Looking Inside for Mapping the Outside: Introspective Simultaneous Localization and Mapping

This thesis focuses on the development and improvement of Simultaneous Localization and Mapping algorithms. In particular, we are interested in reliable and effective techniques for unstructured large scale environments. While previous researcher focused on the analysis of class of algorithms or general effects of different algorithms (e.g. complexity, convergence or consistency), in this thesis we focus on analyzing the mapping process itself. At first, we present an introspection analysis that allows efficient optimizations for RaoBlackwellized SLAM on grid maps. The key idea is based on an analysis of the mapping process which allows us to perform filter updates conditioned to the state of the mapping system: localization, mapping or loop closing . We are able to update the complex posterior with substantially less resources by performing the computations only for a set of representatives instead of for all potential hypotheses. Extending this introspective analysis from a filter perspective to a general one, we find out that the SLAM problem can de decomposed in three main subproblems: a) incremental mapping, that is the process of providing local constraints between consecutive poses, in order to maintain the map locally consistent; b) loop closure: that is the process of finding global constraints among different parts of the map; c) map optimization: that is the process of combining local constraints (provided by the incremental mapper) and global constraints (provided by the loop-closure algorithm) to obtain an overall consistent map. We show solution to this three problems when using a laser range finder in both static and dynamic environmnents.