A Novel Method for Semantic Segmentation on Lidar Point Clouds

Autonomous driving relies on multiple sensors, such as lidar and cameras, to perceive the surrounding environment and the vehicle’s own position. Among them, lidar point cloud segmentation is a crucial and challenging task for 3D scene understanding. In this paper, we propose a novel deep learning method RPNet for lidar point cloud segmentation that combines range image-based segmentation and point based segmentation. Our method extracts point cloud features from range images and predicts 3D point cloud labels from point clouds. The segmentation results of both branches are fused to improve accuracy. We evaluate our method on the Semantic KITTI dataset and show that it outperforms other fusion algorithms in terms of effectiveness and robustness.