Scholarlink

<p>In the realm of computer vision, the duty of multiple objects tracking remains challenging, especially in scenarios involving occlusions and varying object appearances. In this work, we propose an innovative approach leveraging embedded graph matching to address these challenges. The proposed method constructs separate detection and tracklet graphs, to capture contextual relationships and matching constraints. An embedded graph matching network is employed to encode higher-order structural information into vertex features, significantly improving robustness against the cases of occlusions. Incorporating a differentiable Sinkhorn layer enables efficient optimal assignment, enhancing computational efficiency. Our experiments on MOT16, MOT17, and MOT20 datasets demonstrate competitive performance of the proposed method, contributing to smart city surveillance, autonomous driving, and other real-time tracking applications. Here, we achieved a 57.1% MOTA score on MOT17, highlighting the effectiveness of our proposed method.</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

Enhanced multi-object tracking via embedded graph matching and differentiable Sinkhorn assignment: addressing challenges in occlusion and varying object appearances

  • 期刊论文

  • 发表日期：

    2025年1月2日

查看全文
  • Yajuan Zhang,
  • Yongquan Liang,
  • Junjie Wang,
  • Houying Zhu,
  • Zhihui Wang

摘要

In the realm of computer vision, the duty of multiple objects tracking remains challenging, especially in scenarios involving occlusions and varying object appearances. In this work, we propose an innovative approach leveraging embedded graph matching to address these challenges. The proposed method constructs separate detection and tracklet graphs, to capture contextual relationships and matching constraints. An embedded graph matching network is employed to encode higher-order structural information into vertex features, significantly improving robustness against the cases of occlusions. Incorporating a differentiable Sinkhorn layer enables efficient optimal assignment, enhancing computational efficiency. Our experiments on MOT16, MOT17, and MOT20 datasets demonstrate competitive performance of the proposed method, contributing to smart city surveillance, autonomous driving, and other real-time tracking applications. Here, we achieved a 57.1% MOTA score on MOT17, highlighting the effectiveness of our proposed method.