Improving multi‐object tracking by full occlusion handle and adaptive feature fusion

Abstract

Abstract Occlusion has always been a challenging research topic in the field of multi‐target tracking. The invisibility of the target in full occlusion increases the difficulty of continuous tracking, which makes the recovery failure when the target is re‐visible, and ultimately leads to a decrease in tracking accuracy. To address full occlusion problem, an effective multi‐object tracking algorithm with full occlusion handle and adaptive fusion features is proposed. Firstly, a spatio‐temporal model is established for full occlusion, and a simple, efficient and training‐free method is proposed to find full occluded targets. Secondly, local high discrimination features with better stability and independence is proposed to realize effective correlation between targets before and after the full occlusion. Finally, an adaptive feature fusion mechanism is proposed, which can adjust feature structure dynamically according to the occlusion state. The experimental results show that most evaluation metrics of the proposed algorithm are superior to those of some typical algorithms proposed in recent years under full occlusion tracking scenes. The proposed algorithm can realize accurate occluded targets identification and improve tracking robustness under short‐term, long‐term and frequent full occlusion.

Keywords

• computer vision
• image processing
• object tracking