Conventional SLAM (Simultaneous Localization And Mapping) methods do not use prior information about the environment. They estimate the most likely map and position from a set of observations. In order to correct for the cumulative position error, loop closure must be performed. This is not always possible, especially in autonomous driving, rescue, and surveillance operations when a specific position needs to be approached quickly and directly. In addition, highly accurate maps are often required for reliable localization and navigation. However, in unstructured environments with a lot of vegetation and far away from highways and big cities, HD-maps are available only to a limited extent or not at all. In this thesis, a cooperative SLAM method is presented that enables a vehicle to localize at locations that have never been actively observed with sensors before. Inspired by human cognitive abilities, global localization is performed along a path description called a path map. The path map is a topological metric description of the environment and can be generated automatically, for example, based on OpenStreetMap data. By integrating the path map into the SLAM process, the accumulated position error can be compensated and the local SLAM map can be globally referenced to provide the updated map to other vehicles. For this purpose, the results of Monte Carlo localization with novel measurement model for global localization, multi-target tracking method for landmark determination, and B-spline-based road course estimation are probabilistically fused in a graph SLAM approach. For navigation in unstructured environments, scene understanding of the environment is required in addition to landmarks for localization. A novel LiDAR-based method can immediately make the distinction between obstacles, free space, and vertical objects. Using recursive estimation techniques, the states of all objects in the environment are determined. Based on the velocity estimation and state uncertainty, static objects are then detected and considered as landmarks in the SLAM method and Monte Carlo localization. In numerous test drives covering a total distance of 56 km, the system was extensively tested in various areas at different times of the year for the autonomous and split convoy use case. The evaluations demonstrate that topological metric maps for navigation can be combined with a sensor-based perception system to solve the challenges of autonomous navigation in unstructured environments as well as in difficult weather conditions. It has been shown that the selection of abstract landmarks for localization provides a robust map representation. Furthermore, the method is universally applicable and allows for extension to visual object landmarks.
«Conventional SLAM (Simultaneous Localization And Mapping) methods do not use prior information about the environment. They estimate the most likely map and position from a set of observations. In order to correct for the cumulative position error, loop closure must be performed. This is not always possible, especially in autonomous driving, rescue, and surveillance operations when a specific position needs to be approached quickly and directly. In addition, highly accurate maps are often require...
»