Scholarlink

This study investigates the factors affecting crack detection accuracy using three-dimensional point clouds acquired by Terrestrial Laser Scanners (TLS). The research focuses on understanding how measurement conditions—including scanning distance, incidence angle, and point cloud density—influence the ranging results by TLS. In the crack detection procedure, PointNet architecture is employed for binary crack classification, with various input feature combinations evaluated including 3D coordinates, RGB values, intensity data and local geometric features derived from eigenvalue decomposition. Experiments were conducted on a reinforced concrete headworks structure constructed in 1976, using both TLS and Handheld Laser Scanner (HLS) measurements for ground truth validation. Results demonstrate that local geometric features significantly improve detection performance, achieving improvement of F1 score compared using only 3D coordinates. The study reveals that point cloud density, determined by scanning distance and incidence angle, critically affects ranging accuracy in the depth direction. Optimal neighborhood radius for local feature extraction varies between planar and curved surfaces. These findings provide practical guidance for improving crack detection accuracy in civil engineering applications using point cloud data and deep learning methods.

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

Point Cloud-Based Damage Detection for Reinforced Concrete Structures by Deep Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kazuma Shibano,
  • Toma Tsubota,
  • Moeka Mukai,
  • Hiromu Tanaka,
  • Hikaru Umezawa,
  • Ninel Alver,
  • Tetsuya Suzuki

摘要

This study investigates the factors affecting crack detection accuracy using three-dimensional point clouds acquired by Terrestrial Laser Scanners (TLS). The research focuses on understanding how measurement conditions—including scanning distance, incidence angle, and point cloud density—influence the ranging results by TLS. In the crack detection procedure, PointNet architecture is employed for binary crack classification, with various input feature combinations evaluated including 3D coordinates, RGB values, intensity data and local geometric features derived from eigenvalue decomposition. Experiments were conducted on a reinforced concrete headworks structure constructed in 1976, using both TLS and Handheld Laser Scanner (HLS) measurements for ground truth validation. Results demonstrate that local geometric features significantly improve detection performance, achieving improvement of F1 score compared using only 3D coordinates. The study reveals that point cloud density, determined by scanning distance and incidence angle, critically affects ranging accuracy in the depth direction. Optimal neighborhood radius for local feature extraction varies between planar and curved surfaces. These findings provide practical guidance for improving crack detection accuracy in civil engineering applications using point cloud data and deep learning methods.