Tracking multiple targets via Particle Based Belief Propagation

Tracking multiple visual targets involving occlusion and varying number problems is a challenging problem in computer vision. This paper presents a summary of recent works on multiple targets tracking (MTT) in video done by a research group of Institute of Artificial Intelligence and Robotics, XJTU. The summary is given in two parts. The first part presents their work on modelling interactions among objects. MTT problem in video is formulated in a dynamic Markov network. For tracking, a novel sequential stratified sampling belief propagation algorithm is introduced for MAP estimation in dynamic Markov network. It was also shown how to include bottom-up information from a learned detector and belief information for the message updating. The second part presents a sequential Monte Carlo data association algorithm for the MTT problem, which is based on a two-level computational framework. The framework integrates discriminative model learning, Monte Carlo joint data association filtering, and belief propagation algorithm. These methods are realized as different levels of approximation to an ’ideal’ generative model of multiple visual targets tracking, and result in a novel sequential Monte Carlo data association algorithm