T2Track: Multi-Object Tracking by Associating between Tracks

multi-object tracking (MOT) is an essential task in computer vision, such as traffic flow analysis, autonomous driving, and robot vision. However, due to limitations of detector, fast motion of objects, and occlusion, it is difficult to track objects continuously. Therefore, we propose T2Track, a novel track management strategy to restore long-term lost trajectories. Based on the fact that the object cannot suddenly appear or disappear in the center of the image, new tracks appearing in a specified region are marked. When these tracks are active, we re-associate them with lost tracks. Furthermore, a time constraint is adopted to filter out associations that violate correct temporal logic. We have successfully deployed it in our roadside perception project. Compared to state-of-the-art trackers capable of real-time operation on edge computing devices, T2Track performs better, especially impressive ID retention capability. We hope it can be expanded to more fields in the future.