Hybrid Probabilistic Data Association and Variational Filtering for Multi-Target Tracking in Wireless Sensor Networks Jing Teng, Hichem Snoussi, C´ edric Richard and Yi Zhou ICD/LM2S, University of Technology of Troyes 12 rue Marie Curie, 10000, France Email: jing.teng@utt.fr, hichem.snoussi@utt.fr, cedric.richard@utt.fr, yi.zhou@utt.fr Abstract—A hybrid signal processing scheme is proposed for distributed multi-target tracking (MTT). For the sake of resource efficiency in a wireless sensor network (WSN), we reduce the problem to parallel cluster-based single target tracking when the targets are far apart, and switch to MTT only when data association becomes ambiguous. A sequential monte carlo method is employed to assign the ambiguous observations to specific targets or clutter, based on association probabilities. Whereas the rest observations are incorporated by the variational filter, which approximates the distribution of involved particles by a simple Gaussian distribution for each target. The natural and adaptive message compression dramatically reduces the resource consumption of the WSN. The low computation complexity also guarantees the one-line execution of the hybrid MTT scheme. In addition, experimental results prove that the proposed scheme succeeds in distinguishing and tracking multiple targets even during the occlusions. I. I NTRODUCTION Among the potential applications of wireless sensor net- works (WSNs), the tracking of mobile targets has found its major importance in monitoring wildlife animals, vehicles on the freeway, and surveillance in the battle field etc. [1]. Target tracking consists of recursively updating the posterior distribution of the target state given the sequence of sensor observations and the state evolution model [2]. Multi-target tracking (MTT) deals with state estimation of several moving targets, which is not a trivial extension of single target tracking but rather a challenging topic of research. The main difficulty of MTT comes from the assignment of a given measurement to a specific target. Traditionally, the nearest neighbor (NN) approach, which utilizes the closest measurement to the predicted target mea- surement, is the simplest approach for MTT [3]. However, the NN measurements may be originated from a clutter, leading to filter divergence in many situations. As long as the data association is considered in a deterministic way, all possible associations must be exhaustively enumerated [4]. Multiple hypothesis tracking (MHT) [5] recursively builds all possible associations of measurements to existing/new tracks and false alarms, while respecting the mutual exclusion association constraint. MHT is capable of addressing the problems as low detection probability, high false alarm rates, delayed measurements, initiation and termination of tracks. However, it suffers from large storage space requirements, as the number of possible associations increases exponentially with time. The joint probabilistic data association filter (JPDAF) [6] is an alternative solution which consists of updating each individual track state with weighted combinations of all mea- surements. In fact, JPDAF is a particular way of combining the multiple hypotheses generated by MHT into a single hypothesis. Sequential Monte Carlo (SMC) method samples from complex association probability distribution conditioned on observations, where the sample with the highest probability is considered as the best association hypothesis [7], [8]. As the hypotheses are not explicitly enumerated, the large storage space is no longer required compared to MHT. Besides, the SMC method is very easy to implement and can be applied under very general hypotheses to cope with heavy clutters. Due to the consideration of all possible events in the data association phase, MTT is an expensive task in terms of sensing, computation and communication. Concerning the extremely stringent resources in WSNs, an energy-aware dis- tributed signal processing scheme is proposed in this paper. As the targets can travel arbitrarily and no a priori information on targets motion is provided, the general state evolution model proposed in [9], [10] is extended to describe the hidden states. Only the sensors which have detected the appearances of targets are activated to form data processing clusters for energy efficiency, where the cluster heads (CHs) are the ones with the most residual energy in each activated cluster. When the activated clusters are not overlapped, variational filters for single target tracking are parallelly executed in corresponding CHs. Otherwise, the activated CHs exchange ambiguous ob- servations with each other, and invoke the probabilistic data association phase. A SMC method is employed to assign the ambiguous observations to specific targets or clutter based on the association probabilities. Whereas the variational tracking is delayed after the SMC phase to incorporate the rest of observations. Owing to the implicit compression of variational filtering, the temporal dependence of each target is reduce to a Gaussian distribution, which dramatically cuts off the inter-cluster communication. An overview of the hybrid MTT scheme is illustrated by Fig. 1. The rest of the paper is organized as follows. In Section II, we formulate the variational tracking algorithm. Section III is dedicated to a detailed description of the SMC data association 2009 3rd IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing 978-1-4244-5180-7/09/$26.00 ©2009 IEEE 368