Scholarlink

<p>With the rapid advancement of artificial intelligence in UAV applications, this paper proposes an enhanced deep learning-based tracking framework that fundamentally improves occlusion handling through optimized detection network architecture and spatio-temporal information fusion. To achieve robust UAV tracking in complex scenarios, this work centers on a structural enhancement of the YOLO detector within the tracking-by-detection paradigm. On the detection stage, a redesigned backbone that incorporates a novel DSimSPPF module followed by a Receptive Field Enhancement Module for hierarchical feature extraction were integrated, along with refined loss functions for precise localization, collectively achieving 9.3% higher precision than baseline network. On the tracking stage, a discriminative tracker by integrating Felzenszwalb’s Histogram of Oriented Gradients with DSST’s scale adaptation within a Kernel Correlation Filter framework was developed, further augmented by a Color Name-based occlusion judgment mechanism to adaptively assess occlusion scenarios. The proposed spatio-temporal fusion strategy intelligently combines detection and tracking information, demonstrating better performance on OTB100 (70.0% precision, 60.9% success rate at 96.7 FPS over OCC occlusion attributes) and UAV123 (50.0% precision, 32.6% success rate at 80.1 FPS averaged over FOC/POV occlusion attributes) datasets. Edge deployment feasibility was confirmed on Jetson TX2 through TensorRT optimization and CUDA-accelerated computing, achieving real-time tracking performance with lower memory usage. This work provides both theoretical insights and practical solutions for reliable target tracking under occlusion conditions.</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

Deep learning-based real-time spatio-temporal framework for occlusion-robust UAV tracking

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Xiaotong Chen,
  • Jianyuan Wang,
  • Meng Chen,
  • Jinbao Chen

摘要

With the rapid advancement of artificial intelligence in UAV applications, this paper proposes an enhanced deep learning-based tracking framework that fundamentally improves occlusion handling through optimized detection network architecture and spatio-temporal information fusion. To achieve robust UAV tracking in complex scenarios, this work centers on a structural enhancement of the YOLO detector within the tracking-by-detection paradigm. On the detection stage, a redesigned backbone that incorporates a novel DSimSPPF module followed by a Receptive Field Enhancement Module for hierarchical feature extraction were integrated, along with refined loss functions for precise localization, collectively achieving 9.3% higher precision than baseline network. On the tracking stage, a discriminative tracker by integrating Felzenszwalb’s Histogram of Oriented Gradients with DSST’s scale adaptation within a Kernel Correlation Filter framework was developed, further augmented by a Color Name-based occlusion judgment mechanism to adaptively assess occlusion scenarios. The proposed spatio-temporal fusion strategy intelligently combines detection and tracking information, demonstrating better performance on OTB100 (70.0% precision, 60.9% success rate at 96.7 FPS over OCC occlusion attributes) and UAV123 (50.0% precision, 32.6% success rate at 80.1 FPS averaged over FOC/POV occlusion attributes) datasets. Edge deployment feasibility was confirmed on Jetson TX2 through TensorRT optimization and CUDA-accelerated computing, achieving real-time tracking performance with lower memory usage. This work provides both theoretical insights and practical solutions for reliable target tracking under occlusion conditions.