关键词: 三维激光点云, 语义分割, PointNet++, 杆状物, 深度学习

Abstract: Aiming at the problems of manually designed features for data types, poor universality, and low automation in the extraction of existing road rods, a road rod semantic segmentation method based on an improved PointNet++ deep learning network is proposed in this paper, which realizes the segmentation of road rods. First, the parameters of the original network model such as receptive field and block size are adjusted to make the model more suitable for road point cloud data.And then,aiming at the problem of unbalanced point cloud data, the focus loss function is used as the loss function of the model, so that the categories that occupy a relatively small proportion can be fully trained.At last, to address the problem of the PointNet++ network not considering the relationship between the features of each point in the neighborhood when extracting features, a neighborhood feature aggregation module is used to fuse neighborhood information and improve the learning ability of the network model for point cloud features. To verify the effectiveness of the proposed method, an improved network model was used to conduct experiments on a self-built dataset composed of road point clouds. Compared with the classic PointNet++ network, the segmentation accuracy of rod-shaped objects was significantly improved. The intersection over union (IoU) on simple and complex roads increased by 8.44% and 15.25%, respectively, reaching 98.88% and 92.50%.

Key words: 3D laser point cloud, semantic segmentation, PointNet++, rod, deep learning