Multiple Object Tracking algorithm based on cluster cooperative motion constraint

Multi-target tracking is one of the key research fields of computer vision. Existing methods mainly focus on inferring robust and discriminative features for data association based on targets generated by existing detectors. However, these methods independently process each target when generating the trajectory. In contrast, we propose a new multi-target tracking algorithm, which utilizes spatiotemporal cooperative motion constraints. Our algorithm makes use of the cooperative effect of crowd movement, which means that objects in close range tend to move in a more consistent way. We group several objects with consistent mobility into a cluster. We introduce an affinity matrix to measure the synergy and similarity of the movement between targets in the same cluster. When the target is occluded, the affinity matrix can be used to recover the lost track. We evaluated our approach on the MOT challenge benchmark and proved its effectiveness.