Scholarlink

<p>Video person re-identification (ReID) aims to retrieve video sequences of the same individual across non-overlapping camera views. Despite rich temporal cues, cross-camera trajectories often suffer from temporal misalignment and inconsistent action phases due to varying viewpoints and motion states. We propose a robust two-stage video retrieval system tailored for compute-budget-constrained surveillance deployments. The first stage employs a self-attention-based Transformer to learn global video representations and generate a high-recall Top-K candidate set. The second stage introduces a Hierarchical Cross-Attention Reordering (HCAR) module to capture fine-grained spatial correspondences and perform content-based keyframe matching, mitigating occlusion and temporal misalignment. Experiments on MARS, iLIDS-VID, and PRID2011, together with controlled analyses on MARS, show that the proposed sparse refinement improves retrieval performance over the same baseline retriever while avoiding dense full-gallery re-ranking. The source code is publicly available at: <a href="https://github.com/Lwj-68668/HCAR-ReID">https://github.com/Lwj-68668/HCAR-ReID</a>. <a href="https://doi.org/10.6084/m9.figshare.31410519">https://doi.org/10.6084/m9.figshare.31410519</a>.</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

Robust video person re-identification via hierarchical cross-attention mechanisms

  • 期刊论文

  • 发表日期：

    2026年8月5日

查看全文
  • Wenjie Li,
  • Shuai Guo,
  • Ying Li

摘要

Video person re-identification (ReID) aims to retrieve video sequences of the same individual across non-overlapping camera views. Despite rich temporal cues, cross-camera trajectories often suffer from temporal misalignment and inconsistent action phases due to varying viewpoints and motion states. We propose a robust two-stage video retrieval system tailored for compute-budget-constrained surveillance deployments. The first stage employs a self-attention-based Transformer to learn global video representations and generate a high-recall Top-K candidate set. The second stage introduces a Hierarchical Cross-Attention Reordering (HCAR) module to capture fine-grained spatial correspondences and perform content-based keyframe matching, mitigating occlusion and temporal misalignment. Experiments on MARS, iLIDS-VID, and PRID2011, together with controlled analyses on MARS, show that the proposed sparse refinement improves retrieval performance over the same baseline retriever while avoiding dense full-gallery re-ranking. The source code is publicly available at: https://github.com/Lwj-68668/HCAR-ReID. https://doi.org/10.6084/m9.figshare.31410519.