#### Transcript Maximum Likelihood Energy Based Acoustic Source Localization

Sequential Acoustic Energy Based Source Localization Using Particle Filter in a Distributed Sensor Network Xiaohong Sheng, Yu-Hen Hu University of Wisconsin – Madison Dept. Electrical and Computer Engineering [email protected] http://www.ece.wisc.edu/~sensit/ 1 Outline • Wireless Sensor Network – New features of recent sensor devices – Applications – Acoustic Source Localization and Tracking Problems • Available algorithms • Our approach • Source Localization using particle filtering in sensor network – Particle filtering framework – System model – Measurement model • Energy decay model • Cooperate ML Algorithm with particle filtering • Apply particle filter into a distributed framework • Experiments and Simulation • Conclusion 2 Sensor Network • New sensor nodes – Integrating micro-sensing and actuation – On-board processing and wireless communication capabilities – Limited communication bandwidth – Limited power supply • Provides a novel signal processing platform – Detection, classification – Localization, tracking etc Sitex 02 experiment sensir field 3 Localizing and Tracking Targets in Distributed Sensor Network 1200 m Northern Checkpoint ~1300 m Defile Sandia Autonomous Mobile Robot OpArea 1000 m 800 m 100baseT Hardwire Experiment Control Ethernet 600 m Gateway/Imager 400 m RF Ethernet 200 m 25 Nodes @ Intersection Eastern Checkpoint ~500 m//intersection Intersection Western Checkpoint ~ 400 m//intersection Base Camp ~300 m//intersection 4 UWCSP: Univ. Wisconsin Collaborative Signal Processing Node Detection Node Classification • Received Det./Classify report from nodes • N Fault Tolerance Check – Energy Detection – Node target classification Y N Region Detection Fusion Y Classification fusion Target type Current track DB Energy-based localization Target location Track associate N Current target? Create new track Distributed Signal Processing Paradigm (Local) Node signal processing • (Global) Region signal processing – Region detection and classification fusion – Energy based localization – particle filter tracking – Hand-off policy Y Update track and predict Send info to next region Y Handoff N 5 Source Localization and Tracking in wireless Sensor Network • Available Localization and Tracking method – Localization Estimation Modeling • CPA, Beamforming, TDOA – Tracking Method • Sequential Bayesian Estimation – Kalman Filtering, Extended Kalman Filtering – Grid-Based Bayesian Estimation –Exhaustive Search • Our Approach – Previously • Intensity Based Source Localization • ML estimation and Non-Linear estimation – This Paper • Particle Filtering cooperated with ML estimation • Distributed Framework 6 Outline • Wireless Sensor Network – New features of recent sensor devices – Applications – Acoustic Source Localization and Tracking Problems • Available algorithms • Our approach • Source Localization using particle filtering in sensor network – Particle filtering framework – System model – Measurement model • Energy decay model • Cooperate ML Algorithm with particle filtering – Apply particle filter into a distributed framework • Experiments and Simulation • Conclusion 7 System Model for tracking vehicle in sensor field • System Model: a k (t ) w(t ) u k (t ) u k (t 1) a k (t )T 1 ρ k (t ) ρ k (t 1) u t (t 1)T a k (t )T 2 2 • State Vector for source k at time t is: k αt ρk (t ) u k (t ) a k (t ) where: a t (t ) : Acceleration of the source k at time t ρ t (t ) : Velocity of the source k at time t u t (t ) : Location the source k at time t T: Time Interval between two consecutive computation 10 Measurement Model-Acoustic Delay Function • Source Energy attenuates at a rate that is inversely proportional to the Square of the distance to the source • Energy Received by each Sensor is the Sum of the Decayed Source Energy K sk t yi t ys t i t gi i t 2 k 1 ρ t r k i – – – – – gi: gain factor of ith sensor sk(t): energy emitted by the kth source k(t) Source k’s location ri: Location of the ith sensor i(t): sum of background additive noise and the parameter modeling error. – K: the number of the sources 11 Measurement Model-Notation • Let d ij t ρ j t ri be the Euclidean distance xt between sensor i and target j, and 1 d 2 (t ) 11 1 D t d 2 (t ) 21 1 d N2 1 (t ) • Also define y t 1 t 1 1 (t ) y2 t 2 (t ) 2 (t ) y N (t ) N (t ) N (t ) st s1 (t ) s2 (t ) sK (t ) 1 1 g1 gN g2 2 2 d12 (t ) d1K (t ) G t diag , , , N (t ) 1 (t ) 2 (t ) 1 1 2 and d 22 (t ) d 22K (t ) Ht G t Dt 1 1 • Then, the energy attenuation 2 model can be represented as: d N2 2 (t ) d NK (t ) xt G t Dt st ξ t Ht st ξ t ~ N Ht st , I 12 T Cooperating ML estimator with Particle Filtering • Measurement Likelihood for given estimated target locations: ln P(z t | θt ) z t H t s t Γ z t H t s t xTt Pt x t – where θt ρ1 (t ) ρ K (t ) s1 (t ) s K (t ) st Ht xt ; Pt H t : a function of 1:K (t ) 1 T H Ht Ht Ht Ht Therefore: Unknown Parameters p : Projection matrix * x k | k (i ) e xTk Pk ( k* ( i )) x k Need at least K(p+1) sensors, p is the dimension of the location Nonlinear Problem 13 Particle Filter in Distributed Framework Layer 2 sub-region Layer 1 sensor Region Layer 1 Manager node Layer 2 Manager Node Layer 2 Detection Node 14 Distributed Particle Filter-Node Function • Layer 2 Detection Node – BroadCast with Lower Transmission Power • Layer 2 Manager Node – – – – – Encode the data received from its layer 2 detection node BroadCast with higher Transmission Power Distributed Particle Filter Encode Particles Send to Manager Node • Layer 1 Manager Node – Pear to Pear Transmission with the highest Transmission Power, – But only when it predicts the targets will move to its neighboring sensor region 15 Outline • Wireless Sensor Network – New features of recent sensor devices – Applications – Acoustic Source Localization and Tracking Problems • Available algorithms • Our approach • Source Localization using particle filtering in sensor network – Particle filtering framework – System model – Measurement model • Energy decay model • Cooperate ML Algorithm with particle filtering – Apply particle filter into a distributed framework • Experiments and Simulation • Conclusion 16 Application to Field Experiment Data • Sensor Field is divided into two sensor region, i.e., Region 1 and Region 2 • For region 1, Node 1 is manager node, others are detection nodes • For region 2, Node 58 is manager node, others are detection nodes Sensor deployment, road coordinate and region specification for experiments 17 Localization Results (Comparison of ML and Particle Filtering ) 18 Simulation Results for Multiple Targets Tracking • Tracking two targets moving in opposite direction • Bigger random noise are added at random time 19 Future Work – Conclusion • Develop an energy-efficient, band-width efficient, practically applicable, accurate and robust source localization method. • The algorithm can be incorporated in a wireless sensor network to detect and locate multiple sound sources effectively. • The algorithm is activated on demands • The algorithm can be fit into the distributed sensor network framework. – Future Work • Integration EBL with sub-array beam-forming • Distributed Propagating Parameters In Stead of Encoded Particles • Find a better way of brief and state propagating 20 The End http://www.ece.wisc.edu/~sensit/ Thanks 21 Experiments • • • • • Experiment was carried out in Nov. 2001, Sponsored by DARPA ITO SensIT project at 29 Palms California, USA Sensor nodes are laid out along side a road Each sensor node is equipped with 1200 m Northern Checkpoint ~1300 m Defile Sandia Autonomous Mobile Robot OpArea 1000 m – acoustic, seismic and Polorized infrared (PIR) sensors, – 16-bit micro-prcessor, – radio transceiver and modem. 800 m Sensor node is powered by external car battery Military vehicles were driven through the road. 200 m 100baseT Hardwire Experiment Control Ethernet 600 m Gateway/Imager 400 m RF Ethernet 25 Nodes @ Intersection Eastern Checkpoint ~500 m//intersection Intersection Western Checkpoint ~ 400 m//intersection Base Camp ~300 m//intersection – AAV ( Amphibious Assault Vehicle), – DW ( dragon wagon) • Sampling rate : 4960 Hz at 16-bit resolution 22 Significance • Our localization and tracking algorithm will partially address the limitations of the existing algorithms: – Robust to unknown and unexpected disturbance • • • • • Background noise, Interference signals Wind gust, Faulty and drifting sensor readings Failures of sensor nodes and wireless communication network – Less Strict Requirement of Synchronization – Feasible to localize multiple targets 23 Distributed Particle Filter-Node Function • Layer 2 Detection Node – BroadCast with Lower Transmission Power – BroadCast with Delayed Time dt 1 SNR • Layer 2 Manager Node – – – – Forward received data with higher transmission power Distributed Particle Filtering Encode Particles Send encoded particles to Manager Node • Layer 1 Manager Node – Pear to Pear Transmission with the highest Transmission Power, – But only when it predicts the targets will move to its neighboring sensor region 24 Distributed Particle Filter • Parallel Run Particle Filtering at each Layer 2 Manager Node M Layer 2 sub-region Layer 1 sensor Region L M=4, L=2 25 Distributed Particle Filtering • ith Layer2 manager node: – Calculate the number of particles at its sub-region with refined grids, total M2 • Nik, k=1,2,…M2 – Calculate the number of particles at the other sub-region, • Pj, j=1,2,…L2, ji, • Manager Node decode: – For location belongs to sub-region I • Each grid k N i' N ik nik N M2 L2 k 1 j 1, j i N ik Pj N nik – Target Location, rˆ y i 1 k 1r r 2 rˆx i 1 k 1rx rx 2 y y y L i 1 k 1 i 1 k 1 L L2 M 2 R x L2 M 2 R 26 Distributed Particle Filtering • Encoding Particles Layer 2 ID Number of particles occurs in Location ID at the other Layer 2 sub-region Layer 2 Region ( L 1) log 2 N / L2 log 2 L2 log 2 M 2 Number Occurs at the corresponding Location log 2 N / L2 • Maximum Bits Required for Transmission ( L2 1 M 2 )(log • Resolution: r 2 L2 log 2 M 2 log 2 N 2 L ) Rs M 2 L2 – where: • L2: the number of layer 2 • M2: the number of grids at layer 2 • N: the number of total particles used for particle filtering • Rs: Region Size – For N=512, M=4,L=2, Rs=64, R<247 Bits/T, r=8 – For N=512, M=2, L=2, Rs=64, R<77 Bits/T, r=16 27