Abstract
Coalescence, meaning the tracker associates more than one trajectories to some targets while loses track for others, is a challenging problem for visual tracking of multiple targets, especially when similar targets move close or present occlusions. Existing approaches that are based on joint data association are confronted by the combinatorial complexity due to the concatenation of the state spaces of individual targets. This paper presents a novel collaborative approach with linear complexity to the coalescence problem. The basic idea is the collaborative inference mechanism, in which the estimate of an individual target is not only determined by its own observation and dynamics, but also through the interaction and collaboration with the estimates of its adjacent targets, which leads to a competition mechanism that enables different targets to compete for the common image observations. The theoretical foundation of the new approach is based on Markov networks. Variational analysis of this Markov network reveals a mean field approximation to the posterior density of each target, therefore provides a computationally efficient way for such a difficult inference problem. In addition, a mean field Monte Carlo (MFMC) algorithm is designed to achieve Bayesian inference by simulating the competition among a set of low dimensional particle filters. Compared with the existing solutions, the proposed new collaborative approach stands out by its effectiveness and low computational cost to the coalescence problem, as pronounced in the extensive experiments.