ABSTRACT
Addressing the issue of structural deformation in railway tunnel lining due to the long-term influence of geological conditions and vehicle loads, as well as the precision limitations of existing laser scanning detection technology caused by interference from tunnel auxiliary facilities, this study proposes a novel semantic segmentation method for railway tunnel 3D point clouds based on PointNet++. The dataset creation involves stages such as data collection, point cloud segmentation, point cloud subsampling, normal vector calculation, and point cloud annotation. The PointNet++ network, embedded in a deep learning framework, is applied to extract geometric features, refining the lining point clouds. The method’s performance, evaluated using Precision, Recall, F1 score, and Intersection over Union, significantly exceeds that of the traditional PointNet network, highlighting its excellence in point cloud segmentation and its contribution to the safe operation of railway tunnels.
