| Abstrakt: |
To address the problem of occlusion in multi-object tracking (MOT) scenarios, an anti-occlusion multi-object tracking (AMTrack) algorithm is proposed, which is applicable not only to low-viewpoint but also to high-viewpoint MOT. Firstly, an occlusion index (ONI) calculation method is given to quantify object occlusion, helping in measuring the extent of occlusion for each object. Secondly, an anti-occlusion multi-level association (AMLA) mechanism is designed for mitigating the impact of occlusion on data association by hierarchically associating objects containing different levels of noises, ensuring robust tracking in occluded environments. Additionally, an anti-noise trajectory feature update (ATFU) method is presented to reduce the noise from occlusion by selectively updating the trajectory feature. Finally, the experimental results on the MOT16, MOT17, MOT20 and Dancetrack test sets show that AMTrack exhibits state-of-the-art performance, compared with the classic methods such as OsaMOT, DeepSORT, RetinaMOT, TicrossNet, demonstrating its effectiveness and robustness in multi-object tracking scenarios with occlusion. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink