Scholarlink

<p>During the acquisition of tunnel surface point clouds using mobile laser scanning, various types of noise are inevitably introduced, which adversely affect subsequent point cloud processing. Existing point cloud denoising methods typically uniformly apply noise discrimination to all points, which may result in the destruction of structural features or insufficient noise removal. Therefore, this paper proposes an Edge-Focus U-Net (EFU-Net)-based denoising method for underground tunnel point clouds. First, a bidirectional projection method is used to fit the tunnel centerline, and the point cloud is sliced and projected along this axis to construct grid matrices, establishing a mapping between the 3D space and the 2D representation. In this way, the three-dimensional point cloud denoising problem is transformed into a tunnel contour recognition task in the matrix domain. Subsequently, an EFU-Net integrating the Convolutional Block Attention Module (CBAM) and the Channel–Edge Attention Module (CEAM) is proposed to enhance boundary recognition and remove noise meshes from the matrix through an edge-attention learning strategy. Finally, experimental studies are conducted across multiple tunnel scenarios. The results demonstrate that the proposed method effectively eliminates different types of noise across diverse tunnel environments. After denoising, the feature retention rates (FRR) exceed 97%, the intersection over union (IoU) increases by approximately 30%, and the root mean square error (RMSE) ultimately remains below 0.1&#xa0;cm. These findings confirm that the proposed method effectively eliminates various types of noise while preserving structural features, making it well-suited for complex underground tunnel environments.</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

EFU-net-based point cloud denoising method for underground space tunnel

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Yuan Zhang,
  • Shaoyi Xu,
  • Chengtao Wang,
  • Chong Chen,
  • Zuhao Zhu

摘要

During the acquisition of tunnel surface point clouds using mobile laser scanning, various types of noise are inevitably introduced, which adversely affect subsequent point cloud processing. Existing point cloud denoising methods typically uniformly apply noise discrimination to all points, which may result in the destruction of structural features or insufficient noise removal. Therefore, this paper proposes an Edge-Focus U-Net (EFU-Net)-based denoising method for underground tunnel point clouds. First, a bidirectional projection method is used to fit the tunnel centerline, and the point cloud is sliced and projected along this axis to construct grid matrices, establishing a mapping between the 3D space and the 2D representation. In this way, the three-dimensional point cloud denoising problem is transformed into a tunnel contour recognition task in the matrix domain. Subsequently, an EFU-Net integrating the Convolutional Block Attention Module (CBAM) and the Channel–Edge Attention Module (CEAM) is proposed to enhance boundary recognition and remove noise meshes from the matrix through an edge-attention learning strategy. Finally, experimental studies are conducted across multiple tunnel scenarios. The results demonstrate that the proposed method effectively eliminates different types of noise across diverse tunnel environments. After denoising, the feature retention rates (FRR) exceed 97%, the intersection over union (IoU) increases by approximately 30%, and the root mean square error (RMSE) ultimately remains below 0.1 cm. These findings confirm that the proposed method effectively eliminates various types of noise while preserving structural features, making it well-suited for complex underground tunnel environments.