Split and merge data association filter for dense multi-target tracking

Abstract

Bayesian target tracking methods consist in filtering successive measurements coming from a detector. In the presence of clutter or multiple targets, the filter must be coupled with an association procedure. The classical Bayesian multitarget tracking methods rely on the hypothesis that a target can generate at most one measurement per scan and that a measurement originates from at most one target. When tracking a high number of deformable sources, the previous assumptions are often not met that leads to the failure of the existing methods. Here, we propose an algorithm which allows to perform the tracking in the cases when a single target generates several measurements or several targets generate a single measurement. The novel idea presented in this paper is the introduction of a set that we call virtual measurement set which supersedes and extends the set of measurements. This set is chosen to optimally fit the set of the predicted measurements at each time step. This is done in two stages: i) a set of feasible joint association events is built from virtual measurements that are created by successively splitting and merging the real measurements; ii) the joint probability is maximized over all feasible joint association events. The method has been tested on microscopy image sequences which typically contains densely moving objects and gives satisfactory preliminary results.