Nonparametric Bayesian Methods for Large Scale Multi-Target Tracking

We consider the problem of data association for multi-target tracking in the presence of an unknown number of targets. For this application, inference in models which place parametric priors on large numbers of targets becomes computationally intractable. As an alternative to parametric models, we explore the utility of nonparametric Bayesian methods, specifically Dirichlet processes, which allow us to put a flexible, data-driven prior on the number of targets present in our observations. Dirichlet processes provide a prior on partitions of the observations among targets whose dynamics are individually described by state space models. These partitions represent the tracks with which the observations are associated. We provide preliminary data association results for the implementation of Dirichlet processes in this scenario.