Scholarlink

<p>In recent years, Unmanned Aerial Vehicle (UAV) tracking technology has been widely applied in fields such as agriculture and aviation. However, due to limitations in UAV computational resources and battery life, deep learning-based tracking algorithms are difficult to deploy directly. Therefore, efficient and robust tracking methods based on Discriminative Correlation Filters (DCF) have become the mainstream for UAV tracking. The traditional DCF method updates the model with a fixed step size, making it prone to contamination, while the boundary effect further limits its tracking capability. To address the issue of model contamination, an appearance model update and masking mechanism was designed, introducing feature confidence and overall confidence thresholds to evaluate the quality of features and images. Features and images with poor quality are masked, and the learning rate is reduced to prevent model contamination. This paper proposes the Target-Background Feature Block and Aberrance Repressed Correlation Filter (TBFBARCF), which introduces target and background regularization terms into the spatial filter to mitigate boundary effects and incorporates an aberrance repression regularization term into the temporal aberrance repressed filter to reduce response map aberrance. The two filters mutually constrain each other and are jointly trained. Experiments conducted on three challenging datasets, OTB100, UAV123, and UAVDT, demonstrate that the TBFBARCF tracker outperforms 11 other DCF-based trackers. Its average precision and average success rate reach 0.756 and 0.62, respectively, and it operates at a speed of 36.6 frames per second (FPS) on a CPU platform.</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

Target-background feature blocks and aberrance repressed correlation filters for real-time UAV tracking

  • 期刊论文

  • 发表日期：

    2025年5月9日

查看全文
  • Yunwei Jia,
  • Bo Liu,
  • Zhenzhong Liu,
  • Tianyang Wang,
  • Kaiying Yv,
  • Shuo Wang

摘要

In recent years, Unmanned Aerial Vehicle (UAV) tracking technology has been widely applied in fields such as agriculture and aviation. However, due to limitations in UAV computational resources and battery life, deep learning-based tracking algorithms are difficult to deploy directly. Therefore, efficient and robust tracking methods based on Discriminative Correlation Filters (DCF) have become the mainstream for UAV tracking. The traditional DCF method updates the model with a fixed step size, making it prone to contamination, while the boundary effect further limits its tracking capability. To address the issue of model contamination, an appearance model update and masking mechanism was designed, introducing feature confidence and overall confidence thresholds to evaluate the quality of features and images. Features and images with poor quality are masked, and the learning rate is reduced to prevent model contamination. This paper proposes the Target-Background Feature Block and Aberrance Repressed Correlation Filter (TBFBARCF), which introduces target and background regularization terms into the spatial filter to mitigate boundary effects and incorporates an aberrance repression regularization term into the temporal aberrance repressed filter to reduce response map aberrance. The two filters mutually constrain each other and are jointly trained. Experiments conducted on three challenging datasets, OTB100, UAV123, and UAVDT, demonstrate that the TBFBARCF tracker outperforms 11 other DCF-based trackers. Its average precision and average success rate reach 0.756 and 0.62, respectively, and it operates at a speed of 36.6 frames per second (FPS) on a CPU platform.