DipG-Seg: Fast and Accurate Double Image-Based Pixel-Wise Ground Segmentation

Ground segmentation on the 3D point cloud is fundamental to many applications, such as SLAM and object segmentation. As it is usually a preprocessing module of these applications, high efficiency and accuracy are the basic requirements for guaranteeing the whole system's performance. To this end, we avoid ground fitting and region division on the 3D point cloud. We propose a pixel-wise image-based method named DipG-Seg, which projects the 3D point cloud onto two cylindrical images, horizontal range-and z-images, then segments based on them. To realize fast and accurate ground segmentation, we first introduce innovative designs for image-based features. Specifically, we improve the slope feature with consideration of the LiDAR model and propose combining features with different sizes of receptive fields for better recognition of the ground. Then, based on these features, we devise a pre-segmenting pattern for pixel-wise classification. For fine segmentation, we devise a hierarchical refinement framework integrating a nonlinear filter and majority-vote kernel-based convolution, which is demonstrated to enhance the accuracy by over 7% on the basis of pre-segmenting. Comprehensive experiments were conducted on a real-world platform, SemanticKITTI, and nuScenes datasets. The results have demonstrated that our method can achieve an accuracy of 94.41% and a speed of 127 Hz on 64-beams LiDAR, outperforming the state-of-the-art methods and guaranteeing competitive robustness. Our method will be available at: https://github.com/EEPT-LAB/DipG-Seg.