Efficient LiDAR-Based Mapping and Localization in Outdoor Environments

Abstract

Autonomous robotic systems can help people in their daily lives. Robots can do tasks people cannot or simply do not want to do. Service robots like autonomous vacuum cleaner, lawn mower or even autonomous driving taxis are likely becoming an integral part of our lives. Meanwhile, the industry values the precision, efficiency, and capacity of robots to support their production lines. For reliable and safe operation, mobile robots usually rely on some sort of map of the target environment. For that, a robot needs to know its location within the map to utilize the information that is stored in there to the full extent. Many robots are equipped with sensors, like laser scanners, which can be used to figure out their location within the map, based on the current sensor observations. This task of finding the own position within a given map is usually called localization, and is a well studied problem in robotics. Established methods have shown great success, especially in smaller, mostly indoor environments. However, with the rise of advanced systems, like autonomous driving cars, it remains an open question if those techniques are scalable to such an extent at high precision. <br/> In this thesis, we focus on larger-scale maps in outdoor environments such as those encountered in the automotive domain. Due to the sheer size of 3D maps, we have to develop techniques for efficiently representing the data. The resulting map needs to be compact in memory, but also usable for the localization of the robots. Many previous works focus on either compressing the data, or the localization algorithm. However, there is little research tackling both at the same time: trying to build memory efficient maps which are well suited for localization. <br/> To tackle these problems, we propose several algorithms towards city-scale mapping and localization. We start with the fundamental task of calibrating the sensors of robotic platforms to obtain consistent and undistorted data that is needed for all the subsequent tasks. We then describe a method for building consistent point cloud maps using the raw recorded sensor observations. Afterward, we investigate how to utilize machine learning to compress our point cloud maps to be more memory efficient. The remaining thesis focuses on localizing robots in such maps. When navigating at a later point in time through the environment, we want to find out our position with respect to the compressed map without any further cues such as GNSS. For this, we have developed an algorithm that first coarsely figures out in which part of the map the robot is located. Afterward, we look into a fine registration where we aim at centimeter-accurate localization. Note that both, the coarse localization and the fine registration, operate directly on the compressed representation to enable localization in a compressed map. Once we have found our initial position, we can track the robots' movements within the map using our developed pose-tracking method. By this, we can estimate for any point in time the position of the robot in the map. Knowing the robot's position allows relating and fusing the measurements from the robot with the available information that is stored in the map. Additionally, it enables subsequent tasks, like path planning, which require the robot's position. <br/> In this thesis, we advanced towards mapping large-scale environments and localizing in those resulting maps. The methods proposed here have been evaluated on publicly available datasets and are published in peer-reviewed journals and conferences. The software and implementations of our methods are open-source to enable new research to be built upon our works.