Learning-Based Subcategory Ground Segmentation Model for LiDAR Point Cloud

Ground segmentation based on LiDAR point clouds is one of the crucial tasks in 3D perception, directly impacting the recognition of traversable areas for mobile robots. Current mainstream methods rely on manually designed features and parameter tuning, requiring adjustments for different environments. This approach is not only time-consuming but also results in unstable robustness. This paper proposes a model named GroundSeg, which uses deep learning to automatically learn and extract distinguishing ground features. It not only accurately segments the overall ground but also performs semantic segmentation of ground subcategories. In our ground segmentation model, we introduce cylindrical voxelization to match the physical characteristics of LiDAR point clouds. We then propose a Channel-Plane Extraction (CPE) Module to ex-tract both local and global distinguishing features of the ground in three-dimensional voxels. Finally, we incorporate a class-balanced loss group to achieve precise ground segmentation and subcategory classification. Experimental results show that our proposed model has significant advantages in the integrity of ground segmentation and outperforms existing traditional and learning-based methods in F-l score. This is the first work to consider refined category ground segmentation, making it more meaningful for practical robotic applications. All qualitative and quantitative evaluation experiments were conducted on the standard SemanticKITTI dataset.