Motion Estimation Track: Multiple Object Tracking by Using Motion Estimation After Detection Failure

Motion-based multiple object tracking (MOT) methods are often solved with tracking-by-detection (TBD) paradigm, which matches the current track states with new detections based on motion or appearance information. However, this paradigm is limited by detector performance and leading to miss-tracking due to detection failure. To address this problem, we propose a simple and effective method for Motion Estimation Recovery (MER) that recover the trajectories of objects with motion estimation boxes when detections are missing. Additionally, we proposed a novel way to fuse motion and appearance cues with a Difference Weight (DW) which enhances the scores of best matches and differentiates similar objects. Combining them together forms our Motion Estimation Track (MOESTrack) framework. Experiments on two challenging MOT benchmarks (MOT17 and MOT20) demonstrate that our MOESTrack achieve advanced performance.