Simultaneous localisation and mapping with prior information

This thesis is concerned with Simultaneous Localisation and Mapping (SLAM), a technique by which a platform can estimate its trajectory with greater accuracy than odometry alone, especially when the trajectory incorporates loops. We discuss some of the shortcomings of the "classical" SLAM approach (in particular EKF-SLAM), which assumes that no information is known about the environment a priori. We argue that in general this assumption is needlessly stringent; for most environments, such as cities some prior information is known. We introduce an initial Bayesian probabilistic framework which considers the world as a hierarchy of structures, and maps (such as those produced by SLAM systems) as consisting of features derived from them. Common underlying structure between features in maps allows one to express and thus exploit geometric relations between them to improve their estimates. We apply the framework to EKF-SLAM for the case of a vehicle equipped with a range-bearing sensor operating in an urban environment, building up a metric map of point features, and using a prior map consisting of line segments representing building footprints. We develop a novel method called the Dual Representation, which allows us to use information from the prior map to not only improve the SLAM estimate, but also reduce the severity of errors associated with the EKF. Using the Dual Representation, we investigate the effect of varying the accuracy of the prior map for the case where the underlying structures and thus relations between the SLAM map and prior map are known. We then generalise to the more realistic case, where there is "clutter" features in the environment that do not relate with the prior map. This involves forming a hypothesis for whether a pair of features in the SLAM state and prior map were derived from the same structure, and evaluating this based on a geometric likelihood model. Initially we try an incremental Multiple Hypothesis SLAM (MHSLAM) approach to resolve hypotheses, developing a novel method called the Common State Filter (CSF) to reduce the exponential growth in computational complexity inherent in this approach. This allows us to use information from the prior map immediately, thus reducing linearisation and EKF errors. However we find that MHSLAM is still too inefficient, even with the CSF, so we use a strategy that delays applying relations until we can infer whether they apply; we defer applying information from structure hypotheses until their probability of holding exceeds a threshold. Using this method we investigate the effect of varying degrees of "clutter" on the performance of SLAM.