In low-light and weak-texture tunnel environments,3D laser scanning technology is commonly employed to collect point clouds,which are then processed using semantic segmentation algorithms to extract semantic information for environmental perception and understanding.Addressing the complexities of tunnel scenarios,this study introduces an improved tunnel point cloud semantic segmentation network(TDGCNN)based on dynamic graph convolution networks(DGCNN).The network employs a multi-scale strategy to enhance the edge convolution layers,thereby expanding the neighborhood scope to better capture edge features.Additionally,two attention mechanisms are incorporated to aggregate and refine features,effectively reducing redundancy.Experimental results demonstrate that TDGCNN achieves an overall accuracy(OA)of 98.78％and a mean intersection over union(mIoU)of 84.95％on real tunnel point cloud datasets,outperforming existing methods such as DGCNN,PointNet,PointNet++,and GeoSegNet.Validation on the WHU-TLS dataset confirms the robustness of TDGCNN across various tunnel scenarios.Ablation studies further validate the effectiveness of the integrated strategies,indicating that TDGCNN efficiently utilizes multi-level features to enhance the precision of tunnel point cloud semantic segmentation.
关键词
隧道三维点云/语义分割/多尺度/注意力机制
Key words
tunnel 3D point cloud/semantic segmentation/multi-scale/attention mechanism
维普
万方数据