Scholarlink

<p>Accurate semantic segmentation of point cloud data is vital for safety monitoring and intelligent management in tunnel construction. However, challenges such as cluttered environments, occlusions, and the lack of annotated domain-specific datasets hinder effective application of deep learning techniques. To address these issues, this study presents TCS-Net, a novel point cloud segmentation network tailored for under-construction tunnels. A large-scale annotated dataset, named 3D Tunnel, was constructed using handheld laser scanning and contains over 60 million points across eight structural categories, filling a critical data gap in the field. TCS-Net introduces a multi-module fusion framework that combines spatial attention mechanisms, an inverted residual MLP (InvResMLP) for enriched feature representation, and a Kd-tree–based Gaussian upsampling with channel attention for enhanced feature propagation. An optimized training strategy incorporating AdamW, cosine decay, and label smoothing further improves learning robustness. Experimental results on the 3D Tunnel dataset demonstrate that TCS-Net achieves superior segmentation performance, with 94.38% mean IoU and 98.23% overall accuracy, validating its effectiveness and practical potential in tunnel construction scenarios.</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 3D Point Cloud Segmentation for Nondestructive Evaluation and Monitoring of Tunnel Construction

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Lei Kou,
  • Ying Zhuang,
  • Hongzheng Luo,
  • Jian Liu,
  • Feng Guo

摘要

Accurate semantic segmentation of point cloud data is vital for safety monitoring and intelligent management in tunnel construction. However, challenges such as cluttered environments, occlusions, and the lack of annotated domain-specific datasets hinder effective application of deep learning techniques. To address these issues, this study presents TCS-Net, a novel point cloud segmentation network tailored for under-construction tunnels. A large-scale annotated dataset, named 3D Tunnel, was constructed using handheld laser scanning and contains over 60 million points across eight structural categories, filling a critical data gap in the field. TCS-Net introduces a multi-module fusion framework that combines spatial attention mechanisms, an inverted residual MLP (InvResMLP) for enriched feature representation, and a Kd-tree–based Gaussian upsampling with channel attention for enhanced feature propagation. An optimized training strategy incorporating AdamW, cosine decay, and label smoothing further improves learning robustness. Experimental results on the 3D Tunnel dataset demonstrate that TCS-Net achieves superior segmentation performance, with 94.38% mean IoU and 98.23% overall accuracy, validating its effectiveness and practical potential in tunnel construction scenarios.