Scholarlink

<p>Occlusion and motion blur remain major challenges in multi-object tracking (MOT), often resulting in identity switches (IDSW) and trajectory fragmentation. These difficulties are further compounded by unstable camera views, complex pedestrian dynamics, and cluttered visual environments. Although existing tracking-by-detection (TBD) methods—such as SORT, DeepSORT, and ByteTrack—have achieved notable progress, they struggle in dense or occluded scenarios. In this paper, we propose a novel crowded matching strategy to improve tracking robustness under crowded conditions. Furthermore, we introduce a stricter track initialization mechanism to suppress false positives and reduce ID switches, while lowering computational cost by reducing reliance on appearance feature extraction. Built within a standard online tracking-by-detection pipeline, our method extends conventional association with a crowd-specific third-stage recovery, ambiguity-triggered Re-ID, and crowd-aware lifecycle control. This design remains strictly online and improves identity preservation in crowded pedestrian 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

Multi-object pedestrian tracking method based on crowded matching

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Yanyun Liu

摘要

Occlusion and motion blur remain major challenges in multi-object tracking (MOT), often resulting in identity switches (IDSW) and trajectory fragmentation. These difficulties are further compounded by unstable camera views, complex pedestrian dynamics, and cluttered visual environments. Although existing tracking-by-detection (TBD) methods—such as SORT, DeepSORT, and ByteTrack—have achieved notable progress, they struggle in dense or occluded scenarios. In this paper, we propose a novel crowded matching strategy to improve tracking robustness under crowded conditions. Furthermore, we introduce a stricter track initialization mechanism to suppress false positives and reduce ID switches, while lowering computational cost by reducing reliance on appearance feature extraction. Built within a standard online tracking-by-detection pipeline, our method extends conventional association with a crowd-specific third-stage recovery, ambiguity-triggered Re-ID, and crowd-aware lifecycle control. This design remains strictly online and improves identity preservation in crowded pedestrian scenes.