Fast Ground Segmentation for 3D LiDAR Point Cloud with Height-based Candidate Values

LiDAR plays a crucial role in self-driving technology as it provides advanced detection capabilities that enable Autonomous Vehicles (AVs) to gather extensive environmental information. Ground segmentation is of utmost importance for AVs, as it allows them to effectively navigate and recognize neighboring objects. Unfortunately, some current algorithms suffer from issues such as unavailability on complex terrains, and excessive time and memory usage. To tackle these problems, this paper presents a novel ground segmentation framework. First of all, the point cloud is projected onto a 2 D plane, followed by grid region segmentation. Second, candidate grids for the ground are identified based on the flatness criterion, and height-based candidate values are obtained by using Kernel Density Estimation (KDE) on candidate grid heights. Third, the ground segmentation is performed by using region-wise ground plane fitting with height-based candidate values. Finally, experimental verification on SemanticKITTI datasets demonstrates the promising performance of the proposed framework, outperforming three existing methods in terms of segmentation accuracy. Furthermore, the proposed framework achieves an impressive average processing time of only 0.05 s for a single scan data from a 64-beam LiDAR, fully satisfying the real-time performance requirement of AVs.