Integrating SLAM-DUNK and variable rate particle observers for fast multi-hypothesis SLAM

Abstract

In this thesis, the problem of SLAM with some set of prior hypotheses about the map, called Multi-Hypothesis SLAM, was tackled using a combination of an existing landmark-based SLAM algorithm called SLAM-DUNK and a particle filter inspired approach. SLAM-DUNK is a recent Kalman Filter-based algorithm for landmark tracking that scales linearly with the number of landmarks and converges globally to true landmark coordinates, thus far outperforming standard Kalman filter methods such as EKF-SLAM. This was combined with the Hybrid Variable Rate Particle Observer, a particle-filter inspired approach used here for hypothesis discrimination and particularly suited to tracking applications. An algorithm based on combining these two methods was formulated and shown to solve the problem quite successfully, with good resistance to complicating factors such as noise and number of hypotheses. The limitations and peak performance of this algorithm were also investigated by considering the ability of the algorithm to discriminate between very similar hypotheses and the conditions under which the algorithm performed best. Finally, the ability to extend this algorithm to dierent types of hypotheses, including Dense Maps owing to SLAM-DUNK's versatility, was considered and discussed.