Scholarlink

<p>Multi-camera object tracking in complex environments faces significant challenges, including occlusions, appearance variations, and cross-camera association. This paper presents a robust tracking framework that integrates discriminative spatiotemporal and appearance features to address these issues. Our method introduces a novel matching area definition and a similarity function matrix to resolve cross-camera trajectory association, especially for fragmented tracklets. We also employ feature autocorrelation analysis to extract illumination-robust appearance descriptors, thereby enhancing discrimination under variations. Extensive experiments on three airport terminal surveillance datasets and the standard EPFL benchmark demonstrate state-of-the-art performance: a single-camera tracking accuracy of 88.5% (an average improvement of 6% over baseline methods) and a 15% reduction in identity switches. For global trajectory reconstruction, our method achieves an accuracy of 89.4%, a 5% gain over existing approaches. These results highlight the method’s superior robustness to occlusions and appearance changes, offering a practical solution for large-scale surveillance applications. The source code for reproducing our experiments is permanently available at <a href="https://github.com/DangWanli0417/mtmct-terminal-tracking.git">https://github.com/DangWanli0417/mtmct-terminal-tracking.git</a> and archived with the DOI <a href="https://doi.org/10.5281/zenodo.17520331">10.5281/zenodo.17520331</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 multi-camera tracking in terminal environments: a spatio-temporal-appearance fusion approach

  • 期刊论文

  • 发表日期：

    2026年4月4日

查看全文
  • Dang Wanli,
  • Cheng Jian,
  • Luo Qian,
  • Zheng Huaiyu

摘要

Multi-camera object tracking in complex environments faces significant challenges, including occlusions, appearance variations, and cross-camera association. This paper presents a robust tracking framework that integrates discriminative spatiotemporal and appearance features to address these issues. Our method introduces a novel matching area definition and a similarity function matrix to resolve cross-camera trajectory association, especially for fragmented tracklets. We also employ feature autocorrelation analysis to extract illumination-robust appearance descriptors, thereby enhancing discrimination under variations. Extensive experiments on three airport terminal surveillance datasets and the standard EPFL benchmark demonstrate state-of-the-art performance: a single-camera tracking accuracy of 88.5% (an average improvement of 6% over baseline methods) and a 15% reduction in identity switches. For global trajectory reconstruction, our method achieves an accuracy of 89.4%, a 5% gain over existing approaches. These results highlight the method’s superior robustness to occlusions and appearance changes, offering a practical solution for large-scale surveillance applications. The source code for reproducing our experiments is permanently available at https://github.com/DangWanli0417/mtmct-terminal-tracking.git and archived with the DOI 10.5281/zenodo.17520331.