Multiple people tracking with articulation detection and stitching strategy

Multiple people tracking in a monocular video of crowded scenes is a challenging problem, methods of which are mostly based on tracking-by-detection strategies. The result of detection preprocessing used by many tracking methods to avoid creating wrong targets, is likely to be contaminated when there are defective detections in datasets of benchmark. We propose an articulation-based detection selecting method to screen out detections unqualified for further processing. For the association part of tracking workflow, applying minimax operation can minimize the max intra-distance but results in discontinuous trajectories. We design a stitching strategy to link the tracklets created by minimax algorithm. The experimental results will demonstrate that the proposed method outperforms or is comparable to previous approaches.