Occlusion Handling Based on Motion Estimation for Multi-Object Tracking

Although current computer vision-based object tracking has achieved excellent results in various rankings, the completeness and accuracy of the tracking trajectory are unsatisfactory when the objects are occluded. Moreover, object tracking in computer vision associates the objects before and after the occlusion by their appearance features yet lacks an accurate representation of the object's motion when the targets are occluded. We believe that Multi-Object Tracking (MOT) in autonomous driving needs to provide the vehicle with the motion state of the occluded object, including the position and velocity. Target's movement is a critical issue in autonomous driving, and we think it is more important to know how the target moves than what the target is. Therefore, this paper proposes a MOT algorithm for occlusion handling based on the tracking-by-detection strategy for the motion estimation of the occluded targets in autonomous vehicles. We first classify the occluded targets into two groups and track them with different strategies. The effectiveness of the proposed algorithm is validated on the MOTChallenge dataset by quantitative, qualitative, and ablation studies. Experimental results demonstrate that the algorithm can estimate the target motion accurately during occlusion and generate complete and higher quality motion trajectories.