Probabilistic object and label association algorithms for distributed multiobject tracking

Abstract

Most distributed multiobject tracking methods use a fusion of posterior multiobject distributions that involves the association of distribution components describing the same object at different sensors. Most state-of-the-art fusion methods perform a hard association, which may produce incorrect results. By contrast, in this article, we propose a probabilistic (soft) association framework and methodology wherein an association distribution is incorporated in the fusion of the posterior multiobject distributions. We develop this principle for distributed multisensor versions of the joint probabilistic data association (JPDA) filter, the joint integrated probabilistic data association (JIPDA) filter, and the labeled multi-Bernoulli (LMB) filter. A key component of our development is the approximation of the association distribution by the product of its marginals, which is shown to lead to a fused multiobject distribution of the same form as the local multiobject distributions. We also propose efficient implementations using a belief propagation algorithm for fast approximate marginalization and/or a Gaussian approximation. Simulation results demonstrate significant performance gains relative to state-of-the-art methods based on hard association.