Motion Vector Based Data Association for On-Line Multi-object Tracking

On-line multi-object tracking needs to solve the data association problem on each new frame in time-critical video analysis applications. However, associating the new detection responses and existing trajectories under the tracking-by-detection framework is faced with challenges such as mis-detections and false alarms. In order to build a more reliable frame-by-frame association with the given detection results in applications where precision is primarily required, we design a strong associating constraint based on motion vectors computed from uniformly sampled keypoints in the scene while considering spatial information at the same time. With the optical flow analysis between the two successive frames, we propose a new cost function for building the association matrix and solve the multi-object tracking problem in an on-line form. Experimental results on challenging benchmark datasets show that our method achieves overall state-of-the-art performance, especially effective in reducing false alarms.