Scholarlink

<p>RGB-thermal (RGB-T) fusion image object tracking leverages the complementary strengths of both modalities to improve performance in complex environments. However, most existing RGB-T tracking architectures fail to fully exploit the synergistic characteristics of RGB and thermal infrared (TIR) data, leading to suboptimal fusion and reduced effectiveness in challenging unmanned aerial vehicle (UAV) scenarios such as small objects and background clutter. This paper presents a fusion tracking framework based on a dual-stream Siamese network for RGB-T dual-modality object tracking on UAVs. A bimodal information fusion network is devised, where the classification branch differentiates between foreground and background, while the regression branch refines target bounding box precision. Furthermore, an attention mechanism tailored for RGB-T fusion tracking is incorporated to enhance salient feature representations across diverse scenarios, enabling the model to concentrate on the target or regions of interest. The proposed method achieved a precision rate (PR) of 88.8% and a success rate (SR) of 72.2% on the GTOT dataset, with a processing speed of 75 FPS. On the RGB-T234 dataset, it attained a PR of 78.1% and an SR of 53.4%. Furthermore, the method demonstrated superior tracking performance under challenging scenarios such as small objects and background variations.</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

Dual-stream siamese network for RGB-T dual-modal fusion object tracking on UAV

  • 期刊论文

  • 发表日期：

    2025年6月6日

查看全文
  • Hui Zhao,
  • Lei Zhang

摘要

RGB-thermal (RGB-T) fusion image object tracking leverages the complementary strengths of both modalities to improve performance in complex environments. However, most existing RGB-T tracking architectures fail to fully exploit the synergistic characteristics of RGB and thermal infrared (TIR) data, leading to suboptimal fusion and reduced effectiveness in challenging unmanned aerial vehicle (UAV) scenarios such as small objects and background clutter. This paper presents a fusion tracking framework based on a dual-stream Siamese network for RGB-T dual-modality object tracking on UAVs. A bimodal information fusion network is devised, where the classification branch differentiates between foreground and background, while the regression branch refines target bounding box precision. Furthermore, an attention mechanism tailored for RGB-T fusion tracking is incorporated to enhance salient feature representations across diverse scenarios, enabling the model to concentrate on the target or regions of interest. The proposed method achieved a precision rate (PR) of 88.8% and a success rate (SR) of 72.2% on the GTOT dataset, with a processing speed of 75 FPS. On the RGB-T234 dataset, it attained a PR of 78.1% and an SR of 53.4%. Furthermore, the method demonstrated superior tracking performance under challenging scenarios such as small objects and background variations.