关键词: 点云语义分割, 通道注意力机制, 空间几何特征, PointNet++

Abstract: The tunnel point cloud semantic segmentation technology based on deep learning can recognize and classify objects in large-scale point cloud data,and realize the extraction and management of object information in tunnel scenes. A subway tunnel point cloud semantic segmentation network model based on enhanced spatial geometric feature fusion is proposed in this paper,which takes into account the spatial distribution characteristics and geometric features of subway tunnel facilities. A spatial geometric feature extraction module is designed to extract the relative spatial positions and geometric distribution features of tunnel facilities and combines them with the features of the network encoding layer through channel concatenation,enhancing the network′s perception of multi-scale feature information. A feature fusion encoding layer based on channel attention mechanism is designed to extract the weight of features between different channels,and to perform weighted fusion of the information from different spatial scales,so as to fully utilize information of different scales to improve model representation and generalization abilities. The network model proposed in this paper is verified using a dataset made by the point cloud data of a tunnel in Nanjing Subway. The experimental results demonstrate that the proposed network achieved a training mIoU value of 0. 955 6. The predicted results on the test dataset showed a weighted average F1 score of 0. 995 9 and a weighted average IoUvalue of 0. 963 1. For the communication cables,pipeline racks,and contact hangers classes with poor IoU values in the PointNet++ network,the IoU values of the proposed network reached 0. 845,0. 825,and 0. 999,respectively,which effectively improved the overall accuracy of point cloud segmentation of subway tunnel and provided a technical reference for realizing automatic disease inspection and facility management of subway tunnel.

Key words: point cloud semantic segmentation,channel attention mechanism,spatial geometric feature,PointNet++