Semantic Segmentation of Building Components from UAV LiDAR Point Clouds for Structural Analysis Shiori Kubo (a*), Daigo Meguro (b), Hidenori Yoshida(b)
a) Institute of Education, Research and Regional Cooperation for Crisis Management Shikoku, Kagawa University, 1-1 Saiwaicho, Takamatsu 760-8521, Japan
b) Faculty of Engineering And Design, Kagawa University, 2217-20 Hayashicho Takamatsu Kagawa, Japan
Abstract
Delays in issuing Disaster Victim Certificates after major earthquakes in Japan hinder the prompt provision of public support and victim recovery. These delays are caused by the current damage assessment process, which relies on on-site visual inspections that are time-consuming, labor-intensive, subjective, and hazardous for personnel. To address these issues, this study proposes a foundational method for the automated extraction of structural building components from high-precision LiDAR point clouds acquired by drones. We employ the PointNet++ deep learning model, a network designed for unstructured 3D data, to perform semantic segmentation. The model classifies points into four categories, Roof, Wall, Ground, and Others, providing the necessary geometric data for subsequent inclination analysis. The method was validated on a real-world dataset from areas affected by the 2024 Noto Peninsula Earthquake, achieving a mean Intersection over Union of 77.3% and an Overall Accuracy of 88.6%. The Roof and Ground classes yielded excellent results with an IoU of approximately 85%. While the classification of the Wall class, with an IoU of 59.4%, remains a challenge due to class imbalance and geometric constraints, the model successfully recognized the complex 3D structure of both intact and tilted houses, a key capability for damage assessment. This demonstrates the potential of 3D point clouds to overcome the limitations of 2D imagery for structural assessment. The proposed method establishes a robust foundation for the future automated calculation of house inclination, which is expected to contribute to a more rapid, objective, and safe damage assessment process, accelerating the issuance of Disaster Victim Certificates for affected communities.
Keywords: Deep learning- Disaster management- Lidar- Semantic Segmentation- UAV
