Long-term Tracker with Adaptive Occlusion and Recovery Judgment

Abstract

Compared with a short-term task, long-term tracking has received more attention and research in recent years. Long-term tracking is more challenging because it needs to solve two difficult problems: when to update and how to update our model. Many outstanding short-term tracking methods update frame by frame or manually set the threshold to judge if the tracker should be updated, but when the target is blocked or escapes from the field of view, it is easy to get and update wrong samples, resulting in model pollution and drift. Not only that, but due to the lack of a re-detection mechanism, it is difficult for these short-term tracking methods to recover once the target is lost (especially when the target reappears from another location). In this work, we propose a high-speed long-term tracker with adaptive occlusion and recovery judgment (LT-AOR), which comprehensively judge the update chance of the tracker through the discrimination information and appearance information, and re-detects the target in a simplified way to achieve stable tracking in the case of target occlusion and loss.