Pose-Based Multi-Target Tracking

Abstract

Human multi-target tracking in video is an important issue in the field of computer vision. It is necessary to detect the target on each frame, and to connect the targets of all frames into a target sequence. For target matching among different frames, we propose a tracking algorithm for constructing object pose sequence(COPS) based on Openpose. The position status and the ORB feature of the target pose are dynamically weighted and fused into new features. Target pose is searched in the corresponding target pose sequence by comparing the new features between the target pose in the sequence and every pose in current frame. When the target pose is matched, the influence of the position feature on the pose similarity could be enhanced when the target motion is continuously detected. When the target scale changes too much, the method can expand the contribution of the ORB feature to the pose similarity comparison. The experiments of human multitarget tracking algorithm are carried out on the PoseTrack and MOT datasets, and the results show that the proposed tracking algorithm overcomes the problem of target matching between frames.