Adaptive Patchwork: Real-Time Ground Segmentation for 3D Point Cloud With Adaptive Partitioning and Spatial-Temporal Context

Hao Ren, Mingwei Wang, Wenpeng Li, Chen Liu, Mengli Zhang

Abstract: Ground segmentation is a fundamental task in the field of 3D perception using 3D LiDAR sensors. Several ground segmentation methods have been proposed, but they often suffer from mis-segmentation, especially missed detection, due to poor noise removal, unreasonable ground pre-blocking and lack of rectification for segmentation results. To address these issues, a ground segmentation algorithm called Adaptive Patchwork is proposed, which is an extension of Patchwork++. Adaptive Patchwork adopts variable adaptive partitions instead of full a priori partitions, which alleviates the problem caused by point cloud sparsity. Secondly, based on the spatio-temporal correlation of ground point clouds, spatial and temporal correction is proposed, which integrates the spatio-temporal information of the ground with a small cost, significantly reducing missed and false detections. Finally, using the relationship between point cloud intensity and object reflectivity, noise removal based on multi-form features is proposed. The noise that affects ground segmentation is removed by using various point cloud features such as intensity, height and point cloud position prior. Validated experimentally on SemanticKITTI and Koblenz datasets, Adaptive Patchwork achieves satisfactory performance compared to state-of-the-art methods and is fast compared to existing plane fitting based methods.