function forclassification of human activities
The structure of the remaining parts of the paper isas follows Section 2 presents the overview of related workin the smartphone-based HAR domain with a particularemphasis on the features extracted from the sensor dataSection 3 describes the proposedmethod Section 4 evaluatesand discusses the results Finally Section 5 presents theconclusions and discusses future work
2 Overview of HAR Featuresand Related Work
All tasks of the HAR domain require correct identification ofhuman activities from sensor data which in turn requiresthat features derived from sensor data must be properlycategorized and described Next we present an overview offeatures used in the HAR domain
21 Features While numerous features can be extracted fromphysical activity signals increasing the number of featuresdoes not necessarily increase classification accuracy since thefeatures may be redundant or may not be class-specific
(i) Time domain features (such as mean median vari-ance standard deviation minimum maximum androot mean square applied to the amplitude andtime dimensions of a signal) are typically used inmany practical HAR systems because of being lesscomputationally intensive thus they can be easilyextracted in real time
(ii) Frequency-domain features require higher computa-tional cost to distinguish between different humanactivitiesThus they may not be suitable for real-timeAAL applications
(iii) Physical features are derived from a fundamentalunderstanding of how a certain human movementwould produce a specific sensor signal Physicalfeatures are usually extracted from multiple sensoraxes based on the physical parameters of humanmovements
Based on the extensive analysis of the literature andfeatures used by other authors (esp by Capela et al [17]Mathie et al [37] and Zhang and Sawchuk [38]) we haveextracted 99 features of data which are detailed in Table 1
22 Feature Selection Feature selection is the process ofselecting a subset of relevant features for use in constructionof the classification model Successful selection of featuresallows for simplification of models to make them easier tointerpret to decrease model training times and to betterunderstand difference between classes Using feature selec-tion allows removing redundant or irrelevant features with-out having an adverse effect on the classification accuracy
Computational and Mathematical Methods in Medicine 3
Table 1 Catalogue of features
Feature number Description Equation (notation)4ndash6 Acceleration (119909- 119910- and 119911-axes) 119886
119909 119886119910 119886119911
7ndash9 Gyroscope (119909- 119910- and 119911-axes) 119892119909 119892119910 119892119911
10ndash15 Moving variance of 100 samples ofacceleration and gyroscope data
var = 1
119873 (119873 minus 1)
(119873
119873
sum
119894=1
1199092
119894minus (
119873
sum
119894=1
119909119894)
2
) here 119909 = 119886119909 119886119910 119886119911 119892119909 119892119910 119892119911
16-17 Movement intensity of acceleration andgyroscope data
MI119886= radic119886
2
119909+ 1198862
119910+ 1198862
119911
MI119892= radic119892
2
119909+ 1198922
119910+ 1198922
119911
18 Movement intensity of differencebetween acceleration and gyroscope data MI
119892119886= radic(119892
119909minus 119886119909)2
+ (119892119910minus 119886119910)
2
+ (119892119910minus 119886119910)
2
19ndash21 Moving variance of 100 samples ofmovement intensity data
var = 1
119873 (119873 minus 1)
(119873
119873
sum
119894=1
1199092
119894minus (
119873
sum
119894=1
119909119894)
2
) here 119909 = MI119886MI119892MI119892119886
22ndash24 Polar coordinates of acceleration data120593119886= arctan (119886
119910 119886119909)
119903119886= radic119886
2
119909+ 1198862
119910
119911119886= 119886119911
25ndash27 Polar coordinates of gyroscope data120593119892= arctan (119892
119910 119892119909)
119903119892= radic119892
2
119909+ 1198922
119910
119911119892= 119892119911
28ndash30 Polar coordinates of difference betweenacceleration and gyroscope data
82 Simple moving average of cumulativesums of acceleration data
SMAMCS119886=
1
119873
119873
sum
119894=1
MCS119886119894
83 Simple moving average of cumulativesums of gyroscope data
SMAMCS119892=
1
119873
119873
sum
119894=1
MCS119892119894
Computational and Mathematical Methods in Medicine 5
Table 1 Continued
Feature number Description Equation (notation)
84Simple moving average of cumulativesums of difference between accelerometerand gyroscope data
SMAMCS119886119892=
1
119873
119873
sum
119894=1
(MCS119886119894minusMCS
119892119894)
85ndash90 Moving 2nd-order cumulative sum ofacceleration and gyroscope data
MCS1015840 =119873
sum
119895=1
119895
sum
119894=1
MCS119894
91ndash93Moving 2nd-order cumulative sum ofdifferences between cumulative sums ofacceleration and gyroscope data
MCS1015840119886119892=
119873
sum
119895=1
119895
sum
119894=1
(MCS119886119894minusMCS119892
119894)
94ndash96 Polar coordinates of moving cumulativesum of acceleration data
1205931015840
119886= arctan (MCS119886
119910MCS119886
119909)
1199031015840
119886= radic(MCS119886
119909)2
+ (MCS119886119910)
2
1199111015840
119886= MCS119886
119911
97ndash99 Polar coordinates of moving cumulativesum of gyroscope data
1205931015840
119892= arctan (MCS119892
119910MCS119892
119909)
1199031015840
119892= radic(MCS119892
119909)2
+ (MCS119892119910)
2
1199111015840
119892= MCS119892
119911
100ndash102Polar coordinates of moving cumulativesum of differences between accelerationand gyroscope data
1205931015840
119886119892= arctan (MCS119886
119910minusMCS119892
119910MCS119886
119909minusMCS119892
119909)
1199031015840
119886119892= radic(MCS119886
119909minusMCS119892
119909)2
+ (MCS119886119910minusMCS119892
119910)
2
1199111015840
119886119892= MCS119886
119911minusMCS119892
119911
There are four basic steps in a typical feature selectionmethod[58] generation of candidate feature subset an evaluationfunction for feature candidate subset a generation stoppingcriterion and a validation procedure
Further we analyse several feature selection methodsused in the HAR domain
ReliefF [59] is a commonly used filter method that ranksfeatures by weighting them based on their relevance Featurerelevance is based on how well data instances are separatedFor each data instance the algorithm finds the nearest datapoint from the same class (hit) and nearest data points fromdifferent classes (misses)
Matlabrsquos Rankfeatures ranks features by a given classseparability criterion Class separability measures include theabsolute value of a statistic of a two-sample 119905-test Kullback-Leibler distance minimum attainable classification errorarea between the empirical ReceiverOperatingCharacteristic(ROC) curve and the randomclassifier slope and the absolutevalue of the statistic of a two-sample unpairedWilcoxon testMeasures are based on distributional characteristics of classes(eg mean variance) for a feature
Principal component analysis (PCA) is the simplestmethod to reduce data dimensionality This reduced dimen-sional data can be used directly as features for classificationGiven a set of 119873 features a PCA analysis will produce newdata variables (PCA components) as linear combinationsof the features with the highest variance in the subspaceorthogonal to the preceding PCA component As variabilityof the data can be captured by a relatively small number ofPCs PCA can achieve high level of dimensionality reductionSeveral extensions of the PCA method are known such askernel PCA sparse PCA and multilinear PCA
Correlation-based Feature Selection (CFS) [60] is a filteralgorithm that ranks subsets of features by a correlation-based heuristic evaluation function A feature is consideredto be a good one if it is relevant to the target concept but isnot redundant to any of the other relevant features Goodnessof measure is expressed by a correlation between featuresand CFS chooses the subset of features which has the highestmeasure The chosen subset holds the property that featuresinside this subset have high correlation with the class and areunrelated to each other
Table 2 summarizes the feature selectiondimensionalityreduction methods in HAR
A comprehensive review of feature selection algorithmsin general as well as in the HAR domain can be found in [5861ndash63]
23 Summary Related work in the HAR domain is summa-rized in Table 3 For each paper the activities analysed typesof sensor data used features extracted classification methodapplied and accuracy achieved (as given by the referencedpapers) are given
3 Method
31 General Scheme The typical steps for activity recognitionare preprocessing segmentation feature extraction dimen-sionality reduction (feature selection) and classification [24]Themain steps of activity recognition include (a) preprocess-ing of sensor data (eg denoising) (b) feature extraction(c) dimension reduction and (d) classification The prepro-cessing step includes noise removal and representation ofraw data The feature extraction step is used to reduce large
6 Computational and Mathematical Methods in Medicine
Table 2 Summary of feature selectiondimensionality reduction methods in HAR
Method Advantages Disadvantages Complexity
PCA
High dimensionalityreduction reduction of noiselack of redundancy of data dueto orthogonality ofcomponents
The covariance matrix is difficult to beevaluated accurately even the simplestinvariance could not be captured by thePCA unless the training data explicitlyprovides for it
119874(1199012
119899 + 1199013
) where 119899 are datapoints each represented with119901 features
ReliefF Low computationalcomplexity
Unstable due to random selection ofinstances 119874(119901 sdot 119899 sdot log 119899)
RankfeaturesFeatures highly correlatedwith already selected featuresare less likely to be included
It assumes that data classes are normallydistributed
It depends upon classseparability criterion
CFS It evaluates a subset of featuresrather than individual features
It fails to select locally predictive featureswhen they are overshadowed by strongglobally predictive features
119874(119899(
1199012
minus 119901
2
))
input sensor data to a smaller set of features (feature vector)which preserves information contained in the original dataThe dimensionality reduction step can be applied to removethe irrelevant (or less relevant) features and reduce thecomputational complexity and increase the performance ofthe activity recognition processThe classification step is usedto map the feature set to a set of activities
In this paper we do not focus on data preprocessing andfeature extraction but rather on dimensionality reduction andclassification steps since these two are crucial for furtherefficiency of AAL systems The proposed method for humanactivity recognition is based on feature dimensionality reduc-tion using random projections [36] and classification usingkernel density function estimate as a decision function(Figure 1)
32 Description of the Method During random projec-tion the original 119889-dimensional data is projected to a 119896-dimensional (119896 ≪ 119889) subspace using a random 119896 times 119889 matrix119877 The projection of the data onto a lower 119896-dimensionalsubspace is 119883119877119875
119896times119873= 119877119896times119889
119883119889times119873
where 119883119889times119873
is the originalset of 119873 119889-dimensional observations In the derived pro-jection the distances between the points are approximatelypreserved if points in a vector space are projected onto arandomly selected subspace of suitably high dimension (seethe Johnson-Lindenstrauss lemma [64]) The randommatrix119877 is selected as proposed by Achlioptas [36] as follows
119903119894119895=
+1 probability 16
0 probability 23
minus1 probability 16
(1)
Given the low dimensionality of the target space we cantreat the projection of low-dimensional observations ontoeach dimension as a set of random variables for which theprobability density function (PDF) can be estimated usingkernel density estimation (KDE) (or Parzenwindow)method[65]
If 1199091 1199092 119909
119873is a sample of a random variable then
the kernel density approximation of its probability densityfunction is
119891ℎ(119909) =
1
119873ℎ
119870(
119909 minus 119909119894
ℎ
) (2)
where 119870 is some kernel and ℎ is the bandwidth (smoothingparameter) 119870 is taken to be a standard Gaussian functionwith mean zero and variance 1 of the examined data features
119870 (119909) =
1
radic2120587
119890minus(12)119909
2
(3)
For a two-dimensional case the bivariate probabilitydensity function is calculated as a product of univariateprobability functions as follows
119891 (119909 119910) =
119891 (119909) sdot
119891 (119910) (4)
Here 119909 and 119910 are data in each dimension respectivelyHowever each random projection produces a different
mapping of the original data points which reveals only a partof the data manifold in higher-dimensional space In caseof the binary classification problem we are interested in amapping that separates data points belonging to two differentclasses best
As a criterion for estimating the mapping we use theJaccard distance metric between two probability density esti-mates of data points representing each class The advantageof the Jaccard distance metric as compared to other metricsof distance such as Kullback-Leibler (KL) divergence andHellinger distance is its adaptability to multidimensionalspaces where compared points show relations to differentsubsets Therefore it is well adapted to the developed modelof human activity features where according to description inthe previous section we have divided them into some sets ofactions Furthermore the computational complexity of theHellinger distance is very high while KL divergence mightbe unbounded
The Jaccard distance which measures dissimilaritybetween sample sets is obtained by subtracting the Jaccardcoefficient from 1 or equivalently by dividing the difference
Computational and Mathematical Methods in Medicine 7
Table 3 Summary of related works in the HAR domain
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Atallah et al[39]
Lying down preparingfood eating anddrinking socialisingreading dressingwalking treadmillwalking vacuumingwiping tables runningtreadmill runningcycling sittingdowngetting up andlying downgetting up
Accelerationsensors
Averaged entropy over 3axes main FFT frequency(averaged) over 3 axesenergy of the 02Hzwindow centred aroundmain frequency over totalFFT energy (3-axisaverage) and averagedmean of cross covariancebetween every 2 axes
ReliefF Simbaand MRMR
kNN Bayesianclassifier 90
Bayat et al [40]
Running slow walk fastwalk aerobic dancingstairs up and stairsdown
Triaxialaccelerometer
Mean along 119911-axisMinMax STD and RMSfor Am APF along 119909-axis119910-axis and 119911-axis VarAPFSTD along 119909-axis 119910-axisand 119911-axis RMS along119909-axis 119910-axis and 119911-axiscorrelation between 119911-axisand 119910-axis and MinMaxalong 119909-axis 119910-axis and119911-axis
Energy entropy meanvariance mean trendwindowed mean differencevariance trend windowedvariance differencedetrended fluctuationanalysis coefficients119883-119885-energy and maxdifference acceleration
ReliefF SFFS kNN NaiveBayes 98
Henpraserttaeet al [42]
Sitting lying standingand walking Accelerometer Mean and standard
Kose et al [45]Walking runningbiking sitting andstanding
AccelerometerMin max averagevariance FFT coefficientsand autocorrelation
None Clustered kNN 952ndash975
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
82 Simple moving average of cumulativesums of acceleration data
SMAMCS119886=
1
119873
119873
sum
119894=1
MCS119886119894
83 Simple moving average of cumulativesums of gyroscope data
SMAMCS119892=
1
119873
119873
sum
119894=1
MCS119892119894
Computational and Mathematical Methods in Medicine 5
Table 1 Continued
Feature number Description Equation (notation)
84Simple moving average of cumulativesums of difference between accelerometerand gyroscope data
SMAMCS119886119892=
1
119873
119873
sum
119894=1
(MCS119886119894minusMCS
119892119894)
85ndash90 Moving 2nd-order cumulative sum ofacceleration and gyroscope data
MCS1015840 =119873
sum
119895=1
119895
sum
119894=1
MCS119894
91ndash93Moving 2nd-order cumulative sum ofdifferences between cumulative sums ofacceleration and gyroscope data
MCS1015840119886119892=
119873
sum
119895=1
119895
sum
119894=1
(MCS119886119894minusMCS119892
119894)
94ndash96 Polar coordinates of moving cumulativesum of acceleration data
1205931015840
119886= arctan (MCS119886
119910MCS119886
119909)
1199031015840
119886= radic(MCS119886
119909)2
+ (MCS119886119910)
2
1199111015840
119886= MCS119886
119911
97ndash99 Polar coordinates of moving cumulativesum of gyroscope data
1205931015840
119892= arctan (MCS119892
119910MCS119892
119909)
1199031015840
119892= radic(MCS119892
119909)2
+ (MCS119892119910)
2
1199111015840
119892= MCS119892
119911
100ndash102Polar coordinates of moving cumulativesum of differences between accelerationand gyroscope data
1205931015840
119886119892= arctan (MCS119886
119910minusMCS119892
119910MCS119886
119909minusMCS119892
119909)
1199031015840
119886119892= radic(MCS119886
119909minusMCS119892
119909)2
+ (MCS119886119910minusMCS119892
119910)
2
1199111015840
119886119892= MCS119886
119911minusMCS119892
119911
There are four basic steps in a typical feature selectionmethod[58] generation of candidate feature subset an evaluationfunction for feature candidate subset a generation stoppingcriterion and a validation procedure
Further we analyse several feature selection methodsused in the HAR domain
ReliefF [59] is a commonly used filter method that ranksfeatures by weighting them based on their relevance Featurerelevance is based on how well data instances are separatedFor each data instance the algorithm finds the nearest datapoint from the same class (hit) and nearest data points fromdifferent classes (misses)
Matlabrsquos Rankfeatures ranks features by a given classseparability criterion Class separability measures include theabsolute value of a statistic of a two-sample 119905-test Kullback-Leibler distance minimum attainable classification errorarea between the empirical ReceiverOperatingCharacteristic(ROC) curve and the randomclassifier slope and the absolutevalue of the statistic of a two-sample unpairedWilcoxon testMeasures are based on distributional characteristics of classes(eg mean variance) for a feature
Principal component analysis (PCA) is the simplestmethod to reduce data dimensionality This reduced dimen-sional data can be used directly as features for classificationGiven a set of 119873 features a PCA analysis will produce newdata variables (PCA components) as linear combinationsof the features with the highest variance in the subspaceorthogonal to the preceding PCA component As variabilityof the data can be captured by a relatively small number ofPCs PCA can achieve high level of dimensionality reductionSeveral extensions of the PCA method are known such askernel PCA sparse PCA and multilinear PCA
Correlation-based Feature Selection (CFS) [60] is a filteralgorithm that ranks subsets of features by a correlation-based heuristic evaluation function A feature is consideredto be a good one if it is relevant to the target concept but isnot redundant to any of the other relevant features Goodnessof measure is expressed by a correlation between featuresand CFS chooses the subset of features which has the highestmeasure The chosen subset holds the property that featuresinside this subset have high correlation with the class and areunrelated to each other
Table 2 summarizes the feature selectiondimensionalityreduction methods in HAR
A comprehensive review of feature selection algorithmsin general as well as in the HAR domain can be found in [5861ndash63]
23 Summary Related work in the HAR domain is summa-rized in Table 3 For each paper the activities analysed typesof sensor data used features extracted classification methodapplied and accuracy achieved (as given by the referencedpapers) are given
3 Method
31 General Scheme The typical steps for activity recognitionare preprocessing segmentation feature extraction dimen-sionality reduction (feature selection) and classification [24]Themain steps of activity recognition include (a) preprocess-ing of sensor data (eg denoising) (b) feature extraction(c) dimension reduction and (d) classification The prepro-cessing step includes noise removal and representation ofraw data The feature extraction step is used to reduce large
6 Computational and Mathematical Methods in Medicine
Table 2 Summary of feature selectiondimensionality reduction methods in HAR
Method Advantages Disadvantages Complexity
PCA
High dimensionalityreduction reduction of noiselack of redundancy of data dueto orthogonality ofcomponents
The covariance matrix is difficult to beevaluated accurately even the simplestinvariance could not be captured by thePCA unless the training data explicitlyprovides for it
119874(1199012
119899 + 1199013
) where 119899 are datapoints each represented with119901 features
ReliefF Low computationalcomplexity
Unstable due to random selection ofinstances 119874(119901 sdot 119899 sdot log 119899)
RankfeaturesFeatures highly correlatedwith already selected featuresare less likely to be included
It assumes that data classes are normallydistributed
It depends upon classseparability criterion
CFS It evaluates a subset of featuresrather than individual features
It fails to select locally predictive featureswhen they are overshadowed by strongglobally predictive features
119874(119899(
1199012
minus 119901
2
))
input sensor data to a smaller set of features (feature vector)which preserves information contained in the original dataThe dimensionality reduction step can be applied to removethe irrelevant (or less relevant) features and reduce thecomputational complexity and increase the performance ofthe activity recognition processThe classification step is usedto map the feature set to a set of activities
In this paper we do not focus on data preprocessing andfeature extraction but rather on dimensionality reduction andclassification steps since these two are crucial for furtherefficiency of AAL systems The proposed method for humanactivity recognition is based on feature dimensionality reduc-tion using random projections [36] and classification usingkernel density function estimate as a decision function(Figure 1)
32 Description of the Method During random projec-tion the original 119889-dimensional data is projected to a 119896-dimensional (119896 ≪ 119889) subspace using a random 119896 times 119889 matrix119877 The projection of the data onto a lower 119896-dimensionalsubspace is 119883119877119875
119896times119873= 119877119896times119889
119883119889times119873
where 119883119889times119873
is the originalset of 119873 119889-dimensional observations In the derived pro-jection the distances between the points are approximatelypreserved if points in a vector space are projected onto arandomly selected subspace of suitably high dimension (seethe Johnson-Lindenstrauss lemma [64]) The randommatrix119877 is selected as proposed by Achlioptas [36] as follows
119903119894119895=
+1 probability 16
0 probability 23
minus1 probability 16
(1)
Given the low dimensionality of the target space we cantreat the projection of low-dimensional observations ontoeach dimension as a set of random variables for which theprobability density function (PDF) can be estimated usingkernel density estimation (KDE) (or Parzenwindow)method[65]
If 1199091 1199092 119909
119873is a sample of a random variable then
the kernel density approximation of its probability densityfunction is
119891ℎ(119909) =
1
119873ℎ
119870(
119909 minus 119909119894
ℎ
) (2)
where 119870 is some kernel and ℎ is the bandwidth (smoothingparameter) 119870 is taken to be a standard Gaussian functionwith mean zero and variance 1 of the examined data features
119870 (119909) =
1
radic2120587
119890minus(12)119909
2
(3)
For a two-dimensional case the bivariate probabilitydensity function is calculated as a product of univariateprobability functions as follows
119891 (119909 119910) =
119891 (119909) sdot
119891 (119910) (4)
Here 119909 and 119910 are data in each dimension respectivelyHowever each random projection produces a different
mapping of the original data points which reveals only a partof the data manifold in higher-dimensional space In caseof the binary classification problem we are interested in amapping that separates data points belonging to two differentclasses best
As a criterion for estimating the mapping we use theJaccard distance metric between two probability density esti-mates of data points representing each class The advantageof the Jaccard distance metric as compared to other metricsof distance such as Kullback-Leibler (KL) divergence andHellinger distance is its adaptability to multidimensionalspaces where compared points show relations to differentsubsets Therefore it is well adapted to the developed modelof human activity features where according to description inthe previous section we have divided them into some sets ofactions Furthermore the computational complexity of theHellinger distance is very high while KL divergence mightbe unbounded
The Jaccard distance which measures dissimilaritybetween sample sets is obtained by subtracting the Jaccardcoefficient from 1 or equivalently by dividing the difference
Computational and Mathematical Methods in Medicine 7
Table 3 Summary of related works in the HAR domain
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Atallah et al[39]
Lying down preparingfood eating anddrinking socialisingreading dressingwalking treadmillwalking vacuumingwiping tables runningtreadmill runningcycling sittingdowngetting up andlying downgetting up
Accelerationsensors
Averaged entropy over 3axes main FFT frequency(averaged) over 3 axesenergy of the 02Hzwindow centred aroundmain frequency over totalFFT energy (3-axisaverage) and averagedmean of cross covariancebetween every 2 axes
ReliefF Simbaand MRMR
kNN Bayesianclassifier 90
Bayat et al [40]
Running slow walk fastwalk aerobic dancingstairs up and stairsdown
Triaxialaccelerometer
Mean along 119911-axisMinMax STD and RMSfor Am APF along 119909-axis119910-axis and 119911-axis VarAPFSTD along 119909-axis 119910-axisand 119911-axis RMS along119909-axis 119910-axis and 119911-axiscorrelation between 119911-axisand 119910-axis and MinMaxalong 119909-axis 119910-axis and119911-axis
Energy entropy meanvariance mean trendwindowed mean differencevariance trend windowedvariance differencedetrended fluctuationanalysis coefficients119883-119885-energy and maxdifference acceleration
ReliefF SFFS kNN NaiveBayes 98
Henpraserttaeet al [42]
Sitting lying standingand walking Accelerometer Mean and standard
Kose et al [45]Walking runningbiking sitting andstanding
AccelerometerMin max averagevariance FFT coefficientsand autocorrelation
None Clustered kNN 952ndash975
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
82 Simple moving average of cumulativesums of acceleration data
SMAMCS119886=
1
119873
119873
sum
119894=1
MCS119886119894
83 Simple moving average of cumulativesums of gyroscope data
SMAMCS119892=
1
119873
119873
sum
119894=1
MCS119892119894
Computational and Mathematical Methods in Medicine 5
Table 1 Continued
Feature number Description Equation (notation)
84Simple moving average of cumulativesums of difference between accelerometerand gyroscope data
SMAMCS119886119892=
1
119873
119873
sum
119894=1
(MCS119886119894minusMCS
119892119894)
85ndash90 Moving 2nd-order cumulative sum ofacceleration and gyroscope data
MCS1015840 =119873
sum
119895=1
119895
sum
119894=1
MCS119894
91ndash93Moving 2nd-order cumulative sum ofdifferences between cumulative sums ofacceleration and gyroscope data
MCS1015840119886119892=
119873
sum
119895=1
119895
sum
119894=1
(MCS119886119894minusMCS119892
119894)
94ndash96 Polar coordinates of moving cumulativesum of acceleration data
1205931015840
119886= arctan (MCS119886
119910MCS119886
119909)
1199031015840
119886= radic(MCS119886
119909)2
+ (MCS119886119910)
2
1199111015840
119886= MCS119886
119911
97ndash99 Polar coordinates of moving cumulativesum of gyroscope data
1205931015840
119892= arctan (MCS119892
119910MCS119892
119909)
1199031015840
119892= radic(MCS119892
119909)2
+ (MCS119892119910)
2
1199111015840
119892= MCS119892
119911
100ndash102Polar coordinates of moving cumulativesum of differences between accelerationand gyroscope data
1205931015840
119886119892= arctan (MCS119886
119910minusMCS119892
119910MCS119886
119909minusMCS119892
119909)
1199031015840
119886119892= radic(MCS119886
119909minusMCS119892
119909)2
+ (MCS119886119910minusMCS119892
119910)
2
1199111015840
119886119892= MCS119886
119911minusMCS119892
119911
There are four basic steps in a typical feature selectionmethod[58] generation of candidate feature subset an evaluationfunction for feature candidate subset a generation stoppingcriterion and a validation procedure
Further we analyse several feature selection methodsused in the HAR domain
ReliefF [59] is a commonly used filter method that ranksfeatures by weighting them based on their relevance Featurerelevance is based on how well data instances are separatedFor each data instance the algorithm finds the nearest datapoint from the same class (hit) and nearest data points fromdifferent classes (misses)
Matlabrsquos Rankfeatures ranks features by a given classseparability criterion Class separability measures include theabsolute value of a statistic of a two-sample 119905-test Kullback-Leibler distance minimum attainable classification errorarea between the empirical ReceiverOperatingCharacteristic(ROC) curve and the randomclassifier slope and the absolutevalue of the statistic of a two-sample unpairedWilcoxon testMeasures are based on distributional characteristics of classes(eg mean variance) for a feature
Principal component analysis (PCA) is the simplestmethod to reduce data dimensionality This reduced dimen-sional data can be used directly as features for classificationGiven a set of 119873 features a PCA analysis will produce newdata variables (PCA components) as linear combinationsof the features with the highest variance in the subspaceorthogonal to the preceding PCA component As variabilityof the data can be captured by a relatively small number ofPCs PCA can achieve high level of dimensionality reductionSeveral extensions of the PCA method are known such askernel PCA sparse PCA and multilinear PCA
Correlation-based Feature Selection (CFS) [60] is a filteralgorithm that ranks subsets of features by a correlation-based heuristic evaluation function A feature is consideredto be a good one if it is relevant to the target concept but isnot redundant to any of the other relevant features Goodnessof measure is expressed by a correlation between featuresand CFS chooses the subset of features which has the highestmeasure The chosen subset holds the property that featuresinside this subset have high correlation with the class and areunrelated to each other
Table 2 summarizes the feature selectiondimensionalityreduction methods in HAR
A comprehensive review of feature selection algorithmsin general as well as in the HAR domain can be found in [5861ndash63]
23 Summary Related work in the HAR domain is summa-rized in Table 3 For each paper the activities analysed typesof sensor data used features extracted classification methodapplied and accuracy achieved (as given by the referencedpapers) are given
3 Method
31 General Scheme The typical steps for activity recognitionare preprocessing segmentation feature extraction dimen-sionality reduction (feature selection) and classification [24]Themain steps of activity recognition include (a) preprocess-ing of sensor data (eg denoising) (b) feature extraction(c) dimension reduction and (d) classification The prepro-cessing step includes noise removal and representation ofraw data The feature extraction step is used to reduce large
6 Computational and Mathematical Methods in Medicine
Table 2 Summary of feature selectiondimensionality reduction methods in HAR
Method Advantages Disadvantages Complexity
PCA
High dimensionalityreduction reduction of noiselack of redundancy of data dueto orthogonality ofcomponents
The covariance matrix is difficult to beevaluated accurately even the simplestinvariance could not be captured by thePCA unless the training data explicitlyprovides for it
119874(1199012
119899 + 1199013
) where 119899 are datapoints each represented with119901 features
ReliefF Low computationalcomplexity
Unstable due to random selection ofinstances 119874(119901 sdot 119899 sdot log 119899)
RankfeaturesFeatures highly correlatedwith already selected featuresare less likely to be included
It assumes that data classes are normallydistributed
It depends upon classseparability criterion
CFS It evaluates a subset of featuresrather than individual features
It fails to select locally predictive featureswhen they are overshadowed by strongglobally predictive features
119874(119899(
1199012
minus 119901
2
))
input sensor data to a smaller set of features (feature vector)which preserves information contained in the original dataThe dimensionality reduction step can be applied to removethe irrelevant (or less relevant) features and reduce thecomputational complexity and increase the performance ofthe activity recognition processThe classification step is usedto map the feature set to a set of activities
In this paper we do not focus on data preprocessing andfeature extraction but rather on dimensionality reduction andclassification steps since these two are crucial for furtherefficiency of AAL systems The proposed method for humanactivity recognition is based on feature dimensionality reduc-tion using random projections [36] and classification usingkernel density function estimate as a decision function(Figure 1)
32 Description of the Method During random projec-tion the original 119889-dimensional data is projected to a 119896-dimensional (119896 ≪ 119889) subspace using a random 119896 times 119889 matrix119877 The projection of the data onto a lower 119896-dimensionalsubspace is 119883119877119875
119896times119873= 119877119896times119889
119883119889times119873
where 119883119889times119873
is the originalset of 119873 119889-dimensional observations In the derived pro-jection the distances between the points are approximatelypreserved if points in a vector space are projected onto arandomly selected subspace of suitably high dimension (seethe Johnson-Lindenstrauss lemma [64]) The randommatrix119877 is selected as proposed by Achlioptas [36] as follows
119903119894119895=
+1 probability 16
0 probability 23
minus1 probability 16
(1)
Given the low dimensionality of the target space we cantreat the projection of low-dimensional observations ontoeach dimension as a set of random variables for which theprobability density function (PDF) can be estimated usingkernel density estimation (KDE) (or Parzenwindow)method[65]
If 1199091 1199092 119909
119873is a sample of a random variable then
the kernel density approximation of its probability densityfunction is
119891ℎ(119909) =
1
119873ℎ
119870(
119909 minus 119909119894
ℎ
) (2)
where 119870 is some kernel and ℎ is the bandwidth (smoothingparameter) 119870 is taken to be a standard Gaussian functionwith mean zero and variance 1 of the examined data features
119870 (119909) =
1
radic2120587
119890minus(12)119909
2
(3)
For a two-dimensional case the bivariate probabilitydensity function is calculated as a product of univariateprobability functions as follows
119891 (119909 119910) =
119891 (119909) sdot
119891 (119910) (4)
Here 119909 and 119910 are data in each dimension respectivelyHowever each random projection produces a different
mapping of the original data points which reveals only a partof the data manifold in higher-dimensional space In caseof the binary classification problem we are interested in amapping that separates data points belonging to two differentclasses best
As a criterion for estimating the mapping we use theJaccard distance metric between two probability density esti-mates of data points representing each class The advantageof the Jaccard distance metric as compared to other metricsof distance such as Kullback-Leibler (KL) divergence andHellinger distance is its adaptability to multidimensionalspaces where compared points show relations to differentsubsets Therefore it is well adapted to the developed modelof human activity features where according to description inthe previous section we have divided them into some sets ofactions Furthermore the computational complexity of theHellinger distance is very high while KL divergence mightbe unbounded
The Jaccard distance which measures dissimilaritybetween sample sets is obtained by subtracting the Jaccardcoefficient from 1 or equivalently by dividing the difference
Computational and Mathematical Methods in Medicine 7
Table 3 Summary of related works in the HAR domain
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Atallah et al[39]
Lying down preparingfood eating anddrinking socialisingreading dressingwalking treadmillwalking vacuumingwiping tables runningtreadmill runningcycling sittingdowngetting up andlying downgetting up
Accelerationsensors
Averaged entropy over 3axes main FFT frequency(averaged) over 3 axesenergy of the 02Hzwindow centred aroundmain frequency over totalFFT energy (3-axisaverage) and averagedmean of cross covariancebetween every 2 axes
ReliefF Simbaand MRMR
kNN Bayesianclassifier 90
Bayat et al [40]
Running slow walk fastwalk aerobic dancingstairs up and stairsdown
Triaxialaccelerometer
Mean along 119911-axisMinMax STD and RMSfor Am APF along 119909-axis119910-axis and 119911-axis VarAPFSTD along 119909-axis 119910-axisand 119911-axis RMS along119909-axis 119910-axis and 119911-axiscorrelation between 119911-axisand 119910-axis and MinMaxalong 119909-axis 119910-axis and119911-axis
Energy entropy meanvariance mean trendwindowed mean differencevariance trend windowedvariance differencedetrended fluctuationanalysis coefficients119883-119885-energy and maxdifference acceleration
ReliefF SFFS kNN NaiveBayes 98
Henpraserttaeet al [42]
Sitting lying standingand walking Accelerometer Mean and standard
Kose et al [45]Walking runningbiking sitting andstanding
AccelerometerMin max averagevariance FFT coefficientsand autocorrelation
None Clustered kNN 952ndash975
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
Computational and Mathematical Methods in Medicine 5
Table 1 Continued
Feature number Description Equation (notation)
84Simple moving average of cumulativesums of difference between accelerometerand gyroscope data
SMAMCS119886119892=
1
119873
119873
sum
119894=1
(MCS119886119894minusMCS
119892119894)
85ndash90 Moving 2nd-order cumulative sum ofacceleration and gyroscope data
MCS1015840 =119873
sum
119895=1
119895
sum
119894=1
MCS119894
91ndash93Moving 2nd-order cumulative sum ofdifferences between cumulative sums ofacceleration and gyroscope data
MCS1015840119886119892=
119873
sum
119895=1
119895
sum
119894=1
(MCS119886119894minusMCS119892
119894)
94ndash96 Polar coordinates of moving cumulativesum of acceleration data
1205931015840
119886= arctan (MCS119886
119910MCS119886
119909)
1199031015840
119886= radic(MCS119886
119909)2
+ (MCS119886119910)
2
1199111015840
119886= MCS119886
119911
97ndash99 Polar coordinates of moving cumulativesum of gyroscope data
1205931015840
119892= arctan (MCS119892
119910MCS119892
119909)
1199031015840
119892= radic(MCS119892
119909)2
+ (MCS119892119910)
2
1199111015840
119892= MCS119892
119911
100ndash102Polar coordinates of moving cumulativesum of differences between accelerationand gyroscope data
1205931015840
119886119892= arctan (MCS119886
119910minusMCS119892
119910MCS119886
119909minusMCS119892
119909)
1199031015840
119886119892= radic(MCS119886
119909minusMCS119892
119909)2
+ (MCS119886119910minusMCS119892
119910)
2
1199111015840
119886119892= MCS119886
119911minusMCS119892
119911
There are four basic steps in a typical feature selectionmethod[58] generation of candidate feature subset an evaluationfunction for feature candidate subset a generation stoppingcriterion and a validation procedure
Further we analyse several feature selection methodsused in the HAR domain
ReliefF [59] is a commonly used filter method that ranksfeatures by weighting them based on their relevance Featurerelevance is based on how well data instances are separatedFor each data instance the algorithm finds the nearest datapoint from the same class (hit) and nearest data points fromdifferent classes (misses)
Matlabrsquos Rankfeatures ranks features by a given classseparability criterion Class separability measures include theabsolute value of a statistic of a two-sample 119905-test Kullback-Leibler distance minimum attainable classification errorarea between the empirical ReceiverOperatingCharacteristic(ROC) curve and the randomclassifier slope and the absolutevalue of the statistic of a two-sample unpairedWilcoxon testMeasures are based on distributional characteristics of classes(eg mean variance) for a feature
Principal component analysis (PCA) is the simplestmethod to reduce data dimensionality This reduced dimen-sional data can be used directly as features for classificationGiven a set of 119873 features a PCA analysis will produce newdata variables (PCA components) as linear combinationsof the features with the highest variance in the subspaceorthogonal to the preceding PCA component As variabilityof the data can be captured by a relatively small number ofPCs PCA can achieve high level of dimensionality reductionSeveral extensions of the PCA method are known such askernel PCA sparse PCA and multilinear PCA
Correlation-based Feature Selection (CFS) [60] is a filteralgorithm that ranks subsets of features by a correlation-based heuristic evaluation function A feature is consideredto be a good one if it is relevant to the target concept but isnot redundant to any of the other relevant features Goodnessof measure is expressed by a correlation between featuresand CFS chooses the subset of features which has the highestmeasure The chosen subset holds the property that featuresinside this subset have high correlation with the class and areunrelated to each other
Table 2 summarizes the feature selectiondimensionalityreduction methods in HAR
A comprehensive review of feature selection algorithmsin general as well as in the HAR domain can be found in [5861ndash63]
23 Summary Related work in the HAR domain is summa-rized in Table 3 For each paper the activities analysed typesof sensor data used features extracted classification methodapplied and accuracy achieved (as given by the referencedpapers) are given
3 Method
31 General Scheme The typical steps for activity recognitionare preprocessing segmentation feature extraction dimen-sionality reduction (feature selection) and classification [24]Themain steps of activity recognition include (a) preprocess-ing of sensor data (eg denoising) (b) feature extraction(c) dimension reduction and (d) classification The prepro-cessing step includes noise removal and representation ofraw data The feature extraction step is used to reduce large
6 Computational and Mathematical Methods in Medicine
Table 2 Summary of feature selectiondimensionality reduction methods in HAR
Method Advantages Disadvantages Complexity
PCA
High dimensionalityreduction reduction of noiselack of redundancy of data dueto orthogonality ofcomponents
The covariance matrix is difficult to beevaluated accurately even the simplestinvariance could not be captured by thePCA unless the training data explicitlyprovides for it
119874(1199012
119899 + 1199013
) where 119899 are datapoints each represented with119901 features
ReliefF Low computationalcomplexity
Unstable due to random selection ofinstances 119874(119901 sdot 119899 sdot log 119899)
RankfeaturesFeatures highly correlatedwith already selected featuresare less likely to be included
It assumes that data classes are normallydistributed
It depends upon classseparability criterion
CFS It evaluates a subset of featuresrather than individual features
It fails to select locally predictive featureswhen they are overshadowed by strongglobally predictive features
119874(119899(
1199012
minus 119901
2
))
input sensor data to a smaller set of features (feature vector)which preserves information contained in the original dataThe dimensionality reduction step can be applied to removethe irrelevant (or less relevant) features and reduce thecomputational complexity and increase the performance ofthe activity recognition processThe classification step is usedto map the feature set to a set of activities
In this paper we do not focus on data preprocessing andfeature extraction but rather on dimensionality reduction andclassification steps since these two are crucial for furtherefficiency of AAL systems The proposed method for humanactivity recognition is based on feature dimensionality reduc-tion using random projections [36] and classification usingkernel density function estimate as a decision function(Figure 1)
32 Description of the Method During random projec-tion the original 119889-dimensional data is projected to a 119896-dimensional (119896 ≪ 119889) subspace using a random 119896 times 119889 matrix119877 The projection of the data onto a lower 119896-dimensionalsubspace is 119883119877119875
119896times119873= 119877119896times119889
119883119889times119873
where 119883119889times119873
is the originalset of 119873 119889-dimensional observations In the derived pro-jection the distances between the points are approximatelypreserved if points in a vector space are projected onto arandomly selected subspace of suitably high dimension (seethe Johnson-Lindenstrauss lemma [64]) The randommatrix119877 is selected as proposed by Achlioptas [36] as follows
119903119894119895=
+1 probability 16
0 probability 23
minus1 probability 16
(1)
Given the low dimensionality of the target space we cantreat the projection of low-dimensional observations ontoeach dimension as a set of random variables for which theprobability density function (PDF) can be estimated usingkernel density estimation (KDE) (or Parzenwindow)method[65]
If 1199091 1199092 119909
119873is a sample of a random variable then
the kernel density approximation of its probability densityfunction is
119891ℎ(119909) =
1
119873ℎ
119870(
119909 minus 119909119894
ℎ
) (2)
where 119870 is some kernel and ℎ is the bandwidth (smoothingparameter) 119870 is taken to be a standard Gaussian functionwith mean zero and variance 1 of the examined data features
119870 (119909) =
1
radic2120587
119890minus(12)119909
2
(3)
For a two-dimensional case the bivariate probabilitydensity function is calculated as a product of univariateprobability functions as follows
119891 (119909 119910) =
119891 (119909) sdot
119891 (119910) (4)
Here 119909 and 119910 are data in each dimension respectivelyHowever each random projection produces a different
mapping of the original data points which reveals only a partof the data manifold in higher-dimensional space In caseof the binary classification problem we are interested in amapping that separates data points belonging to two differentclasses best
As a criterion for estimating the mapping we use theJaccard distance metric between two probability density esti-mates of data points representing each class The advantageof the Jaccard distance metric as compared to other metricsof distance such as Kullback-Leibler (KL) divergence andHellinger distance is its adaptability to multidimensionalspaces where compared points show relations to differentsubsets Therefore it is well adapted to the developed modelof human activity features where according to description inthe previous section we have divided them into some sets ofactions Furthermore the computational complexity of theHellinger distance is very high while KL divergence mightbe unbounded
The Jaccard distance which measures dissimilaritybetween sample sets is obtained by subtracting the Jaccardcoefficient from 1 or equivalently by dividing the difference
Computational and Mathematical Methods in Medicine 7
Table 3 Summary of related works in the HAR domain
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Atallah et al[39]
Lying down preparingfood eating anddrinking socialisingreading dressingwalking treadmillwalking vacuumingwiping tables runningtreadmill runningcycling sittingdowngetting up andlying downgetting up
Accelerationsensors
Averaged entropy over 3axes main FFT frequency(averaged) over 3 axesenergy of the 02Hzwindow centred aroundmain frequency over totalFFT energy (3-axisaverage) and averagedmean of cross covariancebetween every 2 axes
ReliefF Simbaand MRMR
kNN Bayesianclassifier 90
Bayat et al [40]
Running slow walk fastwalk aerobic dancingstairs up and stairsdown
Triaxialaccelerometer
Mean along 119911-axisMinMax STD and RMSfor Am APF along 119909-axis119910-axis and 119911-axis VarAPFSTD along 119909-axis 119910-axisand 119911-axis RMS along119909-axis 119910-axis and 119911-axiscorrelation between 119911-axisand 119910-axis and MinMaxalong 119909-axis 119910-axis and119911-axis
Energy entropy meanvariance mean trendwindowed mean differencevariance trend windowedvariance differencedetrended fluctuationanalysis coefficients119883-119885-energy and maxdifference acceleration
ReliefF SFFS kNN NaiveBayes 98
Henpraserttaeet al [42]
Sitting lying standingand walking Accelerometer Mean and standard
Kose et al [45]Walking runningbiking sitting andstanding
AccelerometerMin max averagevariance FFT coefficientsand autocorrelation
None Clustered kNN 952ndash975
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
6 Computational and Mathematical Methods in Medicine
Table 2 Summary of feature selectiondimensionality reduction methods in HAR
Method Advantages Disadvantages Complexity
PCA
High dimensionalityreduction reduction of noiselack of redundancy of data dueto orthogonality ofcomponents
The covariance matrix is difficult to beevaluated accurately even the simplestinvariance could not be captured by thePCA unless the training data explicitlyprovides for it
119874(1199012
119899 + 1199013
) where 119899 are datapoints each represented with119901 features
ReliefF Low computationalcomplexity
Unstable due to random selection ofinstances 119874(119901 sdot 119899 sdot log 119899)
RankfeaturesFeatures highly correlatedwith already selected featuresare less likely to be included
It assumes that data classes are normallydistributed
It depends upon classseparability criterion
CFS It evaluates a subset of featuresrather than individual features
It fails to select locally predictive featureswhen they are overshadowed by strongglobally predictive features
119874(119899(
1199012
minus 119901
2
))
input sensor data to a smaller set of features (feature vector)which preserves information contained in the original dataThe dimensionality reduction step can be applied to removethe irrelevant (or less relevant) features and reduce thecomputational complexity and increase the performance ofthe activity recognition processThe classification step is usedto map the feature set to a set of activities
In this paper we do not focus on data preprocessing andfeature extraction but rather on dimensionality reduction andclassification steps since these two are crucial for furtherefficiency of AAL systems The proposed method for humanactivity recognition is based on feature dimensionality reduc-tion using random projections [36] and classification usingkernel density function estimate as a decision function(Figure 1)
32 Description of the Method During random projec-tion the original 119889-dimensional data is projected to a 119896-dimensional (119896 ≪ 119889) subspace using a random 119896 times 119889 matrix119877 The projection of the data onto a lower 119896-dimensionalsubspace is 119883119877119875
119896times119873= 119877119896times119889
119883119889times119873
where 119883119889times119873
is the originalset of 119873 119889-dimensional observations In the derived pro-jection the distances between the points are approximatelypreserved if points in a vector space are projected onto arandomly selected subspace of suitably high dimension (seethe Johnson-Lindenstrauss lemma [64]) The randommatrix119877 is selected as proposed by Achlioptas [36] as follows
119903119894119895=
+1 probability 16
0 probability 23
minus1 probability 16
(1)
Given the low dimensionality of the target space we cantreat the projection of low-dimensional observations ontoeach dimension as a set of random variables for which theprobability density function (PDF) can be estimated usingkernel density estimation (KDE) (or Parzenwindow)method[65]
If 1199091 1199092 119909
119873is a sample of a random variable then
the kernel density approximation of its probability densityfunction is
119891ℎ(119909) =
1
119873ℎ
119870(
119909 minus 119909119894
ℎ
) (2)
where 119870 is some kernel and ℎ is the bandwidth (smoothingparameter) 119870 is taken to be a standard Gaussian functionwith mean zero and variance 1 of the examined data features
119870 (119909) =
1
radic2120587
119890minus(12)119909
2
(3)
For a two-dimensional case the bivariate probabilitydensity function is calculated as a product of univariateprobability functions as follows
119891 (119909 119910) =
119891 (119909) sdot
119891 (119910) (4)
Here 119909 and 119910 are data in each dimension respectivelyHowever each random projection produces a different
mapping of the original data points which reveals only a partof the data manifold in higher-dimensional space In caseof the binary classification problem we are interested in amapping that separates data points belonging to two differentclasses best
As a criterion for estimating the mapping we use theJaccard distance metric between two probability density esti-mates of data points representing each class The advantageof the Jaccard distance metric as compared to other metricsof distance such as Kullback-Leibler (KL) divergence andHellinger distance is its adaptability to multidimensionalspaces where compared points show relations to differentsubsets Therefore it is well adapted to the developed modelof human activity features where according to description inthe previous section we have divided them into some sets ofactions Furthermore the computational complexity of theHellinger distance is very high while KL divergence mightbe unbounded
The Jaccard distance which measures dissimilaritybetween sample sets is obtained by subtracting the Jaccardcoefficient from 1 or equivalently by dividing the difference
Computational and Mathematical Methods in Medicine 7
Table 3 Summary of related works in the HAR domain
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Atallah et al[39]
Lying down preparingfood eating anddrinking socialisingreading dressingwalking treadmillwalking vacuumingwiping tables runningtreadmill runningcycling sittingdowngetting up andlying downgetting up
Accelerationsensors
Averaged entropy over 3axes main FFT frequency(averaged) over 3 axesenergy of the 02Hzwindow centred aroundmain frequency over totalFFT energy (3-axisaverage) and averagedmean of cross covariancebetween every 2 axes
ReliefF Simbaand MRMR
kNN Bayesianclassifier 90
Bayat et al [40]
Running slow walk fastwalk aerobic dancingstairs up and stairsdown
Triaxialaccelerometer
Mean along 119911-axisMinMax STD and RMSfor Am APF along 119909-axis119910-axis and 119911-axis VarAPFSTD along 119909-axis 119910-axisand 119911-axis RMS along119909-axis 119910-axis and 119911-axiscorrelation between 119911-axisand 119910-axis and MinMaxalong 119909-axis 119910-axis and119911-axis
Energy entropy meanvariance mean trendwindowed mean differencevariance trend windowedvariance differencedetrended fluctuationanalysis coefficients119883-119885-energy and maxdifference acceleration
ReliefF SFFS kNN NaiveBayes 98
Henpraserttaeet al [42]
Sitting lying standingand walking Accelerometer Mean and standard
Kose et al [45]Walking runningbiking sitting andstanding
AccelerometerMin max averagevariance FFT coefficientsand autocorrelation
None Clustered kNN 952ndash975
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
Computational and Mathematical Methods in Medicine 7
Table 3 Summary of related works in the HAR domain
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Atallah et al[39]
Lying down preparingfood eating anddrinking socialisingreading dressingwalking treadmillwalking vacuumingwiping tables runningtreadmill runningcycling sittingdowngetting up andlying downgetting up
Accelerationsensors
Averaged entropy over 3axes main FFT frequency(averaged) over 3 axesenergy of the 02Hzwindow centred aroundmain frequency over totalFFT energy (3-axisaverage) and averagedmean of cross covariancebetween every 2 axes
ReliefF Simbaand MRMR
kNN Bayesianclassifier 90
Bayat et al [40]
Running slow walk fastwalk aerobic dancingstairs up and stairsdown
Triaxialaccelerometer
Mean along 119911-axisMinMax STD and RMSfor Am APF along 119909-axis119910-axis and 119911-axis VarAPFSTD along 119909-axis 119910-axisand 119911-axis RMS along119909-axis 119910-axis and 119911-axiscorrelation between 119911-axisand 119910-axis and MinMaxalong 119909-axis 119910-axis and119911-axis
Energy entropy meanvariance mean trendwindowed mean differencevariance trend windowedvariance differencedetrended fluctuationanalysis coefficients119883-119885-energy and maxdifference acceleration
ReliefF SFFS kNN NaiveBayes 98
Henpraserttaeet al [42]
Sitting lying standingand walking Accelerometer Mean and standard
Kose et al [45]Walking runningbiking sitting andstanding
AccelerometerMin max averagevariance FFT coefficientsand autocorrelation
None Clustered kNN 952ndash975
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
8 Computational and Mathematical Methods in Medicine
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Kwapisz et al[46]
Walking jogging stairsupdown sitting andstanding
Accelerometer
Mean std dev averageabsolute difference averageresultant acceleration timebetween peaks and binneddistribution
None
Decision treelogisticregression andMNN
917
Lane et al [47]Driving beingstationary running andwalking
GPSaccelerometerand microphone
Mean variance None Naıve-Bayes 85ndash98
Lee and Cho[48]
Standing walkingrunning stairs updownshopping and taking bus
Accelerometer 119909- 119910- and 119911-axesacceleration values None Hierarchical
HMM 70ndash90
Mannini andSabatini [49]
Walking walkingcarrying items sitting amprelaxing working oncomputer standing stilleating or drinkingwatching TV readingrunning bicyclingstretching strengthtraining scrubbingvacuuming foldinglaundry lying down andrelaxing brushing teethclimbing stairs ridingelevator and ridingescalator
Accelerationsensors
DC component energyfrequency-domain entropyand correlation coefficients
SFFS (Pudilalgorithm)
ContinuousemissionsHidden MarkovModel
991
Mathie et al[37]
Various humanmovements includingresting walking andfalling
Triaxialaccelerationsensor
Integrated area under curve None Binary decisiontree
977(sensitivity)
987(specificity)
Maurer et al[50]
Walking standingsitting running andascending anddescending the stairs
Multiple sensors
Mean root mean squarestandard deviationvariance mean absolutedeviation cumulativehistogram 119899th percentilesinterquartile range zerocrossing rate meancrossing rate and sq lengthof119883 119884
Correlation-based FeatureSelection(CFS)
Decision trees(C45algorithm)119896-NearestNeighborNaıve-Bayesand Bayes Net
80ndash92
Miluzzo et al[51]
Sitting standingwalking and running
AccelerometerGPS and audio
DFT FFT features meanstd dev and number ofpeaks per unit and timedeviation of DFT power
None Decision tree 79
Parkka et al[52]
Lying down rowingex-bikingsittingstandingrunning and Nordicwalking
GPS audioaltitude EKGaccelerometercompasshumidity lighttemperatureheart rate pulserespiratoryeffort and skinresistance
Peak frequency of up-downchest acceleration medianof up-down chestacceleration peak power ofup-down chestacceleration variance ofback-forth chestacceleration sum ofvariances of 3D wristacceleration and powerratio of frequency bands1ndash15Hz and 02ndash5Hzmeasured from chestmagnetometer
Heuristic Decision tree 86
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
Computational and Mathematical Methods in Medicine 9
Table 3 Continued
Author Activities Sensor data Features Featureselection
Classificationmethod Accuracy
Saponas et al[53] Walking jogging Accelerometer
124 features Nike + iPodPacket Payload magnitude(mean std dev minmax and min minusmax) frequency (energy ineach of the first 10frequency components ofDFT energy in each bandof 10 frequencycomponents largestfrequency component andindex of the largestfrequency component)
Magnitude std meanmin max percentiles (1025 50 75 and 90) and sumand square sum ofobservations abovebelowpercentile (5 10 25 75 90and 95) of magnitudeacceleration and squaresum of 119909 amp 119911
None Decision tree +kNNQDA 95
Sohn et al [55] Walking driving anddwelling GPS
Spearman rank correlationvariance and meanEuclidean distance over awindow of measurements
None Logisticregression 85
Yang [56]Sitting Standingwalking runningdriving and bicycling
Accelerometer
Mean std zero crossingrate 75th percentileinterquartile spectrumcentroid entropy andcross-correlation
3D acceleration Mean variance None Neural networkensemble 67ndash98
Data acquisition Denoising Feature
generation
Learning Best projection
Predictionusing KDE Activity
Figure 1 General scheme of the proposed method
of the sizes of the union and the intersection of two sets bythe size of the union
119889119869(119860 119861) = 1 minus 119869 (119860 119861) =
|119860 cup 119861| minus |119860 cap 119861|
|119860 cup 119861|
(5)
In the proposed model the best random projection withthe smallest overlapping area is selected (see an example inFigure 2)
To explore the performance and correlation amongfeatures visually a series of scatter plots in a 2D featurespace is shown in Figure 3 The horizontal and verticalaxes represent two different features The points in differentcolours represent different human activities
In case of multiple classes the method works as a one-class classifier recognizing instances of a positive class whileall instances of other classes are recognized as outliers of thepositive class
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
10 Computational and Mathematical Methods in Medicine
minus200 0 200 400 600 800 1000 1200Feature value
0
001
002
003
004
005
006
007
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
minus1200 minus1000 minus800 minus600 minus400 minus200 0 200 400Feature value
0
1
2
3
4
5
6
7
8
9
Prob
abili
ty d
ensit
y
Walking ForwardWalking Upstairs
times10minus3
Figure 2 Graphical illustration of good separation versus bad separation of kernel density estimation functions (Subject 1 Trial 1 WalkingForward versus Walking Upstairs 2nd dimension)
Figure 3 Example of classification walking versus running (Subject 1 Trial 1) classes randomly projected in a bidimensional feature subspace
33 Algorithm Thepseudocode of the algorithms for findingthe best projection and using it for classification in low-dimensional space is presented in Pseudocodes 1 and 2respectively
4 Experiments
41 Dataset To evaluate the performance of the proposedapproach for HAR from the smartphone data we used thepart of the dataset (USC Human Activity Dataset [38])recorded using the MotionNode device (sampling rate
100Hz 3-axis accelerometer range plusmn6 g 3-axis gyroscoperange plusmn500 dps) The dataset consists of records recordedwith 14 subjects (7 male 7 female age 21ndash49) of 12 activities5 trials each During data acquisition MotionNode wasattached on the front right hip of subjects
The recorded low-level activities are as follows WalkingForward (WF) Walking Left (WL) Walking Right (WR)Walking Upstairs (WU) Walking Downstairs (WD) Run-ning Forward (RF) Jumping Up (JU) Sitting (Si) Standing(St) Sleeping (Sl) Elevator Up (EU) and Elevator Down(ED) Each record consists of the following attributes date
Computational and Mathematical Methods in Medicine 11
ALGORITHM FindBestProjection
INPUT data1 data2 ndash data for class1 and class2 [nxm matrices]
threshold ndash iterating parameter
OUTPUT bestProjection
BEGIN
Jaccard = MAXINT
WHILE (Jaccard gt threshold)
generate Random Projection matrix projectionproject m-dimensional data1 amp data2 into 2D pdata1 amp pdata2FOREACH dimension of pdata1 amp pdata2
calculate kernel density distributions
calculate Jaccard intersection of pdata1 amp pdata2
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
IF (fx1(pSample) fy1(pSample) gt fx2(pSample) fy2(pSample))
LET classLabel = +1
ELSE
LET classLabel = -1
END
RETURN classLabel
END
Pseudocode 2 Pseudocode of binary classification
subject number age height weight activity name activitynumber trial number sensor location orientation and read-ings Sensor readings consist of 6 readings acceleration along119909- 119910- and 119911-axes and gyroscope along 119909- 119910- and 119911-axesEach trial was performed on different days at various indoorand outdoor locations
42 Results In Table 4 we describe the top three best featuresfrom Table 1 (see column Feature number) ranked by theMatlab Rankfeatures function using the entropy criterion
The results of feature ranking presented in Table 5 can besummarized as follows
(i) For Walking Forward Walking Left and WalkingRight the important features are moving variance ofacceleration and gyroscope data movement intensityof gyroscope data moving variance of movementintensity of acceleration data first eigenvalue of mov-ing covariance between acceleration data and polarangle of moving cumulative sum of gyroscope data
(ii) For Walking Upstairs andWalking Downstairs mov-ing variance of gyroscope along 119911-axis movementintensity of gyroscope data and moving variance ofmovement intensity are the most important
(iii) For Running Forward moving variance of 100 sam-ples of acceleration along 119909-axis moving variance of100 samples of gyroscope along 119911-axis and movingenergy of acceleration are distinguishing features
(iv) For Jumping Up the most important features aremoving variance of acceleration moving variance ofmovement intensity and moving energy of accelera-tion
(v) For Sittingmovement intensity of gyroscope data andmovement intensity of difference between accelera-tion and gyroscope data are the most important
(vi) For Standingmoving variance ofmovement intensityof acceleration data moving variance of accelera-tion along 119909-axis and first eigenvalue of moving
12 Computational and Mathematical Methods in Medicine
Table 4 Top features for binary classification of human activities
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
covariance of difference between acceleration andgyroscope data are the most distinctive
(vii) For Sleeping the most prominent features are firsteigenvalue of moving covariance between accelera-tion data andmoving variance of movement intensityof acceleration data
(viii) For Elevator Up and Elevator Down the most com-monly selected feature is moving variance of 119911-axisof gyroscope data Other prominent features arefirst eigenvalue of moving covariance of differencebetween acceleration and gyroscope data andmovingenergy of 119911-axis of gyroscope data
These results can be considered as consistent from whatcan be expected from the physical analysis of humanmotionsin the analysed dataset
The evaluation of HAR classification algorithms is usuallymade through the statistical analysis of the models using the
available experimental dataThemost commonmethod is theconfusion matrix which allows representing the algorithmperformance by clearly identifying the types of errors (falsepositives and negatives) and correctly predicted samples overthe test data
The confusion matrix for within-subject activity recogni-tion using Matlabrsquos Rankfeatures is detailed in Table 5 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for test-ing Grand mean accuracy is 09552 grand mean precisionis 09670 grand mean sensitivity is 09482 grand meanspecificity is 09569 grand mean recall is 09482 grandmean 119865-score is 09482The baseline accuracy was calculatedusing only the top 2 features selected by Rankfeatures butwithout using random projections The results show thatfeatures derived using random projections are significantlybetter than features derived using a common feature selectionalgorithm
Computational and Mathematical Methods in Medicine 13
Table 6 The confusion matrix of within-subject activity classification using ReliefF
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
To take a closer look at the classification result Table 5shows the confusion table for classification of activities Theoverall averaged recognition accuracy across all activities is9552 with 11 out of 12 activities having accuracy valueshigher than 90 If we examine the recognition performancefor each activity individually Running Forward JumpingUp and Sleeping will have very high accuracy values ForRunning Forward the accuracy of 990 is achieved Inter-estingly the lowest accuracy was achieved for Elevator Upactivity only 840 while it was most often misclassifiedwith Sitting and Standing Elevator Down is misclassifiedwith Elevator Up (only 697 accuracy) This result makessense since Sitting on a chair Standing and Standing in amoving elevator are static activities and we expect difficultyin differentiating different static activities Also there is somemisclassification when deciding on a specific direction ofactivity for example Walking Left is confused with WalkingForward (774 accuracy) andWalkingUpstairs (874 accu-racy) Walking Upstairs is also confused with Walking Right(798 accuracy) andWalking Downstairs (708 accuracy)This is due to the similarity of any walk-related activities
For comparison the confusion matrix for within-subjectactivity recognition obtained using the proposed methodwith ReliefF feature selection is detailed in Table 6 Theclassification was performed using 5-fold cross-validationusing 80 of data for training and 20 of data for testingGrandmean accuracy is 0932 grandmean precision is 0944grand mean sensitivity is 0939 grand mean specificity is0933 grand mean recall is 0939 grand mean 119865-score is0922
The baseline accuracy was calculated using only thetop 2 features selected using ReliefF but without usingrandom projections Again the results show that featuresderived using random projections are significantly betterthan features derived using the ReliefF method only
Surprisingly though the classification accuracy of thespecific activities differed the mean accuracy metric results
are quite similar (but still worse if grand mean values areconsidered) The features identified using ReliefF featureselection were better at separating Walking Forward fromWalking Left and Standing from Elevator Up activities butproved worse for separating other activities such as Sittingfrom Standing
For subject identification the data from all physicalactions is used to train the classifier Here we consider one-versus-all subject identification problem Therefore the dataof one subject is defined as positive class and the data of allother subjects is defined as negative class In this case also5-fold cross-validation was performed using 80 of data fortraining and 20of data for testingThe results of one-versus-all subject identification using all activities for training andtesting are presented in Table 7While the results are not verygood they still are better than random baselines grandmeanaccuracy is 0477 precision is 0125 recall is 0832 and 119865-score is 0210
If an activity of a subject has been established separateclassifiers for each activity can be used for subject identifica-tion In this case also 5-fold cross-validation was performedusing 80 of data for training and 20 of data for testingand the results are presented in Table 8 The grand meanaccuracy is 0720 which is better than random baselineHowever if we consider only the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the mean accuracy is 0944
Finally we can simplify the classification problem tobinary classification (ie recognize one subject againstanother)This simplification can bemotivated by the assump-tion that only a few people are living in an AAL home (farless than 14 subjects in the analysed dataset) Then the datafrom a pair of subjects performing a specific activity is usedfor classification and training Separate classifiers are built foreach pair of subjects the results are evaluated using 5-foldcross-validation and the results are averaged The results arepresented in Table 9 Note that the grand mean accuracy has
14 Computational and Mathematical Methods in Medicine
Table 7 Results of one-versus-all subject identification (all activi-ties)
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
Table 8 Results of one-versus-all subject identification for specificactivities
Activity Accuracy 119865
Walking Forward 0947 0727Walking Left 0955 0769Walking Right 0931 0722Walking Upstairs 0857 0551Walking Downstairs 0833 0497Running Forward 0832 0496Jumping Up 0814 0453Sitting 0506 0391Standing 0722 0432Sleeping 0589 0292Elevator Up 0337 0235Elevator Down 0318 0232Mean 0720 0483Random baseline 0071 0133
increased to 0947 while for the top three walking-relatedactivities (Walking ForwardWalking Left orWalkingRight)the grand mean accuracy is 0992
5 Evaluation and Discussion
Random projections have been used in the HAR domain fordata dimensionality reduction in activity recognition fromnoisy videos [69] feature compression for head pose estima-tion [70] and feature selection for activity motif discovery[71] The advantages of random projections are the simplicityof their implementation and their scalability robustness tonoise and low computational complexity constructing therandom matrix 119877 and projecting the 119889 times 119873 data matrix into119896 dimensions are of order 119874(119889119896119873)
Table 9 Accuracy of binary subject identification using separateactivities
Activity Accuracy 119865
Walking Forward 0992 0987Walking Left 0989 0987Walking Right 0993 0993Walking Upstairs 0977 0970Walking Downstairs 0974 0971Running Forward 0980 0974Jumping Up 0983 0980Sitting 0883 0859Standing 0940 0932Sleeping 0956 0953Elevator Up 0856 0847Elevator Down 0846 0822Mean 0947 0939Random baseline 05 05
The HAD dataset has been used in HAR research byother authors too Using the same HAD dataset Zheng[66] has achieved 956 accuracy He used the means andvariances of magnitude and angles as the activity featuresand the magnitude and angles that were produced by atriaxial acceleration vector Classifier used the Least SquaresSupport Vector Machine (LS-SVM) and Naıve-Bayes (NB)algorithm to distinguish different activity classes Sivaku-mar [67] achieved 843 overall accuracy using symbolicapproximation of time series of accelerometer and gyro-scope signal Vaka [68] achieved 907 accuracy for within-person classification and 886 accuracy for interpersonclassification using Random Forest The features used forthe recognition were time domain features mean standarddeviation correlation between119883 and 119884 correlation between119884 and 119885 correlation between119883 and119885 and rootmean squareof a signal Our results (9552 accuracy) obtained using theproposedmethod are very similar to the best results of Zhengfor activity recognition task
The results obtained by different authors using the USC-HAD dataset are summarized in Table 10
We think that it would be difficult to achieve even higherresults due to someproblemswith the analysed dataset whichinclude a set of problems inherent to many Human ActivityDatasets as follows
(i) Accurate Labelling of All Activities Existing activityrecognition algorithms usually are based on super-vised learning where the training data depends uponaccurate labelling of all human activities Collectingconsistent and reliable data is a very difficult task sincesome activities may have been marked by users withwrong labels
(ii) TransitionaryOverlapping ActivitiesOften people doseveral activities at the same time The transitionstates (such as walking-standing lying-standing) canbe treated as additional states and the recognition
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
Computational and Mathematical Methods in Medicine 15
Table 10 Summary of HAR results using USC-HAD dataset
Reference Features Classification method Accuracy
Zheng [66] Means and variances of magnitude andangles of acceleration along 119909- 119910- amp 119911-axes
Least Squares Support Vector Machine(LS-SVM) Naıve-Bayes (NB) 956
Sivakumar [67] Accelerometer and gyroscope data Symbolic approximation 843
Vaka [68] Mean std dev correlation between119883 amp 119884119884 amp 119885 and119883 amp 119885 and RMS Random Forest 907
This paper 99 times frequency and physical features Heuristic (random projections + PDFs +Jaccard distance) 9552
model can be trained with respect to these states toincrease the accuracy
(iii) Context Problem It occurs when the sensors areplaced at an inappropriate position relative tothe activity being measured For example withaccelerometer-based HAR the location where thephone is carried such as in the pocket or in the bagimpacts the classification performance
(iv) Subject Sensitivity It measures dependency of thetrained classificationmodel upon the specifics of user
(v) Weak Link between Basic Activities andMore ComplexActivities For example it is rather straightforwardto detect whether the user is running but inferringwhether the user is running away from danger orjogging in a park is different
(vi) Spurious Data Most published studies handle theproblem of the fuzzy borders by manual data crop-ping
6 Conclusion
Monitoring and recognizing human activities are importantfor assessing changes in physical and behavioural profilesof the population over time particularly for the elderly andimpaired and patients with chronic diseases Although a widevariety of sensors are being used in various devices for activitymonitoring the positioning of the sensors the selection ofrelevant features for different activity groups and providingcontext to sensormeasurements still pose significant researchchallenges
In this paper we have reviewed the stages needed toimplement a human activity recognition method for auto-matic classification of human physical activity from on-bodysensors A major contribution of the paper lies in pursuingthe random projections based approach for feature dimen-sionality reductionThe results of extensive testing performedon the USC-HAD dataset (we have achieved overall accuracyof within-person classification of 9552 and interpersonidentification accuracy of 9475) reveal the advantagesof the proposed approach Gait-related activities (WalkingForward Walking Left and Walking Right) allowed the bestidentification of subjects opening the way for a multitudeof applications in the area of gait-based identification andverification
Future work will concern the validation of the proposedmethod using other datasets of human activity data as wellas integration of the proposed method in the wearable sensorsystemwe are currently developing for applications in indoorhuman monitoring
Competing Interests
The authors declare that there are no competing interestsregarding the publication of this paper
Acknowledgments
The authors would like to acknowledge the contributionof the COST Action IC1303 AAPELE Architectures Algo-rithms and Platforms for Enhanced Living Environments
References
[1] U S State Department and National Institute on Aging (NIA)Why Population Aging Matters A Global Perspective 2007
[2] Department of Economic and Social Affairs and PopulationDivisionWorld Population to 2300 United Nations New YorkNY USA 2004
[3] P Turaga R Chellappa V S Subrahmanian and O UdrealdquoMachine recognition of human activities a surveyrdquo IEEETransactions on Circuits and Systems for Video Technology vol18 no 11 pp 1473ndash1488 2008
[4] R Poppe ldquoA survey on vision-based human action recognitionrdquoImage and Vision Computing vol 28 no 6 pp 976ndash990 2010
[5] J K Aggarwal and M S Ryoo ldquoHuman activity analysis areviewrdquo ACM Computing Surveys vol 43 no 3 article 16 2011
[6] S-R Ke H L UThuc Y-J Lee J-N Hwang J-H Yoo and K-H Choi ldquoA review on video-based human activity recognitionrdquoComputers vol 2 no 2 pp 88ndash131 2013
[7] O C Ann ldquoHuman activity recognition a reviewrdquo in Proceed-ings of the of IEEE International Conference on Control SystemComputing and Engineering (ICCSCE rsquo14) pp 389ndash393 BatuFerringhi Malaysia November 2014
[8] J K Aggarwal and L Xia ldquoHuman activity recognition from 3Ddata a reviewrdquo Pattern Recognition Letters vol 48 pp 70ndash802014
[9] M Vrigkas C Nikou and I A Kakadiaris ldquoA review of humanactivity recognition methodsrdquo Frontiers in Robotics and AI vol2 article 28 2015
[10] M Ziaeefard and R Bergevin ldquoSemantic human activity recog-nition a literature reviewrdquo Pattern Recognition vol 48 no 8pp 2329ndash2345 2015
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
16 Computational and Mathematical Methods in Medicine
[11] O D Incel M Kose and C Ersoy ldquoA review and taxonomy ofactivity recognition on mobile phonesrdquo BioNanoScience vol 3no 2 pp 145ndash171 2013
[12] V Osmani S Balasubramaniam and D Botvich ldquoHumanactivity recognition in pervasive health-care supporting effi-cient remote collaborationrdquo Journal of Network and ComputerApplications vol 31 no 4 pp 628ndash655 2008
[13] O D Lara and M A Labrador ldquoA survey on human activityrecognition using wearable sensorsrdquo IEEE CommunicationsSurveys and Tutorials vol 15 no 3 pp 1192ndash1209 2013
[14] M Shoaib S Bosch O D Incel H Scholten and P J MHavinga ldquoA survey of online activity recognition using mobilephonesrdquo Sensors vol 15 no 1 pp 2059ndash2085 2015
[15] J L R Ortiz Smartphone-Based Human Activity RecognitionSpringer Theses 2015
[16] S Purpura V Schwanda K Williams W Stubler and PSengers ldquoFit4Life The design of a persuasive technology pro-moting healthy behavior and ideal weightrdquo in Proceedings of the29th Annual CHI Conference on Human Factors in ComputingSystems (CHI rsquo11) pp 423ndash432 ACM New York NY USAMay2011
[17] N A Capela E D Lemaire and N Baddour ldquoFeature selectionfor wearable smartphone-based human activity recognitionwith able bodied elderly and stroke patientsrdquo PLoS ONE vol10 no 4 Article ID e0124414 2015
[18] S Patel RHughes THester et al ldquoAnovel approach tomonitorrehabilitation outcomes in stroke survivors using wearabletechnologyrdquo Proceedings of the IEEE vol 98 no 3 pp 450ndash4612010
[19] W Tao T Liu R Zheng and H Feng ldquoGait analysis usingwearable sensorsrdquo Sensors vol 12 no 2 pp 2255ndash2283 2012
[20] G Appelboom B E Taylor E Bruce et al ldquoMobile phone-connected wearable motion sensors to assess postoperativemobilizationrdquo JMIR mHealth and uHealth vol 3 no 3 articlee78 2015
[21] V H Cheung L Gray and M Karunanithi ldquoReview ofaccelerometry for determining daily activity among elderlypatientsrdquo Archives of Physical Medicine and Rehabilitation vol92 no 6 pp 998ndash1014 2011
[22] N Bidargaddi A Sarela L Klingbeil and M KarunanithildquoDetecting walking activity in cardiac rehabilitation by usingaccelerometerrdquo in Proceedings of the International Conferenceon Intelligent Sensors Sensor Networks and Information (ISSNIPrsquo07) pp 555ndash560 December 2007
[23] L Bao and S S Intille ldquoActivity recognition from user-annotated acceleration datardquo in Pervasive Computing A Fer-scha and FMattern Eds vol 3001 of Lecture Notes in ComputerScience pp 1ndash17 Springer Berlin Germany 2004
[24] A Avci S Bosch M Marin-Perianu R Marin-Perianu andP Havinga ldquoActivity recognition using inertial sensing forhealthcare wellbeing and sports applications a surveyrdquo inProceedings of the 23rd International Conference on Architectureof Computing Systems (ARCS 10) pp 1ndash10 Hannover GermanyFebruary 2010
[25] K S Gayathri S Elias and B Ravindran ldquoHierarchical activityrecognition for dementia care using Markov Logic NetworkrdquoPersonal and Ubiquitous Computing vol 19 no 2 pp 271ndash2852015
[26] C Phua P C Roy H Aloulou et al ldquoState-of-the-art assistivetechnology for people with dementiardquo inHandbook of Researchon Ambient Intelligence and Smart Environments Trends and
Perspectives N-Y Chong and F Mastrogiovanni Eds chapter16 pp 300ndash319 IGI Global 2011
[27] P C Roy S Giroux B Bouchard et al ldquoA possibilisticapproach for activity recognition in smart homes for cognitiveassistance to Alzheimerrsquos patientsrdquo in Activity Recognition inPervasive Intelligent Environments vol 4 of Atlantis Ambientand Pervasive Intelligence pp 33ndash58 Springer Berlin Germany2011
[28] A Hildeman Classification of epileptic seizures using accelerom-eters [PhD dissertation] Chalmers University of TechnologyGothenburg Sweden 2011
[29] M R Puyau A L Adolph F A Vohra I Zakeri andN F ButteldquoPrediction of activity energy expenditure using accelerometersin childrenrdquoMedicine and Science in Sports and Exercise vol 36no 9 pp 1625ndash1631 2004
[30] P Gupta and T Dallas ldquoFeature selection and activity recog-nition system using a single triaxial accelerometerrdquo IEEETransactions on Biomedical Engineering vol 61 no 6 pp 1780ndash1786 2014
[31] A T Ozdemir and B Barshan ldquoDetecting falls with wearablesensors usingmachine learning techniquesrdquo Sensors vol 14 no6 pp 10691ndash10708 2014
[32] M Arif and A Kattan ldquoPhysical activities monitoring usingwearable acceleration sensors attached to the bodyrdquo PLoS ONEvol 10 no 7 Article ID e0130851 2015
[33] GGAlvarez andNTAyas ldquoThe impact of daily sleep durationon health a review of the literaturerdquo Progress in CardiovascularNursing vol 19 no 2 pp 56ndash59 2004
[34] S E Crouter J R Churilla and D R Bassett Jr ldquoEstimatingenergy expenditure using accelerometersrdquo European Journal ofApplied Physiology vol 98 no 6 pp 601ndash612 2006
[35] H Yu M Spenko and S Dubowsky ldquoAn adaptive sharedcontrol system for an intelligent mobility aid for the elderlyrdquoAutonomous Robots vol 15 no 1 pp 53ndash66 2003
[36] D Achlioptas ldquoDatabase-friendly random projectionsrdquo in Pro-ceedings of the 20th ACM SIGMOD-SIGACT-SIGART Sympo-sium on Principles of Database Systems (PODS rsquo01) pp 274ndash281Santa Barbara Calif USA May 2001
[37] M JMathie BGCellerNH Lovell andAC F Coster ldquoClas-sification of basic daily movements using a triaxial accelerome-terrdquoMedical and Biological Engineering and Computing vol 42no 5 pp 679ndash687 2004
[38] M Zhang and A A Sawchuk ldquoUSC-HAD a daily activitydataset for ubiquitous activity recognition using wearablesensorsrdquo in Proceedings of the International Conference onUbiquitous Computing (UbiComp rsquo12) pp 1036ndash1043 ACMSeptember 2012
[39] L Atallah B Lo R King and G-Z Yang ldquoSensor positioningfor activity recognition using wearable accelerometersrdquo IEEETransactions on Biomedical Circuits and Systems vol 5 no 4pp 320ndash329 2011
[40] A BayatM Pomplun andDA Tran ldquoA study on human activ-ity recognition using accelerometer data from smartphonesrdquoProcedia Computer Science vol 34 pp 450ndash457 2014
[41] M Berchtold M Budde D Gordon H R Schmidtke and MBeigl ldquoActiServ activity recognition service formobile phonesrdquoin Proceedings of the 14th IEEE International Symposium onWearable Computers (ISWC rsquo10) pp 1ndash8 IEEE Seoul SouthKorea October 2010
[42] A Henpraserttae S Thiemjarus and S Marukatat ldquoAccurateactivity recognition using a mobile phone regardless of device
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
Computational and Mathematical Methods in Medicine 17
orientation and locationrdquo in Proceedings of the InternationalConference on Body Sensor Networks (BSN rsquo11) pp 41ndash46 IEEEDallas Tex USA May 2011
[43] E Hoque and J Stankovic ldquoAALO activity recognition in smarthomes using Active Learning in the presence of Overlappedactivitiesrdquo in Proceedings of the 6th International Conference onPervasive Computing Technologies for Healthcare andWorkshops(PervasiveHealth rsquo12) pp 139ndash146 May 2012
[44] T Iso and K Yamazaki ldquoGait analyzer based on a cell phonewith a single three-axis accelerometerrdquo in Proceedings of the8th International Conference on Human-Computer Interactionwith Mobile Devices and Services (MobileHCI rsquo06) pp 141ndash144Espoo Finland September 2006
[45] M Kose O D Incel and C Ersoy ldquoOnline human activityrecognition on smart phonesrdquo in Proceedings of the WorkshoponMobile Sensing From Smartphones andWearables to BigData(Colocated with IPSN) pp 11ndash15 Beijing China April 2012
[46] J R Kwapisz G M Weiss and S A Moore ldquoActivityrecognition using cell phone accelerometersrdquo ACM SIGKDDExplorations Newsletter vol 12 no 2 pp 74ndash82 2011
[47] N Lane M Mohammod M Lin et al ldquoBewell a smartphoneapplication to monitor model and promote wellbeingrdquo in Pro-ceedings of the 5th International ICST Conference on PervasiveComputing Technologies for Healthcare pp 23ndash26 IEEE 2012
[48] Y S Lee and S Cho ldquoActivity recognition using hierarchicalhidden markov models on a smartphone with 3d accelerom-eterrdquo in Hybrid Artificial Intelligent Systems 6th InternationalConference HAIS 2011 Wroclaw Poland May 23ndash25 2011Proceedings Part I vol 6678 of Lecture Notes in ComputerScience pp 460ndash467 Springer Berlin Germany 2011
[49] A Mannini and A M Sabatini ldquoMachine learning methods forclassifying human physical activity from on-body accelerome-tersrdquo Sensors vol 10 no 2 pp 1154ndash1175 2010
[50] U Maurer A Smailagic D P Siewiorek and M DeisherldquoActivity recognition and monitoring using multiple sensorson different body positionsrdquo in Proceedings of the InternationalWorkshop on Wearable and Implantable Body Sensor Networks(BSN rsquo06) pp 113ndash116 IEEE Cambridge Mass USA April2006
[51] E Miluzzo N D Lane K Fodor et al ldquoSensing meets mobilesocial networks the design implementation and evaluationof the CenceMe applicationrdquo in Proceedings of the 6th ACMConference on Embedded Networked Sensor Systems (SenSysrsquo08) pp 337ndash350 Raleigh NC USA November 2008
[52] J Parkka M Ermes P Korpipaa J Mantyjarvi J Peltola andI Korhonen ldquoActivity classification using realistic data fromwearable sensorsrdquo IEEETransactions on Information Technologyin Biomedicine vol 10 no 1 pp 119ndash128 2006
[53] T S Saponas J Lester J Froehlich J Fogarty and J LandayiLearn on the iPhone Real-Time Human Activity Classificationon Commodity Mobile Phones 2008
[54] P Siirtola and J Roning ldquoRecognizing human activities user-independently on smartphones based on accelerometer datardquoInternational Journal of Interactive Multimedia and ArtificialIntelligence vol 1 no 5 pp 38ndash45 2012
[55] T Sohn A Varshavsky A LaMarca et al ldquoMobility detec-tion using everyday GSM tracesrdquo in UbiComp 2006 Ubiqui-tous Computing 8th International Conference UbiComp 2006Orange County CA USA September 17ndash21 2006 Proceedingsvol 4206 of Lecture Notes in Computer Science pp 212ndash224Springer Berlin Germany 2006
[56] J Yang ldquoToward physical activity diary motion recognitionusing simple acceleration features with mobile phonesrdquo Pro-ceedings of the 1st International Workshop on Interactive Multi-media for Consumer Electronics (IMCE rsquo09) 2009
[57] C Zhu and W Sheng ldquoMotion- and location-based onlinehuman daily activity recognitionrdquo Pervasive and Mobile Com-puting vol 7 no 2 pp 256ndash269 2011
[58] H Liu and L Yu ldquoToward integrating feature selection algo-rithms for classification and clusteringrdquo IEEE Transactions onKnowledge and Data Engineering vol 17 no 4 pp 491ndash5022005
[59] M Robnik-Sikonja and I Kononenko ldquoTheoretical and empir-ical analysis of ReliefF and RReliefFrdquoMachine Learning vol 53no 1-2 pp 23ndash69 2003
[60] M A Hall Correlation-based feature selection for machinelearning [PhD thesis] The University of Waikato 1999
[61] A Jain and D Zongker ldquoFeature selection evaluation appli-cation and small sample performancerdquo IEEE Transactions onPattern Analysis andMachine Intelligence vol 19 no 2 pp 153ndash158 1997
[62] I Guyon and A Elisseeff ldquoAn introduction to variable andfeature selectionrdquo Journal of Machine Learning Research vol 3pp 1157ndash1182 2003
[63] S Pirttikangas K Fujinami andTNakajima ldquoFeature selectionand activity recognition from wearable sensorsrdquo in UbiquitousComputing Systems H Y Youn M Kim and H MorikawaEds vol 4239 of Lecture Notes in Computer Science pp 516ndash527 Springer Berlin Germany 2006
[64] W B Johnson and J Lindenstrauss ldquoExtensions of Lipshitzmapping into Hilbert spacerdquo in Conference in Modern Analysisand Probability vol 26 of Contemporary Mathematics pp 189ndash206 American Mathematical Society 1984
[65] E Parzen ldquoOn estimation of a probability density function andmoderdquo The Annals of Mathematical Statistics vol 33 no 3 pp1065ndash1076 1962
[66] Y Zheng ldquoHuman activity recognition based on the hierar-chical feature selection and classification frameworkrdquo Journalof Electrical and Computer Engineering vol 2015 Article ID140820 9 pages 2015
[67] A Sivakumar Geometry aware compressive analysis of humanactivities application in a smart phone platform [MS thesis]Arizona State University Tempe Ariz USA 2014
[68] P R Vaka A pervasive middleware for activity recognition withsmartphones [MS thesis] University of Missouri-Kansas 2015
[69] D Tran and A Sorokin ldquoHuman activity recognition withmetric learningrdquo in Computer VisionmdashECCV 2008 D ForsythP Torr and A Zisserman Eds vol 5302 of Lecture Notesin Computer Science pp 548ndash561 Springer Berlin Germany2008
[70] D Lee M-H Yang and S Oh ldquoFast and accurate head poseestimation via random projection forestsrdquo in Proceedings of theIEEE International Conference on Computer Vision (ICCV rsquo15)pp 1958ndash1966 IEEE Santiago Chile December 2015
[71] L Zhao X Wang G Sukthankar and R Sukthankar ldquoMotifdiscovery and feature selection for CRF-based activity recog-nitionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 3826ndash3829 Istanbul TurkeyAugust 2010
