+ All Categories
Home > Documents > Dynamic Wireless Indoor Localization Incorporating With an...

Dynamic Wireless Indoor Localization Incorporating With an...

Date post: 04-Aug-2020
Category:
Upload: others
View: 2 times
Download: 0 times
Share this document with a friend
12
1940 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 3, MARCH 2019 Dynamic Wireless Indoor Localization Incorporating With an Autonomous Mobile Robot Based on an Adaptive Signal Model Fingerprinting Approach Ren C. Luo, Fellow, IEEE, and Tung Jung Hsiao AbstractIndoor localization based on received signal strength (RSS) will result in a decreased precision after the environment changes. In this paper, we develop an adaptive wireless indoor localization system (ILS) for dynamic envi- ronments. The system consists of the following two compo- nents: an automated database updating process and a new fingerprinting algorithm called adaptive signal model fin- gerprinting (ASMF). In the ILS, a self-locating mobile robot is set up to continuously collect RSS measurement data within the localization space for autonomously updating the fingerprint database. ASMF is designed to reduce the time consumption and the amount of RSS data needed for up- dating the database. The fingerprint of the signal in ASMF is constructed by the position of the beacons and three signal models, which can be duly corrected based on the regression and optimization algorithm. Finally, we pro- pose experiments for positioning targets in the static and dynamic environments and compare the results of the ASMF algorithm with traditional trilateration and k- nearest-neighbor fingerprinting algorithms. The experimen- tal results demonstrate that the ASMF-based ILS provides much better performance in both static and dynamic en- vironments; furthermore, the positioning accuracy can be actually maintained by the autonomous updated ASMF database. Index TermsAdaptive database, adaptive signal model, dynamic fingerprinting, indoor localization. I. INTRODUCTION W ITH the modernization of society, people spend more than half of their day indoor. Indoor location-based ser- vice (LBS) has become commercial industry with the most po- tential. Additionally, indoor localization, which is the key tech- nology in indoor LBS, has been highly concerned [1]. A global Manuscript received June 13, 2017; revised December 11, 2018 and March 30, 2018; accepted April 12, 2018. Date of publication May 17, 2018; date of current version October 31, 2018. (Corresponding author: Tung Jung Hsiao.) R. C. Luo is with the Center for Intelligent Robotics and Automa- tion Research, National Taiwan University, Taipei 10617, Taiwan (e-mail: [email protected]). T. J. Hsiao is with the Department of Electrical Engineering, Na- tional Taiwan University, Taipei 10617, Taiwan (e-mail: [email protected]. ntu.edu.tw). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TIE.2018.2833021 positioning system is the most popular wireless localization technology in the world. Unfortunately, it has poor performance in indoor environments. So far, many types of sensors have been proposed in literature studies for the implementation of indoor localization, for instance, vision [2], Kinect [3], RFID [4]–[6], ultrasonic [7], [8], and Wi-Fi [9]–[11]. However, considering the device cost and convenience of deployment, radio frequency sensors with received signal strength (RSS) are still the most popular choice for an indoor localization system (ILS). There are two major localization approaches in the RSS-based ILS: ge- ometric and fingerprinting approaches. The geometric approach determines the target’s position based on geometric relations be- tween targets and beacons. The geometric-based ILS is easy to implement but results in a lower accuracy position because the signals are strongly affected by nonline of sight (NLOS). Alter- natively, the fingerprinting approach locates targets according to the previously built database, which records the fingerprints of the signal corresponding to specific reference points (RPs). The fingerprinting approach can mitigate the NLOS effect in static environments and has higher accuracy because the NLOS RSS measurement will be considered for building the fingerprinting database [31], [32]. However, because of the time-consuming process of acquiring signal data, heavy prep work in extracting reliable features from the signals, as well as the less flexible database, the fingerprinting-based ILS is hard to implement. In general, there are three common problems for the fingerprinting-based ILS in dynamic environments: 1) how to reduce positioning errors caused by obstacle shad- ows; 2) how to adjust the database if any beacon position changes; 3) how to maintain the precision in dynamic environments. The first problem in static environments is an issue with NLOS, and many methods have been proposed for dealing with this problem. The statistical methods are commonly used for dealing with the NLOS problem, for example, the bivariate Gaussian mixture model (GMM) can be used to negate the effects of the erroneous distance caused by NLOS [12]. Statisti- cal methods are effective and simple but unsuitable in dynamic environments because the obstacles can move and invalidate the trained model. Moved obstacle detection is important to alleviate the interference of changeable obstacles in dynamic environments. The authors of [13] and [14] utilized the radio tomographic method to detect unknown objects and position device-free targets by calculating the variation of exceptional 0278-0046 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
Transcript
Page 1: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

1940 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 3, MARCH 2019

Dynamic Wireless Indoor LocalizationIncorporating With an Autonomous MobileRobot Based on an Adaptive Signal Model

Fingerprinting ApproachRen C. Luo, Fellow, IEEE, and Tung Jung Hsiao

Abstract—Indoor localization based on received signalstrength (RSS) will result in a decreased precision after theenvironment changes. In this paper, we develop an adaptivewireless indoor localization system (ILS) for dynamic envi-ronments. The system consists of the following two compo-nents: an automated database updating process and a newfingerprinting algorithm called adaptive signal model fin-gerprinting (ASMF). In the ILS, a self-locating mobile robotis set up to continuously collect RSS measurement datawithin the localization space for autonomously updating thefingerprint database. ASMF is designed to reduce the timeconsumption and the amount of RSS data needed for up-dating the database. The fingerprint of the signal in ASMFis constructed by the position of the beacons and threesignal models, which can be duly corrected based on theregression and optimization algorithm. Finally, we pro-pose experiments for positioning targets in the staticand dynamic environments and compare the results ofthe ASMF algorithm with traditional trilateration and k-nearest-neighbor fingerprinting algorithms. The experimen-tal results demonstrate that the ASMF-based ILS providesmuch better performance in both static and dynamic en-vironments; furthermore, the positioning accuracy can beactually maintained by the autonomous updated ASMFdatabase.

Index Terms—Adaptive database, adaptive signal model,dynamic fingerprinting, indoor localization.

I. INTRODUCTION

W ITH the modernization of society, people spend morethan half of their day indoor. Indoor location-based ser-

vice (LBS) has become commercial industry with the most po-tential. Additionally, indoor localization, which is the key tech-nology in indoor LBS, has been highly concerned [1]. A global

Manuscript received June 13, 2017; revised December 11, 2018 andMarch 30, 2018; accepted April 12, 2018. Date of publication May 17,2018; date of current version October 31, 2018. (Corresponding author:Tung Jung Hsiao.)

R. C. Luo is with the Center for Intelligent Robotics and Automa-tion Research, National Taiwan University, Taipei 10617, Taiwan (e-mail:[email protected]).

T. J. Hsiao is with the Department of Electrical Engineering, Na-tional Taiwan University, Taipei 10617, Taiwan (e-mail: [email protected]).

Color versions of one or more of the figures in this paper are availableonline at http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TIE.2018.2833021

positioning system is the most popular wireless localizationtechnology in the world. Unfortunately, it has poor performancein indoor environments. So far, many types of sensors have beenproposed in literature studies for the implementation of indoorlocalization, for instance, vision [2], Kinect [3], RFID [4]–[6],ultrasonic [7], [8], and Wi-Fi [9]–[11]. However, considering thedevice cost and convenience of deployment, radio frequencysensors with received signal strength (RSS) are still the mostpopular choice for an indoor localization system (ILS). Thereare two major localization approaches in the RSS-based ILS: ge-ometric and fingerprinting approaches. The geometric approachdetermines the target’s position based on geometric relations be-tween targets and beacons. The geometric-based ILS is easy toimplement but results in a lower accuracy position because thesignals are strongly affected by nonline of sight (NLOS). Alter-natively, the fingerprinting approach locates targets according tothe previously built database, which records the fingerprints ofthe signal corresponding to specific reference points (RPs). Thefingerprinting approach can mitigate the NLOS effect in staticenvironments and has higher accuracy because the NLOS RSSmeasurement will be considered for building the fingerprintingdatabase [31], [32]. However, because of the time-consumingprocess of acquiring signal data, heavy prep work in extractingreliable features from the signals, as well as the less flexibledatabase, the fingerprinting-based ILS is hard to implement.

In general, there are three common problems for thefingerprinting-based ILS in dynamic environments:

1) how to reduce positioning errors caused by obstacle shad-ows;

2) how to adjust the database if any beacon position changes;3) how to maintain the precision in dynamic environments.

The first problem in static environments is an issue withNLOS, and many methods have been proposed for dealing withthis problem. The statistical methods are commonly used fordealing with the NLOS problem, for example, the bivariateGaussian mixture model (GMM) can be used to negate theeffects of the erroneous distance caused by NLOS [12]. Statisti-cal methods are effective and simple but unsuitable in dynamicenvironments because the obstacles can move and invalidatethe trained model. Moved obstacle detection is important toalleviate the interference of changeable obstacles in dynamicenvironments. The authors of [13] and [14] utilized the radiotomographic method to detect unknown objects and positiondevice-free targets by calculating the variation of exceptional

0278-0046 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Page 2: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

LUO AND HSIAO: DYNAMIC WIRELESS INDOOR LOCALIZATION INCORPORATING WITH AN AUTONOMOUS MOBILE ROBOT 1941

signal attenuation. Moreover, the exceptional signal attenuationof whole localization can be mapped based on the radio tomo-graphic method, and the result can be used to deal with the firstproblem in dynamic environments. In the second problem, themost essential problem is how to continuously obtain beaconpositions. This problem can be regarded as how to continuouslylocalize unknown beacons in the workplace. The most commonmethod is to assign an individual or a group of mobile robotsto collect the signal data from these unknown beacons. Thissignal data, coupled with the measuring pose of the robot(s),can be used to localize the unknown beacons [15], [16]. For thethird problem, a dynamic fingerprinting combination methodhas been proposed to locate targets by dynamically fusing spatialcorrelations from multiple fingerprinting systems [17]. A novelclient/server-based system has been proposed to dynamicallyestimate and calibrate the fine radio map based on a modifiedBayesian regression algorithm [18]. Many dynamic fingerprint-ing algorithms have been proposed, but the computational coston updating changeable parameters is still difficult to overcome.Therefore, we propose an autonomous database updating pro-cess and a new fingerprinting method for the implementationof the ILS in dynamic environments. The contributions of thispaper can be summarized as follows.

1) We construct an ILS based on a new algorithm calledadaptive signal model fingerprinting (ASMF). The ILScan locate targets and update the database simultaneously,which allows the ILS to adapt to dynamic environments.

2) A self-locating robot is set up to patrol the localizationarea. The robot continuously collects RSS measurementdata and maps the localization area. The RSS measure-ment and the map are used to construct and update thedatabase.

3) The ASMF algorithm is developed for constructing a dy-namic database. The fingerprints in the ASMF databaseare determined by the signal models and the beacon po-sitions. The fingerprints are easily adjusted by correctingthe signal models. ASMF takes less time than traditionalalgorithms in rebuilding the fingerprinting database.

4) We design three signal models for the signal noise, thedistribution of the exceptional signal attenuation, and theRSS–distance relationship of the signal.

5) We use regression method to fit the path loss model andthe signal noise, which lead to better positioning accu-racy than the traditional log path loss model and whiteGaussian noise.

6) The shadowing model is designed for mapping the excep-tional signal attenuation of the localization area and usedto compensate for the power loss of the signals producedby obstacle shadows.

7) The experimental results demonstrate that our ILS hasthe better performances than the trilateration-based ILSand the k-nearest neighbors fingerprinting (KNNF)-basedILS in both the static and dynamic environments.

The rest of this paper is outlined as follows. In Section II, weshow the specifications of the system structure. In Section III,we present the signal models used in ASMF. Section IV presentsthe process for building and updating the ASMF database andpositioning targets. In Section V, we perform experiments forcomparing the localization results of our ILS with other commonmethods. Section VI presents the conclusion.

Fig. 1. ASMF-based ILS structure.

Fig. 2. Kangaroo robot.

II. SYSTEM STRUCTURE

The overall ILS structure is shown in Fig. 1. The system isdivided into four segments: beacons, targets, a robot, and a com-puter. First, the beacon used in the ILS is a ZigBee developmentboard following the protocol of IEEE 802.15.4. The beaconsare a set for periodically transmitting signals to the targets andthe robot with a transmitting frequency of 20 Hz. Second, thetargets are people carrying a ZigBee board that receives thesignals from the beacons and transmits the RSS measurementdata to the computer. Third, the robot, called “Kangaroo,” isdesigned and developed by our laboratory, as shown in Fig. 2.The robot is assembled to include a ZigBee board and a Hokuyolaser rangefinder (LRF). The robot can locate itself and map thelocalization area based on the LRF and odometry with the si-multaneous localization and mapping (SLAM) algorithm [19],[20]. The board in the robot is set to transmit the RSS measure-ment data and the robot’s position to the computer every 0.5 s.Finally, in the computer, we position the targets and the robotand update the database as necessary. The task phase of theASMF-based ILS can be divided into two phases, offline andonline, as shown in Figs. 3 and 4. In the offline phase, the robot isarranged to move around the localization area and record RSS

Page 3: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

1942 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 3, MARCH 2019

Fig. 3. Flowchart of the ASMF-based ILS in the offline phase.

Fig. 4. Flowchart of the ASMF-based ILS in the online phase.

measurement data from beacons in RPs. Then, we derive theinitial ASMF database including the beacon positions as wellas three signal models: the fading model, the shadowing model,and the fitting path loss model. During the online phase, theILS locates the targets and the robot and corrects inappropriateparameters in the database. In the right section of Fig. 4, theILS positions the target based on the RSS measurement datafrom the target as well as the ASMF database. At the same time,the ILS compares the RSS measurement data recorded from therobot with the fingerprints in the database every time the robotfinishes one patrol circle. Then, the ILS recognizes the environ-mental change and corrects models in the database. The updateddatabase process is shown in the left section of Fig. 4.

III. ASMF SIGNAL MODEL

The log-distance path loss model, formulated as (1), is themost common model for describing the relationship betweenthe transmission distance and the RSS, where V (d) is the RSScorresponding to the transmission distance d, Vdo

is the refer-ence RSS corresponding to the reference transmission distancedo , γ is the power decay factor, and N (0, σ2), the signal noise,is a zero mean Gaussian distribution with standard deviation(SD) σ. The unit of distance is meter (m) and the unit of RSSis decibel-milliwatts (dBm). However, this model is too sim-ple and not enough to obtain accurate distances in the complexenvironment. We proposed a more appropriate signal model,formulated as (2), where NF (d) is the noise caused by the mea-surement error and multipath effect, S(pt ,pb) is the additionalsignal attenuation by shadowing effects when the transmitterand the receiver are in pt and pb , and VF (d) is the function forthe transformation between the transmission distances and the

Fig. 5. SDs of the RSS datasets used to train the fading model.

RSS values

V (d) = Vdo− 10γ log (d/do) + N (0, σ2) (1)

V (d) = VF (d) − S(pt ,pb) + NF (d). (2)

A. Fading Model

In ASMF, we used a fading model to estimate the noise causedby the measurement noise and multipath effect of the radiosignal. According to the past studies, the GMM is an appropriatemodel for the distribution of the radio signal noise [12], [21].Therefore, the fading model is constructed by the two zero-meanGaussian distributions with two different SDs. The model is asfollows:

NF (d) ∼ φ1(d)N (0, σ21 ) + φ2(d)N (0, σ2

2 ) (3)

where φ1(d) is the weight of the first Gaussian distribution,φ2(d) = 1 − φ1(d) is the weight of the second Gaussian dis-tribution, and σ1 and σ2 are the SDs. To avoid confusion, welet σ1 < σ2 so that the first Gaussian distribution is always thelower variance component of the fading model. In order to de-termine the parameters in the fading model, we investigated thestatistics of the RSS noise by measured statistics of links in thepractical environment. We set up ten beacons in a laboratory andcollected the RSS measurement data from the beacons in 40 dif-ferent positions, totaling 400 sets of RSS measurement data, andabout 1100 RSS values from each beacon were taken in eachposition. Then, the SDs of the RSS measurement data weredivided into 22 sets according to the transmission distance be-tween the measuring point and the beacons, which is from 0.75to 11.25 m with 0.5 m intervals. The distribution of the SDs indifferent transmission distances is depicted in Fig. 5. In orderto determine σ1 and σ2 , we separated the SDs in each datasetinto two clusters by the expectation–maximization algorithm[22]. The probability of the data points for two distributions iscalculated, and the weight φ1 is the ratio of the number of thedata points having higher probability for a low SD distributioncompared to the total number of data points. For example, theclustered result of the dataset with a 2-m transmission distanceis shown in Fig. 6. Then, we obtain σ1,l and σ2,l , which are theSDs of the first Gaussian distribution and the second Gaussiandistribution, respectively, in the lth dataset. σ1 and σ2 are cal-culated as the mean of σ1,l and σ2,l , l = 1, ..., L, where L is 22.The weight φ1(d) is assumed as a linearly independent function(4). The sum of squared errors between φ1(d) and the weightsof the first Gaussian distribution is computed by (5), where dl isthe transmission distance corresponding to the lth dataset, and

Page 4: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

LUO AND HSIAO: DYNAMIC WIRELESS INDOOR LOCALIZATION INCORPORATING WITH AN AUTONOMOUS MOBILE ROBOT 1943

Fig. 6. GMM for the SDs of the RSS measurement.

Fig. 7. Weights of the first Gaussian distribution and the weightfunction.

φ1,l is the weight of the first Gaussian distribution. Then, (5) isreformulated as (6) and (7), and the optimal α minimizing thesum of squared errors is computed as (8):

φ1(d) = a0 + a1d + a2d2 (4)

L∑

l=1

εk =L∑

l=1

[φ1,l − φ1 (dl)]2 (5)

ϕ1 = Dα + ε (6)

ϕ1 =

⎢⎣φ1,1

...φ1,L

⎥⎦ D =

⎢⎣1 d1 d2

1...

. . ....

1 dL d2L

⎥⎦α =

[a0a1a2

]ε =

⎢⎣ε1...

εL

⎥⎦

(7)

α =(DTD

)−1DTϕ1 . (8)

Fig. 7 depicts the weights of the first Gaussian distributionin each dataset and the weight estimated by φ1(d). φ1(d) is adecreasing function, which represents that the distribution of theRSS measurement in short transmission distance is more con-centrative. Fig. 8 plots the distributions of the RSS measurementand the estimated distributions in three different transmissiondistances. The estimated distributions actually conform to thedistributions of the RSS measurement.

B. Shadowing Model

The major reason why the signal rapidly attenuates is theshadowing effect of obstacles. Consequently, we establish ashadowing model mapping the exceptional signal attenuationin the localization area for mitigating the shadowing effect.First of all, we assume that the exceptional signal attenuationhappens when the radio signal passes through obstacles suchas walls or large furniture. The placement of the obstacles can

be obtained from the localization area map obtained by therobot with the LRF. However, it is difficult to directly determinethe exact position of the obstacles in the original localizationarea map shown in Fig. 9(a). These deficiencies are caused byholes in the map as: 1) the LRF can only detect the surfaceof the obstacles; and 2) the LRF is unable to detect surfacesbehind other obstacles. Therefore, in order to acquire the maphaving complete obstacles, we employ the image-processingmethods including image binarization, dilation, erosion, andthe hole-filling algorithm [23]. The processed map is shown inFig. 9(b). Moreover, in order to reduce the computational cost,we divide the map into a grid map with I × J grids. Each gridhas M × N pixels and a shadowing value si,j representing theexceptional signal attenuation when the signals pass through thegrid. The overall shadowing values in the localization area canbe formulated as a (I ∗ J) × 1 vector in (9) and (10), where Hsis a threshold for determining whether the grid has obstacles, inwhich h(m,n) is the gray scale of the pixel within the grid, and0–4 dBm is the limitation for the shadowing value of each grid,which is the normal range of power attenuation caused by thecommon indoor obstacles according to [24]. After constructingthe shadowing model for the localization area, the exceptionalsignal attenuation of the link from a beacon in pb to a target in pt

is calculated by accumulating shadowing values of the grids thelink passes through. Furthermore, according to [25], the shapeof the link of the radio signal can be regarded as an ellipse. Thepassed grids are determined in (11) and (12), where wellipse isthe width of the ellipse of the link, and pgi , j

is the position ofthe center of the grid gi,j . Finally, the total exceptional signalattenuation between the beacon in pb and the target in pt iscalculated in (13). Fig. 10 shows an illustration for calculatingthe total exceptional power attenuation of the link

S = [s1,1 , s1,2 , . . . , s1,J , s2,1 , . . . , sI ,J−1 , sI ,J ]T

(9)

si,j ={

0 − 4,∑M

m=1∑N

n=1 h (m,n) ≥ HS

0, else(10)

ui,j (pt ,pb) =

⎧⎨

1,∣∣pt − pgi , j

∣∣ +∣∣pb − pgi , j

∣∣< |pt − pb | + wellipse

0, otherwise(11)

U(pt ,pb) = [u1,1 . . . uI ,J ] (12)

S(pt ,pb) = U(pt ,pb)S. (13)

C. Fitting Path Loss Model

In general, the most commonly used model for describingthe relationship of RSS and distance is the log-distance pathloss model; however, its distance estimation might be inaccu-rate when the indoor environment is complex because the simpleexponential shape of the model is sometimes inconsistent withpractical Signal propagation. In this paper, we propose the fit-ting path loss model for formulating the practical path loss ofthe signal. The fitting model is composed of several indepen-dent basic functions, and their coefficients are determined bythe optimization approach, which minimizes the error betweenthe model and the practical RSS measurement. We prepare thefollowing four different curve functions to find the best-fittingcurve function: a linearly independent function (F1), a nonlinear

Page 5: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

1944 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 3, MARCH 2019

Fig. 8. Histogram of the RSS measurements and the estimated distributions. The estimated distributions are formulated as φ1 (d)N (V, σ21 ) +

φ2 (d)N (V, σ22 ), where d is the transmission distance (m) and V is the average RSS (dBm). (a) d = 2.24 and V = −53.31. (b) d = 5.13 and

V = −61.0459. (c) d = 10.32 and V = −63.3097.

Fig. 9. Two-dimensional map of the localization area. (a) Original map.(b) Image-processed map.

Fig. 10. Exceptional signal attenuation between two blue stars is 11.

TABLE IFITTING CURVE FUNCTION

function (F2), and two log-distance path loss model functionswith different parameters (F3, F4). The details of the functionsare presented in Table I. The optimal parameters of F1 are ob-tained by solving the least-squares equation. F2, F3, and F4 arenonlinear functions, which do not have a closed-form solution,so we use the downhill simplex (amoeba) algorithm to find theoptimal parameters. For F3, the parameters do , Vdo

, and n arecomputed by the optimal algorithm. For F4, we assume that do

and Vdoare known and only compute the optimal n. This al-

gorithm moves a simplex of points through a high-dimensionalspace based on geometric relationships to find the function min-imums [26].

Fig. 11. Comparison of the fitting curves.

Fig. 12. Mean errors between the RSS estimated by fitting functionsand the RSS measurement.

TABLE IIMEAN ERROR OF THE FITTING FUNCTIONS

In order to determine which curve function is the most suit-able, we recorded RSS measurement data from 12 beacons in500 sample positions and used the RSS measurement data toobtain the parameters of four fitting functions. Fig. 11 plots thefitting curves of four functions based on the RSS measurementdata from one of the beacons, and it is clear that the curves of thelog-distance path log model, F3 and F4, are different from thecurves of the other two functions. Fig. 12 shows the error of the12 beacons between the RSS estimated by fitting functions andthe RSS measurement data. F1 and F2 have fewer errors thanF3 and F4, which means that two nonexponential functions aremore appropriate to the practical RSS measurement. Accordingto Table II, F2 that has the least error is chosen to be the fittingpath loss model in this paper.

Finally, the signal fingerprints in the ASMF database is for-mulated as the probability density functions (PDFs) based onthe fading model, the shadowing model, the fitting path loss

Page 6: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

LUO AND HSIAO: DYNAMIC WIRELESS INDOOR LOCALIZATION INCORPORATING WITH AN AUTONOMOUS MOBILE ROBOT 1945

model, and the beacon positions. The signal fingerprint corre-sponding to the specific position p is computed in (14) and (15),where pb,k is the kth beacon position, NB is the number of thebeacons, d = ‖p − pb,k‖, and V = VF (d) − S(p,pb,k ):

Vfingerprint(p) = {Vpdf (p,pb,1), . . . , Vpdf (p,pb,NB)} (14)

Vpdf (p,pb,k ) ∼ φ1(d)N (V, σ21 ) + φ2(d)N (V, σ2

2 ). (15)

IV. ASMF DYNAMIC DATABASE AND LOCALIZATION

In this section, we depict the process of updating the databaseand positioning objects in the ASMF-based ILS. The updat-ing process includes environmental change recognition, movedbeacon position estimation, and shadowing model optimization.

A. Environmental Change Recognition

For environmental change recognition, we assume two sit-uations: the variation of the beacons and the obstacles. In thebeacon-moved situation, the position of the moved beacons is re-estimated, and the shadowing model is reoptimized as well. Forthe other situation, only the shadowing model is reoptimized.The environmental change recognition is based on comparingthe similarity between the RSS measurement data collected bythe robot and the fingerprints in the ASMF database, which canbe determined as the conditional probabilities of the RSS mea-surement, given fingerprints in the ASMF database as follows:

P (Vrobot,t |Vfingerprint(probot,t)), t = 1, ..., TT (16)

P (Vrobot,t |Vfingerprint(probot,t))

=NB∏

k=1

P (Vrobot,t |Vpdf(probot,t ,pb,k )) (17)

where Vrobot,t is the mean of the RSS measurement in the positionprobot,t , and TT is the number of measuring points in one patrolcircle. Equation (16) can be reformulated as the product ofthe probabilities of the mean of the RSS measurement givenfingerprints corresponding to all beacons formulated in (17)and (18), where dt,k is the distance between the tth measuringpoint and the kth beacon, V is the estimated RSS calculated in(19), and f(Vrobot,t |V, σ2

1 ) and f(Vrobot,t |V, σ22 ) are the normal

distributions formulated in (20) and (21), respectively:

P (Vrobot,t |Vpdf (probot,t ,pb,k ))

= φ1(dt,k )f(Vrobot,t |V, σ21 ) + φ2(dt,k )f(Vrobot,t |V, σ2

2 )(18)

V = VF (dt,k ) − S(probot,t ,pb,k ) (19)

f(Vrobot,t |V, σ21 ) =

1√2σ2

1πe− (V robot, t −V ) 2

2 σ 21 (20)

f(Vrobot,t |V, σ22 ) =

1√2σ2

2πe− (V robot, t −V ) 2

2 σ 22 . (21)

P (Vrobot,t |Vpdf (probot,t ,pb,k )) will be low if the kth beaconhas been moved, so whether the beacon has been re-estimatedcan be determined by (22), where H1 is a threshold. Finally,for the obstacle-moved situation, only the RSS measurementsatisfying (23) will be used to optimize the new shadowingmodel, where H2 is a threshold. H1 and H2 are the adjustable

parameters dependent on φ1(d) and the SDs obtained in thefading model. For example, if the localization environment iscomplex and the SDs of the RSS distribution are high, we canchoose a lower H2 value to avoid the case where the databaseupdate frequency is too high. However, the lower H2 valuealso allows the ILS to have more opportunity to localize thetarget in the wrong RP. It is a tradeoff between the databaseupdate frequency and the localization accuracy. H1 usually canbe chosen as a low value because the movement of the beaconwill lead to RSS measurements becoming very different fromthe database

1TT

TT∑

t=1

P (Vrobot,t |Vpdf (probot,t ,pb,k )) < H1 (22)

P (Vrobot,t |Vfingerprint(probot,t)) < H2 . (23)

B. Position Estimation of Moved Beacons and Targets

A particle filter is a useful algorithm for signal processing tohandle a multimodal PDF. Moreover, the kernel particle filter(KPF) [27], [28], which uses kernel density estimation (KDE)to approximate the posterior PDF, is more effective for allo-cation of particles, as it updates particles states by the meanshift algorithm [29] instead of replacing the original particles.In ASMF-based localization, we use the KPF to estimate notonly the position of the moved beacons, but also the position ofthe targets. The system has two KPFs for separately estimatingthe target and the moved beacons. First of all, we denote the stateof the positioning object as xt , observations Zt = {V1 , ..., Vt},and the particles of the object x(m )

t , and the number of the par-ticles for the object used in the KPF is Np . The motion modelsfor the target and the moved beacons are assumed as the ran-dom walk models [30]. The maximum speed for the target is1.5 m/s, and the maximum speed for the moved beacons is 0.6m/s. Next, the posterior density of the object is estimated basedon KDE in (24), where w

(m )t is an associated weight of the

particle x(m )t , and the bandwidth κ is chosen to minimize the

integrated mean squared error between the posterior density andthe kernel density. Mean shift calculates gradients of each par-ticle and moves each particle along its gradient direction to itssample mean formulated in (25). Finally, after mean shift, theweights of the new particles x′(m )

t have to be redistributed:

p(xt |Zt) =NP∑

m=1

Kκ(xt − x(m )t )w(m )

t (24)

m(x(m )t ) =

∑Np

m ′=1 Hκ(x(m )t − x(m ′)

t )w(m ′)t x(m ′)

t∑Np

m ′=1 Hκ(x(m )t − x(m ′)

t )w(m ′)t

. (25)

C. Shadowing Model Optimization

The grid that has obstacles in the updated map can be for-mulated as a diagonal matrix (26). The exceptional signal at-tenuation between two positions is calculated by subtracting themeasured RSS from the estimated RSS shown in (27), whereVF (dt,k ) is the estimated RSS neglecting exceptional signal at-tenuation from the kth beacon to the position probot,t calculatedby the fitting path loss model. The total exceptional signal atten-uation of the links of the measuring positions and the beaconsis expressed as a vector ∈ �(NB ∗TT )×1 in (28). In contrast, the

Page 7: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

1946 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 3, MARCH 2019

estimated exceptional signal attenuation calculated by the shad-owing model is formulated as (29). The process of optimizingthe shadowing model, which finds an optimal Soptimal minimiz-ing the error between S′ and S, can be regarded as a constrainedlinear least-squares problem and has the closed-form solution(30).

Moreover, in order to reduce the computational time in theupdating process of the obstacle-moved situation, we only use apart of the data satisfying (23) to optimize the shadowing model.S and A can be divided into two parts: one certain and oneuncertain. The grids in the unceratin part are passed in the linksof the RSS measurement satisfying (23). Then, the shadowingmodel can be divided into [S1 S2 ]T , and its correspondingA matrix was divided into [A1 A2 ], and only the shadowingvalues in S2 need adjustment. Then, (30) can be reformulatedas (31), where A1S1 can be regarded as a constant value C, andS′′ is the subpart of S′:

Q =

⎢⎢⎢⎣

q1,1 0 · · · 0

0 q1,2...

.... . . 0

0 · · · 0 qI ,J

⎥⎥⎥⎦ ,

{qi,j = 1, the grid gi,j has obstaclesqi,j = 0, otherwise

(26)

S ′(probot,t ,pb,k ) = VF (dt,k ) − Vrobot,t (27)

S′ =[S ′(probot,1 ,pb,1), ..., S′(probot,TT

,pb,NB)]T

(28)

S =

⎢⎣U(probot,1 ,pb,1)

...U(probot,TT

,pb,N B)

⎥⎦Q

⎢⎣s1,1

...sI ,J

⎥⎦ = AS (29)

Soptimal = arg minS

12

∥∥∥S′ − S∥∥∥

2

2, 0 ≤ si,j ≤ 4 (30)

S2,optimal = arg minS2

12‖S′′ − A2S2 − C‖2

2 , 0 ≤ si,j ≤ 4.

(31)

V. PHYSICAL EXPERIMENT

A. Experimental Setup

The indoor environment used for the experiments is a14 × 10.5 m2 laboratory, which is a space with room partition,and the view of the laboratory is shown in Fig. 13. The velocityof the robot is 0.3 m/s, and the LRF detection range is from0.2 to 4 m. The average self-localization error of the robot isapproximately 0.1 m. The ZigBee board is shown in Fig. 14.There are 12 ZigBee boards deployed in the laboratory tobe the reference beacons. The floor layout of the laboratoryis depicted in Fig. 17. The deployment of the beacons andRPs is shown in Fig. 18. Green crosses represent the beaconpositions; red circles represent the RPs. There are a total of531 RPs. In the online phase, we arrange a person carrying theZigBee board, shown in Fig. 15 to be the target and the robot tomove along the patrolling trajectory that is previously planned.The patrolling trajectory of the robot is shown in Fig. 19. Thereare 145 measuring points in the trajectory. In each measuring

Fig. 13. View of the experimental environment.

Fig. 14. ZigBee board.

Fig. 15. Person carrying a ZigBee board.

Fig. 16. Cabinet and desk used in experiment 2.

Fig. 17. Floor layout mapped by the robot.

Page 8: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

LUO AND HSIAO: DYNAMIC WIRELESS INDOOR LOCALIZATION INCORPORATING WITH AN AUTONOMOUS MOBILE ROBOT 1947

Fig. 18. Deployment of the beacons and RPs.

Fig. 19. Patrolling trajectory of the robot in an online phase.

point, the robot stops for 2 s to collect the RSS measurementdata.

The robot patrolling trajectory is planned based on two im-portant points, namely, maximizing the total mapping cover-age that the target can travel in the environments and collisionavoidance with the obstacles in the experimental environments.The robot patrolling trajectory is planned for covering the mov-ing trajectory of the target person in collecting enough RSSdata to maintain the accuracy of the localization system. In ex-periments, the coverage of the patrolling trajectory does notcover the whole localization area; therefore, the shadowing val-ues of the place where the robot does not go will then not beconsidered in shadowing model optimization. The larger cover-age of the robot trajectory can provide more complete data forupdating the database, but the tradeoff is that the updating in-terval is also increased. For collision avoidance, the position ofthe preplaced obstacles in the experiments has been consideredfor the robot patrolling trajectory planning so that the robot canavoid the collision with the preplaced obstacles. However, usu-ally, the people in the environment are unpredictable and mayblock the robot. We propose two possible reactions to cope withthis blocking problem. In the situation in which the blockingproblem occurs between two measuring points, the robot willavoid the people. In this case, it almost does not affect the sys-tem. In the other situation in which the blocking problem occurson the measuring point, the robot will then stop and wait untilthe people move away. The patrolling time will increase, whichmay lead to the increase in the database updating time.

The experiments are conducted for analyzing the perfor-mances of the ILS based on ASMF, trilateration, and KNNFapproaches. The initial beacon positions are known in ASMF-based and trilateration-based ILSs. The initial databases builtfor ASMF and KNNF approaches are based on the RSS mea-

Fig. 20. Floor layout after the obstacles have changed.

Fig. 21. Position change of two beacons in experiment 3.

Fig. 22. Moving trajectory of the person in experiment 4.

surement data collected in the offline phase. In the experiments,we allow up to three nontarget people arbitrarily moving withinthe localization area; the only limitation is that the nontarget peo-ple cannot collide with the target person and the robot. There arefour experiments conducted in different situations. In the firstexperiment, we position the target in the environment, which isthe same as the environment in the offline phase. The second andthird experiments are localizing the target after environmentalchanges. In the second experiment, we place additional cabinetsand desks, shown in Fig. 16 in the experimental space. The newtwo-dimensional (2-D) map of the changed laboratory is shownin Fig. 20. In the third experiment, we change the position oftwo beacons, as shown in Fig. 21. Experiment 4 is conductedfor continuously positioning the moving target. The trajectoryfor the target is planned beforehand, as shown in Fig. 22. Therobot is assigned to be the target in experiment 4 for obtainingpractical moving positions of the target.

Page 9: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

1948 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 3, MARCH 2019

Fig. 23. Localization results in experiment 1.

Fig. 24. Cumulative localization errors in experiment 1.

TABLE IIICOMPARISON OF LOCALIZATION RESULTS

B. Performance Evaluation

1) Localization Performance in the Static Environment:In the first experiment, we position the targets in 20 points.Fig. 23 shows the positioning results of the three localizationapproaches. Fig. 24 shows the cumulative positioning errors ofthree localization approaches, and the detailed mean error andSD are presented in Table III. According to the results, ASMFhas the best precision in the static environment experiment anddecreases 58.8% and 24.2% error as compared with the trilat-eration and KNNF approaches, respectively. It is obvious thatthe fingerprinting approaches, KNNF and ASMF, have higheraccuracy than the trilateration approach as long as a trustworthydatabase is constructed in advance.

Comparing the results of KNNF and ASMF, the ASMF ap-proach performs better than the KNNF approach. The signalfingerprints directly computed by the RSS measurements areprone to affection from the outlier data. However, the model-based fingerprint database built by the ASMF approach providesa more complete fingerprint of the signal than the database,which is directly constructed from primary RSS measurements.

Fig. 25. Localization results in experiment 2.

Fig. 26. Cumulative localization errors in experiment 2.

Furthermore, the weights function φ1(d) in the fading model isalso the main factor that provides higher confidence to higherRSS measurement from the beacon that is usually close to thetarget and has a lower chance of signal interference.

2) Localization Performance in Dynamic Environments:In the dynamic environment experiments, we analyze the po-sitioning results of the localization approaches if the obstaclesare changed and the beacons are moved. We also discuss theeffect of updating the database of ASMF. ASMF 1 presentsthat the ILS positions the target based on the initial database,and ASMF 2 presents that the ILS positions the target basedon the updated database. In experiment 2, the obstacle-changedcase, the ASMF approach with the updated database has thebest performance. Fig. 25 shows the positioning results of thelocalization approaches in experiment 2, and Fig. 26 plots the cu-mulative positioning errors of the trilateration, KNNF, ASMF 1,and ASMF 2 approaches. According to the experimental results,the change of the obstacles indeed decreases the positioning pre-cision. The fingerprinting approaches have the higher increasein the positioning error. However, even without updating thedatabase, the ASMF approach still has the higher positioningaccuracy than the trilateration and KNNF approaches. The po-sitioning errors of ASMF2 are almost the same as the errors inexperiment 1. This demonstrates that the ASMF database canactually be quickly corrected only based on the RSS measure-ment data, which are collected by the autonomous robot, and theASMF-based ILS is more robust against the change of the ob-stacles in the localization area. Figs. 27 and 28 compare theshadowing model in terms of the nonupdated database and theupdated database. The black circles are the grids that have obsta-cles. The colored squares represent the grids that have a nonzeroshadowing value, and the different colors represent differentranges of the shadowing value for the grid.

In the third experiment, we change the position of two bea-cons. The ASMF approach with the updated database still hasthe best performance. Fig. 29 shows the new estimated position

Page 10: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

LUO AND HSIAO: DYNAMIC WIRELESS INDOOR LOCALIZATION INCORPORATING WITH AN AUTONOMOUS MOBILE ROBOT 1949

Fig. 27. Shadowing model in the initial database.

Fig. 28. New shadowing model in the updated database.

Fig. 29. Estimated position of two moved beacons.

of the two moved beacons by the KPF, and the errors of po-sition estimation of two moved beacons are 1.11 and 0.41 m,respectively. Fig. 30 shows the localization results in experi-ment 3. Fig. 31 plots the cumulative localization errors of thetrilateration, KNNF, ASMF 1, and ASMF 2 approaches. Firstof all, the trilateration approach with wrong beacons’ positionhas 3.64-m mean error, which is too large to be considered.The moved beacons’ position used in the trilateration approachhas been given, which is the same as the estimated positions inthe updated ASMF database. The trilateration approach has theclosest errors to the result in afore-experiments.

The errors of the position estimation of the two moved bea-cons are insufficient to intensely decrease the localization preci-sion in the trilateration approach. KNNF and ASMF 1 have veryhigh growth in the error. This phenomenon shows that the finger-printing approach significantly reduces the positioning accuracyif the beacons have been moved. However, through updating the

Fig. 30. Localization results in experiment 3.

Fig. 31. Cumulative localization errors in experiment 3.

Fig. 32. Localization results in experiment 4.

database, ASMF2 decreases the 48.18% error compared withASMF 1.

In experiment 4, we localize the target moving in the labo-ratory along the prior planned trajectory. The ground truth ofthe moving target is estimated by the LRF and the SLAM ap-proach. The results based on the ASMF and the ASMF withKPF are presented in Figs. 32 and 33. The localization errorsof the ASMF and the ASMF with KPF are 0.85 and 0.81, re-spectively. According to the result, we can find that the KPFcan help the ILS improve the localization result and increase theaccuracy slightly by limiting the moving distance of the targetin the KPF.

3) Comprehensive Analysis: Finally, according to experi-mental results, we can observe the following properties: Fig. 34shows P (Vrobot,t |Vpdf(probot,t ,pb,k )) for the 145 measuringpoints in the experiments, where Vpdf(probot,t ,pb,k ) are the data

Page 11: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

1950 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 3, MARCH 2019

Fig. 33. Localization errors in experiment 4.

Fig. 34. P (Vrobot, t |Vpdf(probot, t , pb ,k )) in the measuring points in theexperiments.

Fig. 35. Histogram of the P (Vrobot, t |Vpdf(probot, t , pb ,k )) in the measur-ing points in the experiments.

in the initial database. The red line plot shows that the environ-ment does not change and the mean value is 0.151. The green lineplot shows that the obstacles in the environment are changed andthe mean value is 0.073. The blue line plot shows that the beaconposition is changed and the mean value is 0.014. According toFig. 34, we can choose H1 to be equal to 0.02 for recognizingwhether the beacon position is changed. Fig. 35 shows the his-togram of the P (Vrobot,t |Vpdf(probot,t ,pb,k )) in the experiments.When the environment remains static, most of the probabili-ties, as shown in the red line plot, are higher than 0.1 (91%) sothat we make an assumption that if the environment does notchange, the lowest value of the P (Vrobot,t |Vpdf(probot,t ,pb,k ))is 0.1. Therefore, we choose H2 to be 0.1 for the case if the

environment has changed and if the database needs to be up-dated. According to the experimental results, it is clear thatP (Vrobot,t |Vpdf(probot,t ,pb,k )) are different in three experimentalenvironments, and these can be used to perform environmentalchange recognition.

The results in experiment 1 demonstrate that the signal mod-els provide the more suitable fingerprint database than thedatabase directly built by the statistic of the RSS measurement.Experiment 2 demonstrates that the effect of the shadowingmodel can compensate for the exceptional power loss for the sig-nals as well as mitigate the interference of obstacles. Experiment3 demonstrates that our ILS can maintain the localization pre-cision, even with beacons that are moved. Finally, experiment 4shows that the KPF can improve the localization result when theILS continuously positions the moving target. The trilateration-based ILS has almost the same localization performances in thefirst three experiments, but the positioning accuracies are low.The fingerprinting-based approach greatly decreases the posi-tioning precision after the environment changes whether it isthe obstacles that are changed or the beacons. It is clear thatthe static fingerprint database is inappropriate in the dynamicenvironments.

The AMSF-based ILS with the updated database has the bestachievement in the first, second, and third experiments. No mat-ter beacon locations change or new obstacles appear in the lo-calization area, the ASMF ILS keeps a certain accuracy byupdating the database. The updating database process in ASMFcan quickly correct the shadowing model and re-estimate theposition of the moved beacons in second and third experimentsand recalculate the fingerprints for 531 RPs only based on theRSS measurement data collected in the robot trajectory. Thegood positioning accuracies of ASMF 2 in second and third ex-periments demonstrate that the ASMF approach can correct thedatabase using less data instead of recording RSS measurementdata in 531 RPs and building a new database. The time cost ofadapting the ILS to the changed environment is significantly re-duced, and the updating process is automatic fashion in the ILS,which allows the ILS to be actually implemented in practice.

VI. CONCLUSION

In this paper, we proposed an adaptive wireless ILS for dy-namic environments based on an autonomous database updatingprocess and the ASMF approach. In the ASMF approach, thefading model models the noise and provides the confidence in-dex, which decreases the influence of the data with low RSSvalues, for the RSS measurement. The shadowing model helpsthe ILS only have to correct a few parameters in the databaseto adapt the environmental changes. The fitting path loss modelconstructed by the nonlinear regression is more suitable fortransmission distance estimation. Finally, the experimental re-sults indicate that the ASMF-based ILS has a better localiza-tion precision than the ILS based on trilateration and KNNFapproaches in static and dynamic environments, and the au-tomatically updating database based on the autonomous robotis effective for maintaining the ILS in dynamic environments.However, the database updating process in the ILS is periodi-cal rather than in immediate fashion. The process needs a fewminutes depending on the patrolling time of the robot. It isunavoidable that the ILS will have a temporary reduction in ac-curacy during the time, in which the environment has changed,but the robot has not finished its patrol. There are some limita-

Page 12: Dynamic Wireless Indoor Localization Incorporating With an …static.tongtianta.site/paper_pdf/8dca7e56-4c46-11e9-885f-00163e08… · Dynamic Wireless Indoor Localization Incorporating

LUO AND HSIAO: DYNAMIC WIRELESS INDOOR LOCALIZATION INCORPORATING WITH AN AUTONOMOUS MOBILE ROBOT 1951

tions that the ASMF approach cannot avoid, for example, theinfluence of moving obstacles and the local optimum problemin optimizing the shadowing model process.

REFERENCES

[1] Z. Deng, Y. Yu, X. Yuan, N. Wan, and L. Yang, “Situation and developmenttendency of indoor positioning,” China Commun., vol. 10, pp. 42–55,2013.

[2] W. Elloumi, A. Latoui, R. Canals, A. Chetouani, and S. Treuillet, “Indoorpedestrian localization with a smartphone: A comparison of inertial andvision-based methods,” IEEE Sens. J., vol. 16, no. 13, pp. 5376–5388,Jul. 2016.

[3] Y. Song, S. Liu, and J. Tang, “Describing trajectory of surface patchfor human action recognition on RGB and depth videos,” IEEE SignalProcess. Lett., vol. 22, no. 4, pp. 426–429, Apr. 2015.

[4] S. S. Saab and Z. S. Nakad, “A standalone RFID indoor positioning systemusing passive tags,” IEEE Trans. Ind. Electron., vol. 58, no. 5, pp. 1961–1970, May 2011.

[5] S. Park and S. Hashimoto, “Autonomous mobile robot navigation usingpassive RFID in indoor environment,” IEEE Trans. Ind. Electron., vol. 56,no. 7, pp. 2366–2373, Jul. 2009.

[6] B. S. Choi, J. W. Lee, J. J. Lee, and K. T. Park, “A hierarchical algorithmfor indoor mobile robot localization using RFID sensor fusion,” IEEETrans. Ind. Electron., vol. 58, no. 6, pp. 2226–2235, Jun. 2011.

[7] S. J. Kim and B. K. Kim, “Dynamic ultrasonic hybrid localization systemfor indoor mobile robots,” IEEE Trans. Ind. Electron., vol. 60, no. 10,pp. 4562–4573, Oct. 2013.

[8] K. H. Choi, W. S. Ra, S. Y. Park, and J. B. Park, “Robust least squaresapproach to passive target localization using ultrasonic receiver array,”IEEE Trans. Ind. Electron., vol. 61, no. 4, pp. 1993–2002, Apr. 2014.

[9] L. H. Chen, E. H. K. Wu, M. H. Jin, and G. H. Chen, “Intelligent fusion ofWi-Fi and inertial sensor-based positioning systems for indoor pedestriannavigation,” IEEE Sens. J., vol. 14, no. 11, pp. 4034–4042, Nov. 2014.

[10] D. Han, S. Jung, M. Lee, and G. Yoon, “Building a practical Wi-Fi-basedindoor navigation system,” Pervasive Comput., vol. 13, pp. 72–79, 2014.

[11] C. Yang and H. R. Shao, “WiFi-based indoor positioning,” IEEE Commun.Mag., vol. 53, no. 3, pp. 150–157, Mar. 2015.

[12] N. Chuku, A. Pal, and A. Nasipuri, “An RSSI based localization schemefor wireless sensor networks to mitigate shadowing effects,” in Proc. IEEEConf. Southeastcon, 2013, pp. 1–6.

[13] J. Wilson and N. Patwari, “Radio tomographic imaging with wirelessnetworks,” IEEE Trans. Mobile Comput., vol. 9, no. 5, pp. 621–632,May 2010.

[14] S. Nannuru, Y. Li, Y. Zeng, M. Coates, and B. Yang, “Radio-frequencytomography for passive indoor multitarget tracking,” IEEE Trans. MobileComput., vol. 12, no. 12, pp. 2322–2333, Dec. 2013.

[15] W. M. Ibrahim, A. E. M. Taha, and H. S. Hassanein, “Robust wirelessmultihop localization using mobile anchors,” in Proc. IEEE Int. Conf.Commun., 2013, pp. 1506–1511.

[16] S. Kuo-Feng, O. Chia-Ho, and H. C. Jiau, “Localization with mobileanchor points in wireless sensor networks,” IEEE Trans. Veh. Technol.,vol. 54, no. 3, pp. 1187–1197, May 2005.

[17] F. Shih-Hau, H. Ying-Tso, and K. Wen-Hsing, “Dynamic fingerprintingcombination for improved mobile localization,” IEEE Trans. WirelessCommun., vol. 10, no. 12, pp. 4018–4022, Dec. 2011.

[18] M. M. Atia, A. Noureldin, and M. J. Korenberg, “Dynamic online-calibrated radio maps for indoor positioning in wireless local area net-works,” IEEE Trans. Mobile Comput., vol. 12, no. 9, pp. 1774–1787,Sep. 2013.

[19] R. C. Luo, M. Hsiao, and C. H. Xie, “Sensor fusion based vSLAM systemfor 3D environment grid map construction,” in Proc. IEEE Symp. Ind.Electron., 2013, pp. 1–6.

[20] R. C. Luo and C. C. Lai, “Multisensor fusion-based concurrent envi-ronment mapping and moving object detection for intelligent servicerobotics,” IEEE Trans. Ind. Electron., vol. 61, no. 8, pp. 4043–4051,Aug. 2014.

[21] R. Bultitude, “Measurement, characterization and modeling of indoor800/900 MHz radio channels for digital communications,” IEEE Commun.Mag., vol. 25, no. 6, pp. 5–12, Jun. 1987.

[22] L. Xu and M. I. Jordan, “On convergence properties of the EM algorithmfor Gaussian mixtures,” J. Neural Comput., vol. 8, pp. 129–151, 1996.

[23] P. Soille, Morphological Image Analysis: Principles and Applications, 2thed. New York, NY, USA: Springer, 2007.

[24] M. Idoudi, H. Elkhorchani, and K. Grayaa, “Performance evaluation ofwireless sensor networks based on ZigBee technology in smart home,” inProc. IEEE Conf. Elect. Eng. Softw. Appl., 2013, pp. 1–4.

[25] P. Agrawal and N. Patwari, “Correlated link shadow fading in multi-hop wireless networks,” IEEE Trans. Wireless Commun., vol. 8, no. 8,pp. 4024–4036, Aug. 2009.

[26] H. Yuguang and W. F. McColl, “An improved simplex method for functionminimization,” in Proc. IEEE Conf. Syst., Man, Cybern., 1996, vol. 3,pp. 1702–1705.

[27] C. Cheng and R. Ansari, “Kernel particle filter: Iterative sampling forefficient visual tracking,” in Proc. IEEE Int. Conf. Image Process., 2003,vol. 2, pp. III-977–III-980.

[28] C. Cheng and R. Ansari, “Kernel particle filter for visual tracking,” IEEESignal Process. Lett., vol. 12, no. 3, pp. 242–245, Mar. 2005.

[29] Y. Cheng, “Mean shift, mode seeking, and clustering,” IEEE Trans. PatternAnal. Mach. Intell., vol. 17, no. 8, pp. 790–799, Aug. 1995.

[30] M. H. Amri, Y. Becis, D. Aubry, and N. Ramdani, “Indoor human robotlocalization using robust multi-modal data fusion,” in Proc. IEEE Int.Conf. Robot. Autom., 2015, pp. 3456–3463.

[31] Z. Xiao, H. Wen, A. Markham, N. Trigoni, P. Blunsom, and J. Frolik, “Non-line-of-sight identification and mitigation using received signal strength,”IEEE Trans. Wireless Commun., vol. 14, no. 3, pp. 1689–1702, Mar. 2015.

[32] S. He and S.-H. G. Chan, “Wi-Fi fingerprint-based indoor positioning:Recent advances and comparisons,” IEEE Commun. Surveys Tuts., vol. 18,no. 1, pp. 466–490, First Quarter 2016.

Ren C. Luo (M’83–SM’88–F’92) received theDipl.-Ing. and Dr.-Ing. degrees in electrical en-gineering from Technische Universitat Berlin,Berlin, Germany, in 1979 and 1982, respectively.

He is currently the Chief Technology Officerof the Fair Friend Group and the Chair Professorwith National Taiwan University, Taipei, Taiwan.He was a Tenured Full Professor of North Car-olina State University, Raleigh, NC, USA, anda Toshiba Chair Professor at the University ofTokyo, Tokyo, Japan. He also served two terms

as the President of National Chung Cheng University. He has authoredmore than 500 papers in international refereed journals and refereed in-ternational conferences and international patents. His research interestsinclude intelligent robotic systems, multisensor fusion and integration,and 3-D printing manufacturing.

Dr. Luo is the Editor-in-Chief of the IEEE TRANSACTIONS ON INDUS-TRIAL INFORMATICS. He was the President of the IEEE Industrial Electron-ics Society. He also served as an Adviser of the Ministry of EconomicAffairs and Science and a Technical Adviser of Prime Minister’s Office inTaiwan. He is a Fellow of the Institution of Engineering and Technology.

Tung Jung Hsiao received the B.S. and M.S.degrees in electrical engineering from NationalCheng Kung University, Tainan, Taiwan, in 2008and 2010, respectively. He is currently workingtoward the Ph.D. degree in electrical engineeringwith National Taiwan University, Taipei, Taiwan.

He is currently a Research Assistant withthe International Center of Excellence in Intel-ligent Robotics and Automation Research, Na-tional Taiwan University. His research interestsinclude wireless indoor localization, computer vi-

sion, and intelligent robotics.


Recommended