Kläß, Jochen; Stückler, Jörg; Behnke, Sven (Autonomous Intelligent Systems, University of Bonn, Germany)
We present robust and efficient means for mobile robot navigation using a 3D representation of the environment. We build 3D surfel grid maps and propose Monte Carlo localization with probabilistic observation models for 2D and 3D sensors in such maps. In contrast to localization methods that utilize a 2D laser scanner in a static horizontal mounting, our method takes advantage of the 3D structure in the environment. This is useful, for instance, to localize in crowds of people: The robot can focus on the static parts of the environment above the person’s height. Finally, we extract navigation maps for 2D path planning from the 3D maps. Our approach avoids discretization effects and considers the complete height range of the robot to estimate traversability. In experiments, we demonstrate the accuracy and robustness of our approach for pose tracking and global localization – even in a crowded environment.