关键词: 深度学习, 点云, 语义分割, LiDAR

Abstract: Point cloud semantic segmentation has become a key technology for achieving multi-semantic visualization in underground alleyway health monitoring.To address the low segmentation accuracy of target edges in alleyway scenarios,this paper constructs a large-scale point cloud semantic segmentation dataset for alleyway scenarios ALSD,and proposes a segmentation method adapted to this scenario.Based on backbone feature extraction,it introduces a local multi-scale neighborhood considering 3D curvature and a global spatial feature enhancement module,combined with an attention mechanism to improve the representation ability of small components and complex boundaries.An evaluation index system for alleyway scenarios is also established.On the ALSD dataset of real alleyway scenarios,we systematically analyzed the impact of training set size,input dimension,and hyperparameter settings on model performance.Experiments show that the proposed method achieves IoUs of 0.858,0.883,0.933,and 0.822 for pipes,supports,ground,and columns,respectively,with an mIoU of 0.865 and an overall accuracy (OA)of 98.7%.Compared with typical deep learning baseline methods such as PointNet++ and RandLA-Net,the proposed model achieves higher semantic segmentation accuracy on the ALSD dataset,which can provide high-precision 3D semantic support for the structural health monitoring of underground alleyways.

Key words: deep learning, 3D alleyway point cloud, semantic segmentation, LiDAR