Self-adaptive and robust ground segmentation for autonomous driving system in outdoor environment

Abstract

Ground segmentation is one of the essential techniques for perceiving the surroundings using 3D LiDAR (Light Detection and Ranging). It is usually utilized as a preprocessing stage for various tasks such as SLAM (Simultaneous Localization and Mapping), object detection and semantic segmentation. Therefore the following three conditions are generally considered important. First, it should guarantee high recall with little variance. It is because a ground segmentation failure, specifically an under-segmentation, may result in failures of subsequent algorithms. Second, it should be applicable to various environments. It is because the ground in the real world is not always flat, but it has a variety of shapes, such as bumpy roads, rough terrain, and steep slope. Finally, it should be fast enough to be used as a preprocessing stage in real-time applications. In this thesis, a fast, robust, and adaptive ground segmentation method, which is applicable to various environments while avoiding under-segmentation, is proposed. This method divides a point cloud into multiple regions using Concentric Zone Model, then estimates the ground plane based on Principal Component Analysis using points at low heights as inliers for each region. Moreover, it estimates the ground likelihood for the estimated ground plane using the previous ground information in an adaptive manner. Finally, it reverts the ground plane using neighborhood ground estimation results to avoid under-segmentation of ground with different characteristics from previous cases. The proposed method is evaluated quantitatively using publicly available datasets, SemanticKITTI and NuScenes, and qualitatively using real-world data acquired by a quadruped robot in an outdoor environment.