An Adaptive and Scalable Multi-Object Tracker Based on the Non-Homogeneous Poisson Process

Abstract

This paper proposes a new adaptive framework for tracking multiple objects in the presence of data association uncertainty and heavy clutter, either with or without knowledge of the measurement rates and/or target shapes. Built upon an online Gibbs sequential Markov chain Monte Carlo sampling scheme, the adaptive tracker is Bayesian optimal and robust, requiring no additional approximations or measurement partition steps. With a non-homogeneous Poisson process measurement model, our tracker can tackle the data association task with linear computational complexity. Meanwhile, we study generalised inverse Gaussian and inverse Wishart distributions for modelling Poisson rates and object shapes, respectively; these prior models ensure closed-form full conditionals in our online Gibbs sampling steps, under which object states and shapes can be jointly estimated with associations and Poisson rates in a parallel fashion. Furthermore, a fast Rao-Blackwellisation scheme for linear Gaussian dynamics is designed and demonstrated to significantly improve both tracking efficiency and accuracy. We validate the efficacy of our method on real and simulated data.