Scholarlink

<p>In RGBT tracking, the TIR modality complements the RGB modality by providing robust information in challenging scenarios such as low-light conditions and occlusions, thereby enhancing the robustness and stability of visual tracking. However, existing RGBT tracking methods often extract features separately, and then simply concatenate or fuse them without fully exploring the inherent relationships between the modalities. To address these issues, a novel RGBT object tracking framework is proposed in this paper. It employs a hierarchical multi-modal feature fusion strategy to boost tracking accuracy. Specifically, X-Modal Bridge Attention (XMBA) mechanism is introduced to integrate features of different modalities, realizing coarse-grained and effective interactions. Furthermore, a Multi-modal Deep Interaction Module (MDIM) is designed to perform multi-level feature fusion, dynamically guiding fine-grained interactions via templates. This hierarchical fusion strategy enables the model to gain a comprehensive understanding of target features in complex scenarios, ultimately enhancing its tracking performance. Extensive experimental results on three popular RGBT tracking benchmarks demonstrate the superiority of our proposed tracker. It achieves state-of-the-art performance, exhibiting robust results even under challenging conditions such as fast motion, abrupt illumination variation, motion blur, similar appearance and partial occlusion. The implementation and training information will be released in the future at: <a href="https://github.com/TrackingR/HMFF.git.">https://github.com/TrackingR/HMFF.git.</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

Hierarchical multi-modal feature fusion for RGBT tracking

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Na Li,
  • Kai Huang,
  • Zihang Wang,
  • Yuquan Gan,
  • Jinglu He

摘要

In RGBT tracking, the TIR modality complements the RGB modality by providing robust information in challenging scenarios such as low-light conditions and occlusions, thereby enhancing the robustness and stability of visual tracking. However, existing RGBT tracking methods often extract features separately, and then simply concatenate or fuse them without fully exploring the inherent relationships between the modalities. To address these issues, a novel RGBT object tracking framework is proposed in this paper. It employs a hierarchical multi-modal feature fusion strategy to boost tracking accuracy. Specifically, X-Modal Bridge Attention (XMBA) mechanism is introduced to integrate features of different modalities, realizing coarse-grained and effective interactions. Furthermore, a Multi-modal Deep Interaction Module (MDIM) is designed to perform multi-level feature fusion, dynamically guiding fine-grained interactions via templates. This hierarchical fusion strategy enables the model to gain a comprehensive understanding of target features in complex scenarios, ultimately enhancing its tracking performance. Extensive experimental results on three popular RGBT tracking benchmarks demonstrate the superiority of our proposed tracker. It achieves state-of-the-art performance, exhibiting robust results even under challenging conditions such as fast motion, abrupt illumination variation, motion blur, similar appearance and partial occlusion. The implementation and training information will be released in the future at: https://github.com/TrackingR/HMFF.git.