A Randla-Net-Based Approach to Segment Roadway Objects in Large-Scale Point Clouds

Mobile Laser Scanning (MLS) systems provide fast and easy access to dense and accurate point clouds of roadways. Currently, semantic segmentation of urban roadway objects in these data plays a significant role in urban modelling, autonomous driving, intelligent transportation, etc. However, the semantic segmentation of large-scale urban roadway point clouds is still challenging due to the unstructured and disordered nature of point cloud and the enormousness of points acquired during practical applications. Based on that, we propose an automated target objects semantic segmentation approach based on RandLA-Net. With this procedure, we first perform sample annotation, then train a RandLA-Net for semantic segmentation, and finally evaluate the model on test sets. The mean IoU reaches 76.54%, and the recall score metrics for semantic segmentation of buildings and rods-like objects reach 96.44% and 93.66%, respectively. From the visualized results, we achieved clean building contour segmentation and robust, consistent rods-like object segmentation in up to 10 ⁸ points urban roadway scenarios.