Multiple target tracking with structural inference

In this paper we propose a novel multiple target tracking model composed of two detectors and a tracker. An on-line detector and a tracker are used to generate target candidates, whose confidence scores are then evaluated by the off-line trained detectors. In the data association stage, the high-efficient inference in a structural model leads to the optimal tracking result. The experimental results demonstrate that our tracking model can overcome the occlusion and appearance changing problems. The tracking results can be applied to analyze the structural information of a single target or among multiple targets.