Novel Method for Feature-set Ranking Applied to Physical Activity Recognition

Post on 15-Jun-2015

107 views 1 download

Tags:

description

The benefits arising from proactive conduct and subject-specialized healthcare have driven e-health and e-monitoring into the forefront of research, in which the recognition of motion, postures and physical exercise is one of the main subjects. We propose here a multidisciplinary method for the recognition of physical activity with the emphasis on feature extraction and selection processes, which are considered to be the most critical stages in identifying the main unknown activity discriminant elements. Efficient feature selection processes are particularly necessary when dealing with huge training datasets in a multidimensional space, where conventional feature selection procedures based on wrapper methods or ‘branch and bound’ are highly expensive in computational terms. We propose an alternative filter method using a feature quality group ranking via a couple of two statistical criteria. Satisfactory results are achieved in both laboratory and semi-naturalistic activity living datasets for real problems using several classification models, thus proving that any body sensor location can be suitable to define a simple one feature-based recognition system, with particularly remarkable accuracy and applicability in the case of the wrist. This presentation illustrates part of the work described in the following articles: * Banos, O., Damas, M., Pomares, H., Prieto, A., Rojas, I.: Daily Living Activity Recognition based on Statistical Feature Quality Group Selection. Expert Systems with Applications, vol. 39, no. 9, pp. 8013-8021 (2012) * Banos, O., Pomares, H., Rojas, I.: Ambient Living Activity Recognition based on Feature-set Ranking Using Intelligent Systems. In: Proceedings of the 2010 International Joint Conference on Neural Networks (IJCNN 2010), IEEE, Barcelona, July 18-23, (2010)

transcript

Novel Method for Feature-set

Ranking Applied to Physical Activity

Recognition

IEA-AIE 2010

Córdoba (SPAIN)

O. Baños, H. Pomares, I. Rojas

Health Sector Today

• Innovations in Technology and Globalization have transformed health services

• Medical interventions have changed from “direct and specific person treatment” to “continuous and spatio-independent interaction”

2

• Acute diseases have evolved to chronic diseases, while World population is becoming older

AmiVital Project

• Create an integral and consistent approach for the provision of AmI (Ambient Intelligence) services to citizens, from both a social and health care perspective

3

• Merge concepts from the AmI paradigm and the current framework for health assistance into a more general and integral model of services

Activity Recognition

• Fundamental part of medical/health assistant work, being applicable to other areas (sport efficiency, videogames industry, robotics, etc.)

• Changeableness due to capability for discovering and identifying actions, movements and gestures than normally are unnoticed

• Objectives

4

Define an original methodology Identify the main characteristics Improve results in unsupervised monitoring studies

Experimental setup • Five accelerometers

Walking Sitting and relaxing Standing still Running

5

• Four activities

• Twenty subjects

• Two monitoring methodologies

Data preprocessing

• Different approximations were studied

• Best results “a posteriori” using a LPF+HPF (IIR elliptic)

6

ORIGINAL MEAN FILTERING LPF+HPF

Feature extraction

Magnitudes

Amplitude Autocorrelation Cepstrum Correlation lags Cross correlation Energy Spectral Density Spectral coherence Spectrum amplitude/phase Histogram Historical data lags Minimum phase reconstruction Wavelet decomposition

Statistical functions

4th and 5th central statistical moments Energy Arithmetic/Harmonic/Geometric/ Trimmed mean Entropy Fisher asymmetry coefficient Maximum / Position of Median Minimum / Position of Mode Kurtosis Data range Standard deviation Total harmonic deviation Variance Zero crossing counts

7

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 21

1.5

2

2.5

3

3.5

4

Walking

Sitting and relaxing

Standing still

Running

Why feature selection is needed?

• Influence on classification process

OPTIMUM

Few Features Good classification

0 500 1000-1

-0.5

0

0.5

1x 10

4 Thigh accelerometer

Features

Fe

atu

re v

alu

e8

• Huge feature set (861 parameters 2861 1.5 x 10259 possible combinations)

Feature selection

0

5

10

15

20

25

30

Wavelet coef. (a5) geometric mean

Fe

atu

re v

alu

e

Discriminant

capacity Robustness

Quality

group

4 5 1

4 4 2

4 3 3

4 2 4

4 1 5

3 5 6

3 4 7

3 3 8

3 2 9

3 1 10

2 5 11

2 4 12

2 3 13

2 2 14

2 1 15

1 5 16

1 4 17

1 3 18

1 2 19

1 1 20

0 5 21

Overlapping criteria

Robustness criteria

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 21

1.5

2

2.5

3

3.5

4

Walking

Sitting and relaxing

Standing still

Running

9

Feature selection

0 0.2 0.4 0.6 0.8 10

200

400

600

800

1000

Overlapping Threshold

No

. D

iscri

min

an

t F

ea

ture

s

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

100

200

300

400

500

600

700

800

900

Overlapping Threshold

No

. D

iscri

min

an

t F

ea

ture

s

Walking

Sitting and relaxing

Standing still

Running

All activities

All activities & all accelerometers

10

• Features extracted from the complete signal • Data corresponding to hip accelerometer

thf

thf

okpifk class discrim. no

okpifk class discrim.f

)(

)(

Feature selection

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

100

200

300

400

500

600

700

800

900

Overlapping Threshold

No

. D

iscri

min

an

t F

ea

ture

s

Walking

Sitting and relaxing

Standing still

Running

All activities

All activities & all accelerometers

0 0.2 0.4 0.6 0.8 10

200

400

600

800

1000

Overlapping Threshold

No

. D

icri

min

an

t F

ea

ture

s

11

• Features extraction based on a windowing method • Data corresponding to hip accelerometer

thf

thf

okpifk class discrim. no

okpifk class discrim.f

)(

)(

Classification (SVM)

12

• Fast

• Simple solutions

• Good precedents

• Binary multiclass models based on

• Different kernels (linear, quadratic, RBF, MPL, etc.)

Classification (SVM)

13

• Fast

• Simple solutions

• Good precedents

• Binary multiclass models based on

• Different kernels (linear, quadratic, RBF, MPL, etc.)

Classification (DT)

14

• Very fast

• Easy interpretability

• Entropy related

Test

15

• Cross validation

▫ Leave-one-subject-out

▫ 50% training – 50% test

SVM DT

LAB 96.37 ± 4.58 98.92 ± 1.08

SEM 75.81 ± 0.90 95.05 ± 1.20

Mean (%) ± standard deviation (%)

Comparison with other studies

16

Work Accuracy rates

S.W. Lee and K. Mase. Activity and location recognition using wearable sensors. 92.85% a 95.91%

J. Mantyjarvi, J. Himberg, and T. Seppanen. Recognizing human motion with multiple acceleration sensors.

83% a 90%

K. Aminian, P. Robert, E. E. Buchser, B. Rutschmann, D. Hayoz, and M. Depairon. Physical activity monitoring based on accelerometry: validation and comparison with video observation.

89.30%

L. Bao and S.S. Intille. Physical Activity Recognition from Acceleration Data under Semi-Naturalistic Conditions

89%

THIS WORK 95.05% (SEM), 98.92(LAB)

Source: L. Bao and S.S. Intille. Physical Activity Recognition from Acceleration Data under Semi-Naturalistic Conditions

Conclusion

• Only a source of data (accelerometer ) is necessary for inferring the considered activities

• Best results (≈ 100%) for laboratory data:

• Seminaturalistic accuracy rates are highly improved with respect to prior works (≈ 95%)

17

Filtering

Feature extraction over

the complete signal

Features selected: coef. wavelets,

autocorrelación or amplitude

geometric mean

Classification based on DT

Future work

• Analyze other methods and compare with the presented work

• Study other activities and apply this methodology to other kind of problems

• Define new approaches for other physiological parameters (ECG, PPG, body temperature,…)

18

Thank you for your attention

Questions?

19