Scholarlink

<p>The railway catenary is central to railway electrification. Because of its exposed location and long spans, the catenary system is susceptible to the intrusion of foreign objects such as plastic bags, bird nests, and kites, which reduce the current collection performance of electric locomotives and increase the safety risks associated with railway operations. Due to the small target size and complex background environment, detecting foreign objects on the catenary poses significant challenges. For catenary foreign object detection which is challenged by small targets and complex backgrounds, this paper presents LKAM-RCatenaryDet, which incorporates a reparameterization-optimized large convolution kernel and a global attention mechanism. To enhance the capability of small-target detection, the LRFSRBackbone network is designed. With the goal of strengthening complex background detection, the Global Attention Mechanism is integrated into the small-target detection head. Furthermore, the Wise-IoU model is adopted as the loss function, which addresses anchor box overlap and refines detection accuracy. To verify the effectiveness of the proposed LKAM-RCatenaryDet, experiments were conducted on the self-constructed railway catenary foreign object dataset, with results showing that LKAM-RCatenaryDet surpasses YOLO11s which is currently recognized as the detection model in this field in Precision (+ 3.2%), Recall (+ 4.2%), mAP@50 (+ 4.7%), and mAP@95 (+ 2.6%). Dataset and code location: <a href="https://github.com/DJTUResearches/LKAM">https://github.com/DJTUResearches/LKAM</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

LKAM-RCatenaryDet: enhanced railway catenary foreign object detection via large kernel fusion and global attention mechanism

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Yanjuan Wang,
  • Xianxin Ji,
  • Jiatong Li,
  • Jun Zhao,
  • Yuxin Hu,
  • Fengqiang Xu,
  • Xiaohong Yan,
  • Fengqi Li

摘要

The railway catenary is central to railway electrification. Because of its exposed location and long spans, the catenary system is susceptible to the intrusion of foreign objects such as plastic bags, bird nests, and kites, which reduce the current collection performance of electric locomotives and increase the safety risks associated with railway operations. Due to the small target size and complex background environment, detecting foreign objects on the catenary poses significant challenges. For catenary foreign object detection which is challenged by small targets and complex backgrounds, this paper presents LKAM-RCatenaryDet, which incorporates a reparameterization-optimized large convolution kernel and a global attention mechanism. To enhance the capability of small-target detection, the LRFSRBackbone network is designed. With the goal of strengthening complex background detection, the Global Attention Mechanism is integrated into the small-target detection head. Furthermore, the Wise-IoU model is adopted as the loss function, which addresses anchor box overlap and refines detection accuracy. To verify the effectiveness of the proposed LKAM-RCatenaryDet, experiments were conducted on the self-constructed railway catenary foreign object dataset, with results showing that LKAM-RCatenaryDet surpasses YOLO11s which is currently recognized as the detection model in this field in Precision (+ 3.2%), Recall (+ 4.2%), mAP@50 (+ 4.7%), and mAP@95 (+ 2.6%). Dataset and code location: https://github.com/DJTUResearches/LKAM.