AbstractsComputer Science

Robust graph-based localization and mapping

by Pratik Agarwal




Institution: Universität Freiburg
Department: Technische Fakultät (bisher: Fak. f. Angew. Wiss.)
Degree: PhD
Year: 2015
Record ID: 1106041
Full text PDF: http://www.freidok.uni-freiburg.de/volltexte/2015/10033/


Abstract

Localization and mapping are fundamental requirements to enable mobile robots to operate robustly in their environments. The solution of these two problems will allow robots to assist humans in important tasks such as autonomous driving or hazardous search-and-rescue missions. These complex navigation tasks require a robot to build a map while simultaneously localizing itself in it. It is relatively straightforward to build a map with perfect sensors. The challenge arises when a mobile robot needs to perform both mapping and localization with noisy sensors. This requires a robot to reliably recognize a previously visited place, while being robust to noise in sensor measurements and perceptual aliasing due to repetitive structures in the environment. In this thesis, we have proposed novel simultaneous localization and mapping (SLAM) algorithms which are more robust to place recognition and sensor errors. Our methods can recover from errors resulting from visual aliasing due to ambiguity in the environment. Additionally, our methods are more robust to gross non-Gaussian errors in sensor measurements. As map building can be cast as an optimization problem, mapping algorithms should be robust to poor initialization. Furthermore, the methods we have developed enable robots to mitigate the effects of poor initializations. In contrast to other methods that first find a good initialization to seed the optimization or that remove outliers before optimization, our method is more elegant as it does not require an additional pre-processing step. We also survey several geodetic mapping methods with an aim at providing a geodetic perspective of SLAM, as both fields share similarities when it comes to building large maps. We show that many state-of-the-art robot mapping algorithms have a direct relation to geodetic mapping techniques developed many decades earlier. Our survey will possibly inspire new solutions to large-scale, autonomous robotic SLAM methods. Finally, we believe that robots should be able to navigate with already available maps even if they are built for humans or for other purposes. This will save significant computational and sensor requirements on the robot, reduce the setup time, and allow them to operate in environments without previously visiting them. In this regard, we outline a novel approach to metric localization that does not require the construction of a new map when one already exists. The central idea is to leverage geotagged imagery on Google Street View as an accurate source for global positioning and localize a robot using only a monocular camera with odometric estimates. We have released open source implementations of our methods and tested them extensively on simulated as well as real-world datasets collected on mobile robots. Our claims are supported by comparisons to existing state-of-the-art methods and verified by other researchers. We believe that our methods will bring us closer to a future with mobile robots navigating autonomously and assisting humans in their daily tasks. Die…