Point cloud semantic segmentation based on graph convolutional networks

Aiming at the problems of ground point redundancy, insufficient local feature representation and missing global context in airborne laser point cloud semantic segmentation, a semantic segmentation framework combining improved progressive triangulation filtering and hierarchical graph convolution is proposed in this paper. Experiments on DALES data set show that the overall accuracy (OA) of the proposed method is 97.5%, and the mean crossover ratio (mIoU) is 78.3%, which is significantly improved compared with KPConv, PointNet++ and other benchmark models. Ablation experiments verified that ground point pretreatment reduced the calculation time by 22% without affecting the accuracy, and multi-scale feature fusion improved mIoU by 14.5%. The research proves the technical advantages of point cloud semantic analysis based on graph convolutional networks.