ISBN 978-1-60558-336-5ACM Order Number 533083

Sponsored by: ACM SIGMOBILE

with support from: National Science Foundation,

Microsoft Research, & Cisco

MobiCom’08

San Francisco, CA, USASept. 8-12, 2008

Co-Located WorkshopsCHANTS’08 MELT’08, VANET’08,

WiNS-DR’08, WiNTECH’08, & WuWNeT’08

MobiCom 2008 Compilation Proceedings (MELT�’08) file:///C:/Documents%20and%20Settings/Bob/Desktop/Upload/Mobi...

1 of 5 8/19/2008 4:43 PM

MobiCom�’08 Co-Located Workshops

CHANTS�’08 MELT�’08, VANET�’08, WiNS-DR�’08, WiNTECH�’08, & WuWNeT�’08

Sept. 8-12, 2008 San Francisco, CA, USA

Main PageCHANTS�’08

MELT�’08VANET�’08

WiNS-DR�’08WiNTECH�’08WuWNeT�’08

MELT 2008 Author IndexA B C D E F GH J K L M N OP Q S T V W

Aksu, AylinReduction of Location Estimation Error Using Neural Networks (Page 103)

(Return to Top)

Amundson, IsaacMobile Sensor Localization and Navigation Using RF Doppler Shifts (Page 97)

Ananthasubramaniam, BharathCooperative Localization Using Angle of Arrival Measurements in Non-line-of-sight Environments (Page 117)

Baek, YunjuPrecise Location Tracking System Based on Time Difference of Arrival over LR-WPAN (Page 67)

Baggio, AlineA Geometrical Perspective on Localization (Page 85)

Bargh, Mortaza S. Indoor Localization Based on Response Rate of Bluetooth Inquiries (Page 49)

Beauregard, StéphaneA Novel Backtracking Particle Filter for Pattern Matching Indoor Localization (Page 79)

Bolliger, PhilippRedpin - Adaptive, Zero-Configuration Indoor Localization through User Collaboration (Page 55)

Bonamico, CarloProximity Classification for Mobile Devices Using Wi-Fi Environment Similarity (Page 43)

Burbey, IngridPredicting Future Locations Using Prediction-by-Partial-Match (Page 1)

Carlotto, AlessandroProximity Classification for Mobile Devices Using Wi-Fi Environment Similarity (Page 43)

(Return to Top)Cho, Hyuntae

Precise Location Tracking System Based on Time Difference of Arrival overLR-WPAN (Page 67)

MobiCom 2008 Compilation Proceedings (MELT�’08) file:///C:/Documents%20and%20Settings/Bob/Desktop/Upload/Mobi...

5 of 5 8/19/2008 4:43 PM

(Return to Top)

Santucci, FortunatoLocating ZigBee® Nodes Using the TI®'s CC2431 Location Engine: A TestbedPlatform and New Solutions for Positioning Estimation of WSNs in DynamicIndoor Environments (Page 37)

Son, SanghyunPrecise Location Tracking System Based on Time Difference of Arrival overLR-WPAN (Page 67)

Spring, Michael B.Reduction of Location Estimation Error Using Neural Networks (Page 103)

Swangmuang, NattapongOn Clustering RSS Fingerprints for Improving Scalability of PerformancePrediction of Indoor Positioning Systems (Page 61)

Szcodronski, RickAAMPL: Accelerometer Augmented Mobile Phone Localization (Page 13)

Tennina, StefanoLocating ZigBee® Nodes Using the TI®'s CC2431 Location Engine: A TestbedPlatform and New Solutions for Positioning Estimation of WSNs in DynamicIndoor Environments (Page 37)

Valla, MassimoProximity Classification for Mobile Devices Using Wi-Fi Environment Similarity (Page 43)

Wang, ZizhuoReal-time Tracking for Sensor Networks via SDP and Gradient Method (Page 109)

Wicke, MartinLocalization of Mobile Users Using Trajectory Matching (Page 123)

WidyawanA Novel Backtracking Particle Filter for Pattern Matching Indoor Localization (Page 79)

Wolkowicz, HenrySensor Network Localization, Euclidean Distance Matrix Completions, and GraphRealization (Page 129)

(Return to Top)

A Novel Backtracking Particle Filter for Pattern MatchingIndoor Localization

WidyawanCork Institute of TechnologyCenter for Adaptive Wireless

SystemCork, Ireland

widyawan@cit.ie

Martin KlepalCork Institute of TechnologyCenter for Adaptive Wireless

SystemCork, Ireland

martin.klepal@cit.ie

Stephane BeauregardTechnologie-Zentrum

InformatikUniversitat Bremen

28359 Bremen, Germanybogie@tzi.de

ABSTRACTParticle Filter (PF) techniques has been widely used in in-door localization systems. They are often used in conjunc-tion with pattern matching based on Received Signal StrengthIndication (RSSI) fingerprinting. Several variants of theparticle filter within a generic framework of the SequentialImportance Sampling (SIS) algorithm have been described.The purpose of this paper is to show how a variant of PF, theso-called Backtracking Particle Filter (BPF), can be used toimprove indoor localization performance.

The BPF is a technique for refining state estimates basedon exclusion of invalid particle trajectories. Categorizationof invalid trajectory determined during importance samplingstep of the PF. The BPF can also take advantage of availablebuilding plan information using the so-called Map Filtering(MF) technique. The incorporation of MF allows the BPFto exploit long-range geometrical constraints.

This paper evaluates BPF with indoor localization based onWLAN RSSI fingerprinting. The filtering schema is evalu-ated using the propagation simulation in an o!ce building,a typical environment for fingerprinting technique. Favor-able result are obtained, showing positioning performance(1.34 m mean 2D error) superior to the PF-only no MF case(1.82 m mean 2D error), or up to 25% improvement. It isalso shown that the performance is far better than the po-sition estimates from conventional Nearest-Neighbour (NN)and Kalman Filter (KF) approaches using the same RSSImeasurements.

Categories and Subject DescriptorsG.3 [Probability and Statistics]: [Probabilistic algorithms]

General TermsExperimentation

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.MELT’08, September 19, 2008, San Francisco, California, USA.Copyright 2008 ACM 978-1-60558-189-7/08/09 ...$5.00.

Keywordsindoor localization, backtracking particle filter

1. INTRODUCTIONThe Global Positioning System (GPS) has emerged as main-stream technology for outdoor localization. However, sinceGPS signals are heavily attenuated and reflected by buildingstructures, the system does not provide su!cient availabilityor accuracy indoor. Several technologies have been proposedfor indoor localization, such as UWB, infra-red sensors, in-ertial sensors (e.g. PDR), wireless sensor nodes (WSN) andwireless LAN (WLAN) [7].

The ubiquity of WLAN infrastructure makes this technol-ogy an attractive base on which to build indoor localizationsystems. The advantage of WLAN is that the technologyprovides a wireless communication infrastructure and RSSImeasurements from the transceivers can readily be used forindoor localization.

Particle Filtering (PF) has been a prominent filtering schemafor indoor RSSI-based localization. The non-linearity andnon-gaussian nature of RSSI measurements make the PFtechnique well-adapted for estimating the user position. Sev-eral variants of the PF within a generic framework of theSequential Importance Sampling (SIS) algorithm are widelyused. These are often combined with a technique called MapFiltering (MF) which can take advantage of building plansinformation. Previous research showed that WLAN/RSSI-based indoor localization provides interesting results [1]. Thetypical room-level accuracy may be su!cient for some ap-plication domains. However, if the accuracy and resolutioncan still be improved, a larger number of application do-mains could be served.

The aim of this paper is to propose a novel variant of the Par-ticle Filter, called the Backtracking Particle Filter (BPF).BPF is a technique for refining state estimates based onexclusion of invalid particle trajectories. In a previous pub-lication [10], the authors described a framework for fusingbuilding plans and PDR motion measurements with BPF fil-ter. It was shown that the BPF can take advantage of long-range (geometrical) constraint information provided by var-ious levels of building plan detail and provide high-accuracyposition estimates in many instances.

In this paper, we evaluate the BPF framework with patternmatching localization technique based on RSSI fingerprint-ing. With this pattern matching technique, the BPF cantake advantage of long-term likelihood sequences and canalso use building plan information if MF is used.

The remainder of the paper is organized as follow. In Section2, a pattern matching based positioning system is briefly in-troduced. The filtering algorithm and the novel Backtrack-ing Particle Filter will be proposed in the Section 3. Section4 will briefly describe the tools and simulation conducted.Section 5 will present the results of the experiments. Finallysection 6 will conclude the paper.

2. INDOOR LOCALIZATIONThe pattern matching indoor location system generally worksin two phases. First, there is a calibration phase and thecreation of a database of RSSI values, also known as fin-gerprints. The RSSI fingerprints are tuples consisting of anaccess point (AP) MAC address, a measured RSSI valueand a position. A tuple is created for every available APat the measurement position. Secondly, during the onlinetracking, the client’s mobile device will scan for all availableWLAN access points and measure their RSSI values. TheRSSI data in turn is input to the PF estimator to estimateuser position.

The calibration phase can be conducted manually by mea-suring RSSI at every grid point in the floor plan. While thiscalibration approach is accurate, it is expensive and timeconsuming. Re-calibration will be needed if there is a majorchange in the propagation environment, such as the reloca-tion or addition of walls, furniture, or APs.

To overcome this problem, indoor propagation models canbe used to predict the fingerprints, either entirely in simu-lation [11] using theoretical models, or alternatively usinga hybrid predict/fit approach with small number of sam-pled fingerprints, propagation models and some data fittingalgorithm [4].

There are two conventional approaches to the propagationmodeling: empirical and deterministic. The advantages ofempirical approaches are speed, simplicity of the input data(only a simple building plan is required) and straightforwardprediction formulae. The main disadvantages are poor site-specific accuracy and the impossibility of predicting wide-band communication channel parameters. In contrast, de-terministic models can be more accurate but they are con-siderably more complicated to implement [6].

In this paper, we use a simple empirical approach called theMulti-Wall Model (MWM) for fingerprint prediction [11].In the MWM, the signal loss is given with following equa-tion [9]:

LMWM = L1 + 20 log(d) + nW ! LW + nF ! LF (1)

where LMWM denotes the predicted signal loss, L1 is thefree space loss at a distance of 1m from the transmitter, dis the distance from the transmitter to the receiver, LWall

is the contribution of each of nW walls to the total signalloss, LF loor is the contribution of each of nF floors to totalsignal loss.

Figure 1: Particle Transition Near Obstacles: If aparticle tries to move across walls or other obstaclesdefined in the map, it will be killed o!.

3. FILTERING ALGORITHMParticle Filtering (PF) is a technique that implements arecursive Bayesian filter using the Sequential Monte-Carlomethod [2]. It is particularly good for dealing with non-linear and non-Gaussian estimation problems. It is basedon a set of random samples with weights, called particles,for representing a probability density. The Particle Filterdirectly estimates the posterior probability density function(pdf) of the state using the following equation [8]:

p(xt|Zt) "N

!

i=1

wt!(xt # xi

t) (2)

where xit is the i-th sampling point or particle of the poste-

rior probability and wit is the weight of the particle.

For indoor positioning, building plans are very useful infor-mation that can be used to enhance location accuracy andreduce uncertainty of walking trajectories. Particle Filterscan take into account building plan information during theindoor positioning process with a technique called Map Fil-tering [5]. Map Filtering implements a fairly straightforwardidea. New particles should not move to impossible positionsgiven the map constraints. For example, particles are notallowed to cross directly through walls. Particle that tran-sition through such obstacles are downweighted or deletedfrom the set of particles, as seen in Figure 1.

3.1 Particle Filter ImplementationDuring the prediction stage, each particle will have dynamicsaccording to a motion model that represents the estimatedobject. Particles state can be modeled with:

xi

t =

"

xit

yit

#

=

"

xit!1 + vi

tcos("it)"t + nt!1

yit!1 + vi

tsin("it)"t + nt!1

#

(3)

Where vt denotes velocity; "t describes particle heading atthe time t; nt!1 is a noise with Gaussian distribution.

In some applications, estimates of both the particle veloc-ity and heading can be obtained directly from an inertialsensor measurement. In the absence of these measurements,the particle velocity and heading are modeled through aheuristic approach. The particle velocity is given by follow-ing equations:

v = [0, 10ms!1]; vt = |N(vt!1, 1ms!2"t)| (4)

Particle heading is given by:

" = [0, 2#]; "t = N("t!1, 2# # arctan(

$vt

2)"t) (5)

The inclusion of vt in particle heading is to limit headingvariation based on particle speed (people tend to slow downwhen he/she want to change direction of walk).

The new particle position, which is determined by the mo-tion model, should not cross walls or other obstacles. Ifseveral crossing attempts fail, the particle will be catego-rized as invalid and the particle weight will also be changedaccording to the following rule:

wi

t =

$

0, crossing wall particlewi

t!1 p(zt|xit), otherwise

(6)

where p(zt|xit) is the likelihood function. In the case of pat-

tern matching localization, the likelihood function p(zt|xit)

describes the probability of receiving a set of signal leveltuples (signature) in a specific location. Example on Likeli-hood function p(zt|xi

t) formula is as follow [3]:

p(zt|xi

t) =1$2#$

exp

%

#(Xzt

# Xxit)2

2$2

&

(7)

with Xztbeing the position returned by the database, Xxi

t

the position of the ith particle at time step t, and $ the mea-surement standard deviation. Xzt

can be determined fromNN algorithm, which calculate minimum signature distance(the minimum signal space distance between RSSI measure-ment and RSSI in fingerprint database) [1]:

d" = argmin(!

|Td # Tz|) (8)

where d" being the minimum signature distance, Td is atuple from the database and Tz is a tuple from received RSSImeasurement. Figure 2 shows typical likelihood function ofRSSI measurement of an access point in open space.

Figure 2: Typical likelihood function of RSSI mea-surement in open space, an access point is placed inthe middle of the horizontal plane

3.2 Backtracking Particle FilterBacktracking Particle Filter is a technique for refining stateestimates based on particle trajectory histories. The incor-poration of the Map Filtering technique allows the BPF toexploit the long-range geometrical constraints of a buildingplan and long-term likelihood function weighting. If someparticles xi

t are not valid at some time t, the previous stateestimates back to xt!k can be refined by removing the in-valid particle trajectories. This is based on assumption that

an invalid particle is the result of a particle that follows aninvalid trajectory or path. Therefore, recalculation of theprevious state estimation xt!k without invalid trajectorieswill produce better estimates. In order to enable backtrack-ing, each particle has to remember its state history or tra-jectory.

(a) Detecting the invalid particles

(b) Backtracking the invalid trajectories

(c) Backtracking the estimated states

Figure 3: BPF for pattern matching localization

The BPF implementation for pattern matching localizationis illustrated in the following figures. Figure 3(a) shows atypical phenomenon when a standard Particle Filter is usedfor pattern matching indoor localization. It illustrates pos-terior density of particles in four time steps. The positionestimates and the ground truth are shown in the image aswell.

Map Filtering and the likelihood function categorize someparticles as invalid between the 3rd and 4th step and the in-valid particles are not subsequently resampled. Figure 3(b)shows how the Backtracking Particle Filter removes the in-valid trajectories. Figure 3(c) illustrates the recalculatedstate estimates after backtracking. It can be seen that undercertain conditions backtracking can improve state estimatesrelative to a normal PF.

The pseudocode below describes the complete BPF algo-rithm for state refinement.

Backtracking-PF(N, tail)

1 sampling N particles from initial pdf2 tailcount % 03 repeat4 get zt

5 for i % 1 to N6 do get xt

i from p(xt|xt!1)7 calculate wi

t = p(zt|xit)

8 for i % 1 to N9 do normalize wi

t = wit/

'

N

i=1wi

t

10 resample and inherit state history11 estimate state xt = 1

N

'

N

i=1xi

t

12 if tailcount & tail13 then xt!tail = 1

N

'

N

i=1xi

t!tail

14 increment tailcount15 increment t16 until stop

The main features of the BPF for pattern matching local-ization can be seen in steps 6,10, and 13 of the pseudocode.During prediction sampling in step 6, a new particle is sam-pled from the transition pdf p(xt|xt!1) with the Map Filter-ing technique. In step 10, the resampling is followed by theinheritance of the state history. The resampling algorithmis taken from [8]. The inheritance step made the newly sam-pled particle xi

t

"

will inherit its parent state (position andheading) Xi

t!1 ! {xi, i = 1, 2, 3, ...t # 1}. This inheritancestep will enable the backtracking of invalid trajectories andalso the calculation of the backtracking state (step 13).

4. TOOLS AND SIMULATIONSimulations were run using the floor plan of a small, onestorey o!ce building, measuring approximately 50m x 50m(2,500m2). The fingerprint database, i.e RSSI measure-ments for the entire simulation area, was estimated usingthe MWM. A ground truth path, approximately 150 metersin length was drawn on the floor plan through the o!cecorridors, rooms and outdoor area. A nominal walking be-havior was generated by advancing the current location of asimulated user along a ground truth path at a rate of 1 m/sat each simulation time step (1 second).

Noisy RSSI scan measurements were generated for each sim-ulation time step by adding Gaussian noise with standarddeviation of 5 dBm to the nominal, noise-free RSSI values forcurrent position from fingerprint DB. The simulated RSSIvalues were then used as input to a probabilistic positionestimating routines, which was then used as input to thePF and BPF applications, both of which were implementedin C++. One thousand particles were used during the fil-tering. For comparison purposes, position estimates werealso calculated using the same noisy, simulated RSSI val-ues with conventional nearest neighbour and Kalman filterapproaches.

5. RESULTS AND ANALYSISAs was to be expected, the conventional nearest-neighbourestimation method shows a large position scatter all along

the simulated path. The primary reason for this relativelypoor performance is that it is a memory-less algorithm andconsequently cannot filter out any noise in the estimates.The Kalman filter, which try to refine NN estimation, hasbetter trajectory. It can filter out noise in successive mea-surements, and can incorporate a motion model. See Figure4(a) and Table 1 for details.

The BPF that can take advantage of long-term likelihoodfunction and long-range (geometrical) constraint informa-tion yields excellent positioning performance (1.34 m mean2D error), it shows enhancement up to 25% compare to PFonly (1.82 m mean 2D error) and more than 3 times bettercompare to NN performance. More significantly, the BPFwithout MF yields improved positioning performance (1.62m mean 2D error) compare to a PF-only (1.82 m mean 2Derror). This result confirms that BPF can be performedvia the elimination of trajectory errors based on likelihoodfunction. The positioning accuracy is summarized in Table1.

Table 1: Positioning AccuracyNN KF PF BPF

without MF µ =4.37 µ =3.57 µ =1.82 µ =1.62$ =4.66 $ =2.90 $ =1.50 $ =0.97

with MF - - µ =1.67 µ =1.34$ =1.40 $ =0.80

The trajectory evolution over time is shown in 4(b) for es-timation without MF. It can be seen that PF trajectory ismore jagged than BPF. Trajectories of filtering with MF 4(c)are better since they are constrained by the building walls.

The tail value of the BPF is established empirically. Thevalue is optimized by considering several parameters, mostnotably building plans dimension and trial duration. Thesimulation is performed with the tail value of 5 and below.While using bigger tail values are possible as shown in [10],smaller tail has more practical significance for real-time lo-calization.

6. CONCLUSIONIn this paper a novel Backtracking Particle Filter algorithmis proposed. The filter is evaluated with pattern matchinglocalization.

It has been shown that BPF with building constraint in-formation yields excellent positioning performance (1.34 mmean 2D error). It shows enhancement up to 25% compareto PF only (1.82 m mean 2D error). More significantly, theBPF without MF yields improved positioning performance(1.62 m mean 2D error) relative to a PF-only. This resultshow that BPF can be performed via the elimination of tra-jectory error based on likelihood function.

It is expected that this performance can be reproduced formany other environment and data encountered during in-door localization. Further experiments will be performed inthe future to test this hypothesis.

(a) Trajectory of Nearest-Neighbour and Kalman Filter

(b) PF and BPF Trajectory without Map Filtering

(c) PF and BPF Trajectory with Map Filtering

Figure 4: Trajectory for di!erent methods

7. ACKNOWLEDGMENTThe authors wish to acknowledge the funding support ofthe Enterprise Ireland under the Proof of Concept programPC/2007/0073 for the work reported in this paper.

8. ADDITIONAL AUTHORSAdditional authors: Dirk Pesch (Cork Institute of Technol-ogy, email: dpesch@cit.ie)

9. REFERENCES[1] P. Bahl and V. Padmanabhan. Radar An in-building

rf-based user location and tracking system. InProceedings of the IEEE INFOCOM 2000, pages775–784, March 2000.

[2] A. Doucet, N. de Freitas, and N. Gordon, editors.Sequential Monte Carlo Methods in Practice. Statisticsfor Engineering and Information Science. Springer,New York, 1 edition, June 2001.

[3] F. Evennou and F. Marx. Advanced integration of wifiand inertial navigation systems for indoor mobilepositioning. EURASIP Journal on Applied SignalProcessing, 2006:1–11, January.

[4] B. Ferris, D. Hahnel, and D. Fox. Gaussian processesfor signal strength-based location estimation. InProceedings of Robotics Science and Systems,Philadelphia, USA, August 16 2006.

[5] P.-Y. Gillieron, I. Spassov, and B. Merminod. IndoorNavigation Enhanced by Map-Matching. EuropeanJournal of Navigation, 3(3), 2005.

[6] M. Klepal. Novel Approach to Indoor ElectromagneticWave Propagation Modelling. PhD thesis, Departmentof Electromagnetic Field, Czech Technical University,2003.

[7] H. Liu, H. Darabi, P. Banerjee, and J. Liu. Survey ofwireless indoor positioning techniques and systems.IEEE Transaction on Systems, Man, and Cybernetics,37(6):1067–1077, November 2007.

[8] B. Ristic, S. Arumpalampam, and N. Gordon. Beyondthe Kalman Filter: Particle Filters for TrackingApplications. Artech House Radar Library. ArtechHouse Publishers, February 2004.

[9] C. Telecommunications. Cost231 final report, digitalmobile radio: Cost231 view on the evolution towards3rd generation systems. Technical report, EuropeanCommission/COST Telecommunications, Brussel,1998.

[10] Widyawan, M. Klepal, and S. Beauregard. Abacktracking particle filter for fusing building planswith pdr displacement data. In Proceedings of the 5thWorkshop on Positioning, Navigation andCommunication (WPNC 2008), Hannover, Germany,March 27 2008.

[11] Widyawan, M. Klepal, and D. Pesch. Influence ofpredicted and measured fingerprint on the accuracy ofrssi-based indoor location systems. In Proceedings ofthe 4th Workshop on Positioning, Navigation andCommunication (WPNC 2007), Hannover, Germany,March 22 2007.