Most applications for mobile robots rely on the fundamental challenge to determine the robots' poses within some reference frame. In the last decades, research on mapping and localization with mobile robots has produced functioning solutions for many situations. However, as robotic systems are exposed to a wide range of uncertainties\,---\,and, therefore, are typically modeled probabilistically\,---\,common approaches rely on various approximations. In this thesis, we investigate several of these approximations typically assumed for mapping and localization. This leads to a better understanding of existing methods and gives rise to new methods with improved accuracy and applicability. The first contribution of this thesis examines and advances the Bayes filter for Markov localization, the foundational framework for basically all recursive state-of-the-art robot localization algorithms. Despite its omnipresence in the context of robot localization and beyond, there has been a gap in the examination of a particular independence assumption in its derivation. We close this gap by deriving a more general form of the filter and demonstrate that it yields improved localization accuracy. Moreover, we carry out a theoretical examination to determine in which cases it is valid to use our augmented filter and in which cases it reduces to the well-known version of the Bayes filter. The second contribution is a decentralized algorithm for collaborative multi-robot localization that can cope with constraints that state-of-the-art approaches cannot satisfy. Existing methods typically neglect the correlations between individual robots' poses\,---\,which can lead to over-confidence\,---\,or introduce substantial communication or memory overhead. We avoid these drawbacks by introducing an approximate, decentralized update of the cross-correlation matrices. Our algorithm is the first fully recursive method that estimates the correlations under the constraint that robots can only exchange information at the instance of a relative detection. Thirdly, this thesis presents several improvements over state-of-the-art grid mapping methods for lidar sensors. We introduce a ray tracing based sensor model, namely the decay-rate model, and its consistent inverse model with the corresponding map representation. In contrast to many existing approaches, our model accounts for the exact distances the laser beams travel in each cell. Moreover, we develop a framework to compute the posterior distributions over the introduced decay-rate maps, as well as the posteriors over a general class of existing map representations. We derive the formulas to leverage the posterior distributions for localization and demonstrate that our approach improves the localization accuracy as compared to existing methods that assume a single map to be given. Lastly, we examine a potential use case for map posteriors beyond localization: detecting changes in grid maps based on information-theoretic scores on the map posteriors. In summary, this thesis introduces probabilistic methods for various aspects of mobile robot localization and mapping that improve and generalize existing approaches with minimal overhead regarding implementation, computation, and memory consumption.
Efficient generalizations of probabilistic methods for robot localization and mapping
2020-01-01
Hochschulschrift
Elektronische Ressource
Englisch
DDC: | 629 |
A probabilistic model for appearance-based robot localization
British Library Online Contents | 2001
|A Probabilistic Approach to Concurrent Mapping and Localization for Mobile Robots
British Library Online Contents | 1998
|Mobile Robot Simultaneous Localization and Mapping in Dynamic Environments
British Library Online Contents | 2005
|A Set Theoretic Approach to Dynamic Robot Localization and Mapping
British Library Online Contents | 2004
|