3D semantic segmentation using deep learning for large-scale indoor point cloud

Abstract

3D laser point cloud can express complex large-scale 3D scenes. And yet, it is difficult to obtain the local structural model of each spatial point as the input feature to semantic segmentation. To address this issue, this work proposes a new 3D semantic segmentation method based on PointNet and PointSIFT model for large-scale indoor point cloud. First, several different sensing radius modules are built by PointSIFT model to extract the local features for 3D laser point cloud and form a multi-dimensional input features through the full connected layer. Then, the connection features of the PointNet network are full connected again, and the classification score of each point is obtained. Finally, the proposed deep neural network model is validated by the indoor dataset S3DIS. Experiments show that the overall and average accuracy of the proposed method for 3D laser point cloud classification are increased by 1.22% and 3.06%, verifying the accuracy of 3D semantic segmentation in complex indoor scenes.