Cross-camera multi-person tracking by leveraging fast graph mining algorithm

Multiple person tracking is a very useful task in intelligent video surveillance, which is hindered by many challenges such as the variations of illumination, the irregular changes of human shapes and the particle occlusions. To tackle these challenges, in this paper, we propose a new online learning tracking system to generate the complete trajectory for each tracking object. In detection stage, we build a classifier for each tracking object by online learning in order to provide more accurate detection results. Online learning could real-time update the classifier for an accurately tracking results in the future. In the tracklet generation stage, we apply the spatiotemporal constrain to generate a set of reliable tracklets. Finally, we propose a new Part-based matching method to get the correlation between different tracklets and apply linear programming and greedy algorithm to handle the data association problem to generate the complete trajectory for each tracking object. In particular, our approach is able to cast the multiple cameras tracking problem as a data association problem. The experiments on our proposed method demonstrate state-of-the-art performance in multiple person tracking.