Robot localisation algorithms with improved robustness

Abstract

This thesis addresses a number of difficult issues caused by data association failure in different robot localisation algorithms. The data association task identifies the correspondence between sensor measurements and physical objects in the environment . The focus of this thesis is on two specific problems . First, many indoor environments are filled with very similar artificial objects, such as rows of chairs or tables. In these environments, it is impossible to recognise each of the individual objects uniquely. A successful localisation algorithm should be robust to these non-unique object matches. Second, for Simultaneous Localisation And Map-building (SLAM), neither the robot nor the obstacle positions are precisely known at any given moment. The estimation algorithm can be vulnerable to occasional data association problems if only a single hypothesis is being maintained, as with the most prevalent Extended Kalman Filtering (EKF) based SLAM methods . The following principal theoretical contributions are presented in this thesis. First, the Landmark, 'friangulation, Reconstruction then Comparison (LTRC) algorithm is developed for panoramic vision based global localisation. It is a two-stage localisation algorithm, which combines the strengths of the feature and iconic localisation approaches. It improves the overall localisation performance by comparing higher dimensional image representations of the current and reference images, rather than low dimensional position estimates. Second, the Multiple Particle Filtering (MPF) approach is devised for SLAM. By exploring the conditional independence indicated by the factorised SLAM posterior, the robot pose and obstacle positions are estimated with separated low dimensional particle filters . The flexible MPF architecture benefits SLAM estimation. It allows the use of non-Gaussian noise models and maintains different hypotheses for each of the particles. At the same time, the time complexity is also improved to linear in the number of obstacles within the sensor coverage in comparison to the quadratic EKF SLAM algorithm. The practical implementation results support the theoretical analysis of the proposed algorithms .