On line Bayesian tracking and detection of multiple objects

Sequential Monte Carlo (SMC) methods such as particle filters have been used in tracking problems for moving from an intractable distribution to a density that is closer to the actual posterior distribution. These methods makes use of stochastic simulations that can approximate non-linear and non-Gaussian posterior distributions via importance sampling. Since standard SMC methods only allows to track a single target, it becomes necessary find representations of the multiple object posterior density. The filtering distribution must take into account measurement-to-track data association and the probability of objects appearing and disappearing in the field of view. The problem can be even more challenging when considering misdetections and clutter, so methods based on random sets and the point process theory have been proposed as suitable representations for the intractable distribution. A method that uses a random set formalism called the probability hypothesis density (PHD) filter is considered for the joint detection and tracking problem. The PHD filter has been previously used in tracking in video sequences where the detection step is given by a separate process, so a particle PHD filter is approximating the multiple object posterior density. In this paper we show the SMC implementation of the PHD filter in video sequences and present work in progress results for the joint tracking and detection problem. The method proposed is able to model appearing and disappearing objects in the presence of occlusion and overlapping measurements under heavy clutter.