Applications & Theory Azadeh Kushki azadeh.kushki@ieee.org Professor K N Plataniotis Professor K.N. Plataniotis Professor A.N. Venetsanopoulos
Presentation Outline 2 Part I: The case for WLAN positioning History Applications Overview & challenges Part II: Theory Memoryless positioning Tracking A cognitive design
Part I: The Case for WLAN Positioning i i History Applications Overview & challenges
Positioning 4 Objective: Determine physical coordinates of a mobile terminal Historical perspective: Studied d widely for the past five decadesd Limited to military/civilian target tracking & navigation Renewed interest: mobile computing
Mobile Computing 5 Motivated by advances in wireless that allow computing anywhere, anytime Mobility has led to new needs Location-dependent resource & information needs Mobility has sparked new applications Location-based services
Location-Based Services 6 Network Provider Location server Location-based management Proactive resource deployment Radio/map information Position information Authentication Resource allocation Social networking Friend Finder services Location metadata Content Provider Geo-tagging/geo-blogging User generated content t Content server Location-based information
Location-Based Services 7 300 250 Number of subscribers (Millions) 200 150 100 50 Revenue in $ (Billions) 0 8 7 6 5 4 3 2 1 0 2007 2008 2011 2007 2008 2011 Figures obtained from Gartner.
Positioning Technology 8 Motivation: To enable location-based service, accurate and timely location information is needed Example technologies: Global positioning system Cellular-based methods In this talk, we focus on positioning in indoor environments
Indoor Positioning 9 Motivation: GPS & cellular systems provide limited coverage in indoors Objective: Determine physical coordinates of a pedestrian carrying a wireless device in an indoor environment
Indoor Positioning Solutions 10 Technology Accuracy Cost Complexity Invasive RFIDs <10m Medium Low Yes Visual centimeters High High Yes surveillance Radio (WLAN) <10m Low Low No tracking
WLAN Tracking: Basic Idea 11 WLAN radio signal features depend on distance between receiver & transmitter Measure signal features to determine location Time of Arrival Require additional Time dff difference of Arrival hardware Angle of Arrival Received Signal Strength th (RSS)
RSS-Based Tracking: Motivation 12 Inexpensive No additional hardware needed Scalable Ubiquitous deployment Non-invasive Requires cooperation of mobile device
The Setup 13 Pedestrian carries a WLAN-capable device L access points Unknown positions r 2 ( k) 3 r 3 ( k) r 4 ( k) 1 ( k r ) r L (k) Mobile measures RSS vector at time k 1 r( k ) = [ r ( k), L, r L ( k)] T
The Problem 14 Given a sequence of RSS measurement over time R( k) = { r(1), L, r( k)} Estimate a sequence of position estimates pˆ (1), L, pˆ( k)
Technical Challenges 15 Functional form of RSS-position relationship generally unknown Severe multipath, shadowing Propagation models insufficient i to describe spatial variations A. Kushki, K.N. Plataniotis, A.N. Venetsanopoulos, Indoor Positioning with Wireless Local Area Networks, in the Encyclopedia of Geographical Information Sciences, 2008.
Technical Challenges 16 RSS measurements depend on unpredictable environmental factors Moving people, doors, humidity, etc. RSS measurements vary over time at fixed locations Variations do not obey well-known n distributionstions
Location Fingerprinting 17 Characterize RSS-position dependency through training-based i method Construct a radio map y R = { ( p F( ( p ) ), L, ( p, F ( p ) ) } 1, 1 N N p i x i y i T = [ p p ]: anchor point N: Number of anchor points [ (1) L ( ) ] F ( p i ) = r i r i n : fingerprint matrix x n: Number of RSS samples per anchor points
Outline of Solutions 18 Kernel density estimation for fingerprinting- based positioning i Nonparametric Information Filter Improve positioning accuracy by incorporating knowledge of pedestrian motion dynamics Cognitive design to deal with unpredictable RSS Cognitive design to deal with unpredictable RSS variations through sensor selection
Part II: Theory Memoryless positioning ii i Tracking A cognitive design Conclusion and future work
Memoryless Positioning 20 Objective: given an RSS measurement, determine a position estimate t r(k)? p(k) ˆ( Radio map Optimality criterion: minimum mean square error (MMSE) ( 2 pˆ ( k ) = arg min p ( E { p ( k ) p ( k ) }
MMSE Estimation 21 MMSE estimate is given as p ˆ( k) = E{ p( k) r( k)} = p( k) f ( p( k) r( k) ) dp( k) unknown Approximate the posterior density Histogram Kernel density estimator pˆ( k) N i = 1 N i= 1 w p A. Kushki, K.N. Plataniotis, A.N. Venetsanopoulos, Kernel-based Positioning in Wireless Local Area Networks, IEEE Transactions on Mobile Computing, 6(6), pp.689-705, 2007. i w i i
Memoryless MMSE Estimator 22 Radio map RSS observation RSS rep extraction r 1 (1) r r 1 1 ( n ) 1 n w 1 p 1 r N (1) r N (n) RSS rep extraction r N KDE wnp N + pˆ r ( k) Temporal processing Spatial processing n: Number of RSS samples per anchor points N: Number of anchor points w i = N ( r k); r, ) ( i Σ r
Performance Evaluation 23 Evaluation data collected in a real office RSS measured using public software on a laptop 46m 42m Performance measure: root mean square positioning error
Test Conditions 24 Capture environmental variations Training & testing sets collected on different days Orientation mismatch Two motion scenarios considered Stationary user 352 test cases (44 locations) Mobile user Mobile user 34 paths
Experimental Results 25 Method Stationary user (Average RMSE) Mobile user (Average RMSE) Complexity KNN 3.18m 5.85m O(dN) Histogram 3.22m 5.68m O(bdN) Kernel Density 2.90m 5.70m O(dN) b: Number of histogram bins d: Number of access points n: Number of RSS samples per anchor points N: Number of anchor points
Part II: Theory Memoryless positioning ii i Tracking A cognitive design Conclusion and future work
Tracking 27 Objective: given the RSS observation record, determine positioning i estimates t over time Dynamic model R ( k ) = { r(1), L, r ( k )} pˆ (0), L, pˆ( k 1)? Radio map p( ˆ( k ) Exploit knowledge of pedestrian motion dynamics to refine RSS-based estimates
Tracking 28 Traditional approach: Bayesian filtering Estimate the hidden state of system given observable RSS measurements Kalman filter & extensions, particle filter Challenge: Lack of an explicit relationship between RSS & positions Computational complexity A. Kushki, K.N. Plataniotis, A.N. Venetsanopoulos, Nonparameteric Techniques for Pedestrian Tracking in Wireless Local Area Networks, to appear in the Handbook on Sensor and Array Processing.
Bayesian Filtering: State Vector 29 Contains all variables needed to describe the evolution of the state t of a system In general, many parameters needed to describe pedestrian motion Simplifying assumption: In indoor office spaces Simplifying assumption: In indoor office spaces, movements constrained by physical structure
The State Vector 30 Assuming linear motion, define the state vector x x y y T x( k ) = [ p ( k) v ( k) p ( k) v ( k) ], where x y [ p ( k ) p ( k )] is pedestrian coordinates at time k [ v x ( k) v y ( k)] is pedestrian velocity at time k The dynamic model is x ( k + 1) = Fx( k) + ω( k) Initial state: x ( 0) ~ N ( x, P 0 0), System matrix: F, System noise: ω ( k ) ~ N (0, Q ).
MMSE Tracking 31 MMSE estimate of the state is defined as xˆ( k) = arg min ( E{ x( k) x ( x( k) 2 } MMSE estimate is given as x ˆ( k k) = E{ x( k) R( k)} ( x( k) R( k) ) = x( k) f dx( k) unknown
Bayesian Filtering 32 Estimate the posterior density recursively in two steps Prediction Correction Estimate at k-1 prediction Predicated estimate at k correction Estimate at k Dynamic model Measurement model RSS observation
Bayesian Filtering: Prediction 33 Use the dynamic model to predict the state given the previous estimate t xˆ ( k 1 k 1) xˆ( k k 1) Since a linear-gaussian dynamic model is assumed, prediction is the same as traditional Kalman filteringi
Bayesian Filtering: Correction 34 Use measurements to refine predicted estimate Requires measurement model that relates RSS observations to the state Explicit measurement model not available in fingerprinting! The Nonparametric Information (NI) Filter A. Kushki, K.N. Plataniotis, A.N. Venetsanopoulos, Cognitive Dynamic Radio Tracking in Indoor Wireless Local Area Networks, submitted to the IEEE Transactions on Mobile Computing.
The Nonparametric Information Filter 35 Radio map RSS observation x ˆ( k 1 k 1) P( k 1 k 1) Memoryless estimator Dynamic model x ˆ ( k ) (k ) r P r ) ) x ˆ ( k k 1) P ( k k 1) NI filter x ˆ( k k) P( k k) P 1 x( k ( k k) = P k) = P( k 1 ( k k) k 1) + P ( k) ( r 1 P ( k k 1)ˆ( x k k 1) + P ( k)ˆ x ( k) ) ˆ 1 1 r r
Experimental Results 36 Method Stationary user (Average RMSE) Mobile user (Average RMSE) Complexity Memoryless 2.90m 5.70m O(dN) Kalman filter 2.75m 5.41m O(dN) Particle filter 2.44m 5.16m O(dNNpart) NI filter 2.29m 4.58m O(dN) d: Number of access points N: Number of anchor points Npart: Number of particles (Npart =1000) All filters use same memoryless estimator All filters use same motion model
Part II: Theory Memoryless positioning ii i Tracking A Cognitive design Conclusion and future work
A Cognitive Design 38 Motivation: NI filter builds its knowledge of the environment through RSS observations & radio map Conditions during tracking may be different than those learned from fingerprints Objective: Mitigate adverse effects of unpredictable Mitigate adverse effects of unpredictable environmental variations
A Cognitive Design 39 Basic idea: Proactively adapt sensing and estimation parameters based on predicated operating conditions Approach: adaptive radio scene analysis Anchor point selection RSS-position relation is many-to-many Access point selection Number of available access points >>3
Adaptive Radio Scene Analysis 40 Determine region of interest (ROI) using feedback Use only anchor points in ROI for positioning Evaluate access point selection criterion i over ROI A. Kushki, K.N. Plataniotis, A.N. Venetsanopoulos, Cognitive Dynamic Radio Tracking in Indoor Wireless Local Area Networks, submitted to the IEEE Transactions on Mobile Computing.
The Cognitive Design 41 Anchor point selection Memoryless estimator Outlier Mitigation State Prediction Access Point selection State Estimation Position Estimate Adaptive Scene Analysis NI filter Radio map RSS observation Two levels of feedback Local (NI filter) Global l (Scene analysis)
Experimental Results 42 Method Stationary user Mobile user Complexity (Average RMSE) (Average RMSE) Memoryless 2.90m 5.70m O(dN) NI filter 2.29m 4.58m O(dN) NI filter + anchor point selection 2.31m 3.96m O(dN ) NI filter + anchor point selection + access point selection 2.07m 2.51m O(dN ) d : Number of access points N : Number of anchor points N : Number of selected anchor points (N <N)
Example 43
Part II: Theory Memoryless positioning ii i Tracking A cognitive design Conclusion and future work
Conclusions 45 Location-based services (LBS) emerging area with significant ifi commercial impact WLAN positioning is an enabling technology for indoor LBS Inexpensive & scalable Accuracy limited it by quality of propagation channel Use of motion dynamics, sensor selection
Future Directions 46 Fusion of multiple technologies to provide reliable positioning i in indoor/outdoor environments GPS, radio, video Privacy security an anonymity in positioning Privacy, security, an anonymity in positioning systems
Related Publications 47 A. Kushki, K.N. Plataniotis, "Nonparametric Techniques for Pedestrian Tracking in Wireless Local Area Networks", to appear in Handbook on Sensor and Array Processing, S. Haykin and K.J.R. Liu, Eds., IEEE-Wiley, 2009. A. Kushki, K.N. Plataniotis, A. N. Venetsanopoulos, "Indoor Positioning with Wireless Local Area Networks (WLAN)", in the Encyclopedia of Geographical Information Science, S. Shekhar and H. Xiong, Eds., Springer, pp.566-571, 571 2007. A. Kushki, K.N. Plataniotis, A.N. Venetsanopoulos, "Kernel-based Positioning in Wireless Local Area Networks", IEEE Transactions on Mobile Computing, 6(6), pp.689-705, 2007. A. Kushki, K.N. Plataniotis, and A.N. Venetsanopoulos, "Sensor Selection for Mitigation of RSS-based Attacks in Wireless Local Area Network Positioning", in the proceedings of the International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 2065-2068, 2008.
Thank You. Contact: t azadeh.kushki@ieee.org hki@i