Tracking Targets Via Particle Based Belief Propagation

We first formulate multiple targets tracking problem in a dynamic Markov network(DMN)which is derived from a MRFs for joint target state and a binary process for occlusion of dual adjacent targets. We then propose to embed a novel Particle based Belief Propagation algorithm into Markov Chain Monte Carlo approach (MCMC) to obtain the maximum a posteriori (MAP) estimation in the DMN. In the message propagation,a stratified sampler incorporates information both from a learned bottom-up detector (e.g. SVM classifier) and a top-down dynamic behavior model. Experimental results show that the proposed method is able to track varying number of targets and handle their interactions.