1. Detection, Classification and Tracking in Distributed Sensor Networks D. Li, K. Wong, Y. Hu and A. M. Sayeed Dept. of Electrical & Computer Engineering University of Wisconsin-Madison Akbar@engr.wisc.edu Akbar@engr.wisc.edu http://dune.ece.wisc.edu/

2. Overview -Small, densely distributed wireless sensors. -Collaboration necessary for tracking and classification, but not for detection. -Multi-modal sensors (potentially) -Tracking vehicles (tanks, trucks).

3. Space-Time Sampling Sensors sample the spatial signal field in a particular modality (e.g., acoustic) –Sensor density commensurate with spatial signal variation Sampling of time series from each sensor commensurate with signal bandwidth Signal field decomposed into space-time cells to enable distributed signal processing Time Space Time Space A moving object corresponds to a spatial peak moving with time Target tracking corresponds to determining the peak location over time Uniform space-time cellsNon-uniform space-time cells

4. Detection Simple energy detector –detect a target/event when the output exceeds a threshold –Otherwise, update the threshold. Detector output: – at any instant is the average energy in a certain window is sampled at a certain rate based on a priori estimate of target velocity

5. Target Localization at a Time Instant CPA/Energy based: 1.Location of the sensor with largest output 2.Using attenuation exponent and 4 or more sensor measurements.

7. Single Target Tracking Initialization: Put all locations where vehicle can enter (in this case the corners) onto detection alert. Three-step procedure: 1.A track is initiated when a target is detected in a region 2.Target locations are estimated for N successive time instants. These positions are used to predict target location at M<N future instants 3.Predicted positions are used to create new regions that are put on detection alert

8. Multiple Targets Can track multiple targets if they are sufficiently separated in space and/or time –separate track for each target Can track spatio-temporally overlapping targets with appropriate classification algorithms

9. Classification Three types of classifiers under investigation: –K-nearest neighbor (KNN) –Maximum likelihood (ML) (Gaussian mixture) –Support Vector Machine (SVM) Three target classes: –Wheeled –Tracked –Unknown

10. Some Issues and Challenges Variability in measurements and conditions –Have to rely on data from prior experiments to train the classifiers for new experiments Variations in spectral signatures due to motion (Doppler), acceleration, gear shifts Timing synchronization for collaborative processing

11. Summary Framework for detection, classification and tracking of targets Single-node algorithms: –Energy detection –Classification (of single target events) Collaborative Processing: –Localization –Location prediction for tracking