Scholarlink

<p>Overhead catenary system (OCS) condition monitoring is essential for the safe operation of high-speed railways. Owing to its rich geometric information and robustness to illumination variation and poor visibility conditions, 3D LiDAR has become an effective sensing modality for OCS inspection. Current point cloud segmentation methods in OCS scenarios face challenges due to uneven point cloud distribution, geometric feature differences, and complex structures, limiting segmentation performance. To overcome these challenges, this study proposes PGADNet, a segmentation network developed to enhance accuracy. The network uses the Point Cloud Refinement Attention (PCRA) module to reduce confusion among slender components, employs the Geometric Feature Enhancement (CGF) module to enhance vital structural features, and achieves dynamic feature fusion combination across multiple scales through the Dynamic Feature Fusion (DFF) module. Experimental results show that PGADNet achieves an overall accuracy (OA) of 99.44%, a mean accuracy (mAcc) of 98.67%, and a mean intersection over union (mIoU) of 96.54% on OCS point clouds. The method facilitates precise segmentation of critical regions in OCS point clouds and exceeds the performance of existing methods, providing a component level semantic basis for intelligent OCS inspection and downstream engineering analysis.</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

A Novel 3D LiDAR Point Cloud Semantic Segmentation Method for High-Speed Railway Overhead Contact Systems

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Deli Yan,
  • Zikun Shi,
  • Zhaohui Chen,
  • Xiaodong Ji,
  • Qian Li

摘要

Overhead catenary system (OCS) condition monitoring is essential for the safe operation of high-speed railways. Owing to its rich geometric information and robustness to illumination variation and poor visibility conditions, 3D LiDAR has become an effective sensing modality for OCS inspection. Current point cloud segmentation methods in OCS scenarios face challenges due to uneven point cloud distribution, geometric feature differences, and complex structures, limiting segmentation performance. To overcome these challenges, this study proposes PGADNet, a segmentation network developed to enhance accuracy. The network uses the Point Cloud Refinement Attention (PCRA) module to reduce confusion among slender components, employs the Geometric Feature Enhancement (CGF) module to enhance vital structural features, and achieves dynamic feature fusion combination across multiple scales through the Dynamic Feature Fusion (DFF) module. Experimental results show that PGADNet achieves an overall accuracy (OA) of 99.44%, a mean accuracy (mAcc) of 98.67%, and a mean intersection over union (mIoU) of 96.54% on OCS point clouds. The method facilitates precise segmentation of critical regions in OCS point clouds and exceeds the performance of existing methods, providing a component level semantic basis for intelligent OCS inspection and downstream engineering analysis.