Scholarlink

<p>Multi-object tracking (MOT) is a core task in video signal processing and visual surveillance. However, it still faces challenges such as occlusion, dense interaction, and similar appearance in complex scenes. This paper presents AFGM-SORT, an online tracking-by-detection framework with adaptive feature fusion and global matching for robust MOT. First, a dynamic weight fusion strategy is designed to adjust the contributions of motion and appearance features according to trajectory missing frames. Second, a confidence-aware global simultaneous matching scheme is proposed to unify high- and low-confidence detections and avoid suboptimal association. Third, a composite geometric cost function is constructed by integrating center distance, weighted Height IoU, IoU, and ReID cost to improve spatial discrimination. Extensive experiments on DanceTrack and MOT20 show that AFGM-SORT achieves competitive performance with 67.7 HOTA and 70.2 IDF1 on DanceTrack, and 64.4 HOTA and 80.2 IDF1 on MOT20. The proposed method effectively reduces identity switches and improves tracking robustness in crowded and occluded scenes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AFGM-SORT: Adaptive feature fusion and global matching for robust multi-object tracking

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Furong Peng,
  • Ning Wang,
  • Linjun Wu,
  • Zhen Peng

摘要

Multi-object tracking (MOT) is a core task in video signal processing and visual surveillance. However, it still faces challenges such as occlusion, dense interaction, and similar appearance in complex scenes. This paper presents AFGM-SORT, an online tracking-by-detection framework with adaptive feature fusion and global matching for robust MOT. First, a dynamic weight fusion strategy is designed to adjust the contributions of motion and appearance features according to trajectory missing frames. Second, a confidence-aware global simultaneous matching scheme is proposed to unify high- and low-confidence detections and avoid suboptimal association. Third, a composite geometric cost function is constructed by integrating center distance, weighted Height IoU, IoU, and ReID cost to improve spatial discrimination. Extensive experiments on DanceTrack and MOT20 show that AFGM-SORT achieves competitive performance with 67.7 HOTA and 70.2 IDF1 on DanceTrack, and 64.4 HOTA and 80.2 IDF1 on MOT20. The proposed method effectively reduces identity switches and improves tracking robustness in crowded and occluded scenes.