SCPL: Indoor Device-Free Multi-Subject Counting and Localization Using Radio Signal Strength Chenren Xu†, Bernhard Firner†, Robert S. Moore∗, Yanyong Zhang† Wade Trappe†, Richard Howard†, Feixiong Zhang†, Ning An§ †WINLAB, Rutgers University, North Brunswick, NJ, USA ∗Computer Science Dept, Rutgers University, Piscataway, NJ, USA §Gerontechnology Lab, Hefei University of Technology, Hefei, Anhui, China IPSN 2013 About This Paper • Indoor localization technique – RF-based device-free passive localization – Fingerprinting based approach – Count and track multiple subjects • Result – Counting accuracy: 86% – Localization accuracy: 1.3m Contributions • The first work to simultaneous counting and localizing – Up to 4 objects – Only using RF-based technique • Relying on data collected by single subjects • Trajectory constraints to improve tracking accuracy • Recognize the nonlinear fading effects – Cause by multiple subjects Problem Formulation • Partition into K cells • Training phase – Measure ambient RSS value for L links – A single subject appear in single cell (randomly walk within cell) • Take N measurement for L links • Subtract ambient RSS • Dataset D: K * N * L matrix – Subject’s present in Cell i: State Si • DS1, DS1, DS1 ,……, DSk Problem Formulation • Testing phase – Measure ambient RSS for L links – A subject appears in random cell • Measure RSS for all L links • Subtract ambient • Form an RSS vector O • Compare D and O – Classification algorithm Outline • • • • • Counting multiple subjects Localizing multiple subjects Experimental setup and result Limitation Conclusion Impact of Multiple Subject • Hypothesis: more subjects – Not only affect more links – But also higher level of RSS change • Infer the number of subjects by RSS change – Total energy change: – Absolute RSS mean difference • Distance between subjects – Distance > 4m faraway – Else closeby Counting Subjects • Successive cancellation – In each round, estimate the strongest subject’s cell number – Subtract it share of RSS change • If (Impact from multiple subjects is linear) – Subtract the mean vector • But the impact is Nonlinear – Need an coefficient Location-Link Coefficient Matrix • For each link, calculate the correlation between a cell pair (i,j) ℎ𝑙 ij • Coefficient Matrix • When two cell close to each other – High correlation • When only one cell affect link l – Low correlation Successive Cancellation • Constructing upper and lower bound • Iteration 1. 2. If (energy change < C0 upper bound) count = 0 Presence detection 1. If (energy change >= C1 upper bound) 1. 2. 3. Estimate the occupied cell Contribution Substracting 1. 5. Else (goto End) Cell Identification 1. 4. Increment count by one, goto next Substracting from O End 1. 2. If (remained energy change < C1 upper bound) Increase count Outline • • • • • Counting multiple subjects Localizing multiple subjects Experimental setup and result Limitation Conclusion Conditional Random Field Formulation • Transition model • Define – Cell neighbors: adjacent cells which can be entered – Order of Neighbor: neighbor distance – Trajectory ring: • Radius r: area consist of up to r-order neighbors • Let Ω𝑟(i) be the cells in i’s r-trajectory • Nr(i) be the size of Ω𝑟(i) , thus Localization Algorithm • Viterbi algorithm: find highest probably path • Denote Q = {q1,…,qc}, C is total number of subjects 𝐾 • For current state Qt, permutation 𝐶 • For each permutation, compute Viterbi score Outline • • • • • Counting multiple subjects Localizing multiple subjects Experimental setup and result Limitation Conclusion Experiment Setup • CC1100 transceiver – 909.1MHz – Broadcast 10-byte packet every 0.1s • RSS collected as a mean value over 1s • Training phase: 30s in each cell • Performance metrics – Counting percentage – Error distance Office environment – 13 transmitter, 9 receiver – 150 m^2, divided into 37 cell – Movement scenarios Counting Percentage Location-Link Coefficient Counting Result Localization Result Open Floor Space • 12 transmitter, 8 receiver • 400 m^2, 56 cells • Movement scenarios Location-Link Coefficient Counting Result Localization Error Outline • • • • • Counting multiple subjects Localizing multiple subjects Experimental setup and result Limitation Conclusion Limitation • Computation complexity – 0.87s and 0.88s for 4 objects – More that 1s for 5 objects or above • Long-term test – Suffer from environmental change – Fingerprint aging Conclusion • • • • • Device free localization system Track multiple subjects Average 86% counting accuracy ?? Average 1.3m localization accuracy ?? Test in two different environments – How many iteration? • Not very successful with more objects