Scholarlink

<p>Existing RGBT trackers usually improve tracking performance through designed sophisticated multimodal fusion strategies. However, in complex scenarios with dynamically varying modality quality, how to mine robust target representations and achieve reliable multimodal fusion remains a significant challenge for RGBT tracking. To handle these problems, we propose a novel Uncertainty-Aware RGBT Tracking framework called UATrack, which models consecutive template frames as Gaussian distributions to capture latent template states for robust representation and exploits the reliability of multimodal template representations to estimate modality-specific uncertainty for reliable multimodal fusion. In particular, we devise an uncertainty-based representation enhancement module that models template features at different time steps as Gaussian distributions and exploits uncertainty to characterize the latent state space of the template to enhance the robustness and generalization of template representations. Furthermore, we design an uncertainty-aware multimodal fusion module, which models modality-specific uncertainty in the subjective logic framework by leveraging the correlation between multimodal template and search region representations to enable reliable multimodal fusion. In addition, existing trackers mainly focus on designing effective template update strategies but overlook the potential of the initial template during the updating process. To fill this gap, we propose a simple yet effective contrastive template update module, which determines template updates by jointly evaluating the relative quality of the candidate and initial templates and the classification scores predicted from their respective same search regions. Extensive experiments conducted on four RGBT tracking benchmarks demonstrate that our method outperforms existing state-of-the-art approaches. The code and results of UATrack are publicly available at <a href="https://github.com/dongdong2061/UATrack">https://github.com/dongdong2061/UATrack</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

Uncertainty-Aware RGBT Tracking

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Zhaodong Ding,
  • Chenglong Li,
  • Futian Wang,
  • Jin Tang

摘要

Existing RGBT trackers usually improve tracking performance through designed sophisticated multimodal fusion strategies. However, in complex scenarios with dynamically varying modality quality, how to mine robust target representations and achieve reliable multimodal fusion remains a significant challenge for RGBT tracking. To handle these problems, we propose a novel Uncertainty-Aware RGBT Tracking framework called UATrack, which models consecutive template frames as Gaussian distributions to capture latent template states for robust representation and exploits the reliability of multimodal template representations to estimate modality-specific uncertainty for reliable multimodal fusion. In particular, we devise an uncertainty-based representation enhancement module that models template features at different time steps as Gaussian distributions and exploits uncertainty to characterize the latent state space of the template to enhance the robustness and generalization of template representations. Furthermore, we design an uncertainty-aware multimodal fusion module, which models modality-specific uncertainty in the subjective logic framework by leveraging the correlation between multimodal template and search region representations to enable reliable multimodal fusion. In addition, existing trackers mainly focus on designing effective template update strategies but overlook the potential of the initial template during the updating process. To fill this gap, we propose a simple yet effective contrastive template update module, which determines template updates by jointly evaluating the relative quality of the candidate and initial templates and the classification scores predicted from their respective same search regions. Extensive experiments conducted on four RGBT tracking benchmarks demonstrate that our method outperforms existing state-of-the-art approaches. The code and results of UATrack are publicly available at https://github.com/dongdong2061/UATrack.