Scholarlink

<p>RGB-T tracking leverages the complementary properties of visible (RGB) and thermal modalities to enhance tracking robustness in complex environments. However, modality and temporal unreliability remain major obstacles. To address these, we propose ECT-tracker (evidential collaboration and temporal gating), a unified framework incorporating evidential fusion and temporal gating. Specifically, our evidential dynamic fusion module uses Dempster–Shafer theory to reconcile conflicting modality information. Concurrently, the information-temporal collaboration head (ITC-head) employs a current-frame-driven gating strategy to mitigate temporal information pollution. Furthermore, a frequency–spatial feature enhancement module is integrated to refine feature representations and improve localization accuracy. Experiments on three challenging benchmarks—RGBT234, LasHeR, and VTUAV—demonstrate state-of-the-art accuracy and robustness, while achieving real-time performance at 55 FPS. Code and data are available at: <a href="https://github.com/lakewo0d/ECT-tracker.">https://github.com/lakewo0d/ECT-tracker.</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 RGBT tracking via evidential fusion and dynamic temporal gating

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Yuhang Deng,
  • Chengfang Zhang,
  • Ziliang Feng

摘要

RGB-T tracking leverages the complementary properties of visible (RGB) and thermal modalities to enhance tracking robustness in complex environments. However, modality and temporal unreliability remain major obstacles. To address these, we propose ECT-tracker (evidential collaboration and temporal gating), a unified framework incorporating evidential fusion and temporal gating. Specifically, our evidential dynamic fusion module uses Dempster–Shafer theory to reconcile conflicting modality information. Concurrently, the information-temporal collaboration head (ITC-head) employs a current-frame-driven gating strategy to mitigate temporal information pollution. Furthermore, a frequency–spatial feature enhancement module is integrated to refine feature representations and improve localization accuracy. Experiments on three challenging benchmarks—RGBT234, LasHeR, and VTUAV—demonstrate state-of-the-art accuracy and robustness, while achieving real-time performance at 55 FPS. Code and data are available at: https://github.com/lakewo0d/ECT-tracker.