Robust LiDAR SLAM framework for autonomous vehicles leveraging ground segmentation

Abstract

This dissertation proposes a robust simultaneous localization and mapping~(SLAM) framework leveraging ground segmentation. SLAM is an integral part of autonomous driving because a 3D map provides information about surroundings that helps robots to localize themselves with respect to the map coordinates. Recently, light detection and ranging~(LiDAR) sensors have been widely used to achieve mapping, localization, and perception due to their precise measurements compared with other sensors. This thesis begins with the observation that most cloud points within the 3D point cloud captured by a LiDAR sensor are from the ground. Since these ground points are redundant and featureless, so these points potentially degrade the performance of feature matching and object clustering. Therefore, in this thesis, various studies to tackle the problem above are proposed. This thesis mainly consists of four parts: a)~a fast and robust region-wise ground segmentation method as a preprocessing step, called \textit{Patchwork}, b)~LiDAR odometry utilizing the robust ground segmentation, c)~robust global registration using ground segmentation to perform loop closing in SLAM, called \textit{Quatro++}, and d)~static map building, called \textit{ERASOR}, based on the premise that most of the dynamic objects in the urban environments come into contact with the ground. In particular, this thesis emphasizes that ground segmentation should be robust against uneven environments and applicable without prior knowledge of the surroundings. Through these four steps, a 3D static map for robotic navigation can be successfully achieved. As verified by various experiments, it was shown that our proposed methods outperform the state-of-the-art methods.