Multi-Target Tracking and Occlusion Handling with Learned Variational Bayesian Clusters and a Social Force Model

This paper considers the problem of tracking multiple humans in video. A solution is proposed which is able to deal with the challenges of a varying number of targets, occlusions and interactions when every target gives rise to multiple measurements. The developed novel framework comprises a variational Bayesian clustering approach combined with a social force model. It can handle a time-varying number of targets and can cope with complex inter-target occlusions by maintaining the identities of targets during their close physical interactions. It performs measurement to-target association by automatically detecting the measurement relevance. A variational Bayesian clustering technique clusters the measurements and thereby provides an elegant solution to the measurement origin uncertainty problem. A particle filter is developed in which the clustering algorithm and the social force model enhance the prediction step. The performance of the proposed framework is evaluated over several sequences from publicly available data sets: AV16.3, CAVIAR and PETS2006, which demonstrate that the proposed framework successfully initializes and tracks a variable number of human targets in the presence of complex occlusions. The proposed algorithm is compared with state-of-the-art techniques due to Khan et al., Laet et al. and Czyz et al. and is shown to achieve improved tracking performance.