Cross-entropy method for K-best dependent-target data association hypothesis selection

Abstract

This paper is concerned with probabilistic evaluation of multiple-frame data association hypotheses in multiple-target tracking problems, in particular, when targets are not necessarily independent a priori. Multiple-target tracking problems with dependent targets naturally arise whenever targets interact with each other, as they move in congested traffic, or as they actively coordinate their movements in other situations. This paper develops a Bayesian data association hypothesis evaluation formula for dependent targets. Because the resulting formula does not have a multiplicative or log-linear form, the best hypothesis cannot be selected by integer linear programming or multi-dimensional assignment algorithms commonly used to solve data association problems in multiple target tracking. Instead, we propose to use Reuven Rubinstein's cross-entropy method as a possible solution. A K-best hypothesis selection extension will be discussed as an application of the generalized Murty's algorithm. This paper focuses on the theoretical aspects as the first step of a solution concept development.