A Novel Framework for Ground Segmentation Using 3D Point Cloud

Abstract

Ground segmentation is an essential preprocessing task for autonomous driving. Most existing 3D LiDAR-based ground segmentation methods segment the ground by fitting a ground model. However, these methods may fail to achieve ground segmentation in some challenging terrains, such as slope roads. In this paper, a novel framework is proposed to improve the performance of these methods. First, vertical points in the point cloud are filtered out by a gradient-based method. Second, a polar grid map is built to extract the seed points for model fitting. Moreover, the fitting-based method is used to model the ground. And a coarse segmentation result can be obtained by the fitted model. Next, the coarse segmentation result is used to update the ground height value for each grid in the grid map. Finally, the segmentation result is refined by the grid map. Experiments on the SemanticKITTI dataset have shown that the fitting-based method can achieve more accurate segmentation results by integrating with our proposed framework.