Long-short time association algorithm: a robust data association algorithm

Abstract

The main challenge of multi-object tracking (MOT) is identity switch caused by severe occlusion. The solution to identity switching is video object association, which assigns an identity number to the same target in different frames. In this paper, a long-short time association algorithm is proposed for identity switching. In the short-time, that is, the motion features between adjacent frames are used to match, and in the long-time, that is, the non-adjacent frames are directly added to the appearance features for association to rematch the object detected after occlusion. Besides, the Kalman filter is improved and the frame width parameter is added to make the predicted frame more accurate; appearance features use average appearance features and increase detection confidence as update parameters to make appearance more robust and can still work in complex scenes. The new tracker, LSATrack, achieves 81.3MOTA and 81.3IDF1 in the MOT17 and achieves stable tracking in severe occlusion scenarios.

Keywords

• video object association|multi-object tracking|tracking-by-detection|average appearance feature