PSTG-based multi-label optimization for multi-target tracking

摘要

Many recent advances in multi-target tracking have grown concern over latent corresponding relation among observations, e.g. social relationship. To handle long-term occlusion within group and tracking failure caused by interaction of targets, various correlations among tracklets need to be exploited. In this paper, a paratactic–serial tracklet graph (PSTG) theory is proposed for inter-tracklet analysis in multi-target tracking to avoid tracking failure caused by long-term occlusion within group or crossing trajectories. Contrary to recent approaches, a novel PSTG is defined to describe the correlation among all tracklets in spatio-temporal domain to model the mutual influence among trajectories. Paratactic–tracklet graph extends the potential relationship among tracklets which show similar motion patterns in spatio-temporal neighbor. Serial–tracklet graph enhances the integrity and continuity of trajectories which represent two trajectory fragments of a certain target in different periods. Furthermore, a PSTG-based multi-label optimization algorithm is presented to make the trajectory estimation more accurate. A PSTG energy is minimized by multi-label optimization, including group, integrity and spatio-temporal constraints. Experiments demonstrate the anti-occlusion performance of the proposed approach on several public datasets and actual surveillance sequences, and achieve competitive results by quantitative evaluation.