Abstract
Mobile robot localisation has been, and continues to be, a very active research area. Estimating the position of a mobile robot is fundamental for its navigation and map-building. This thesis addresses some of the problems associated with mobile robot localisation. Three distinct items of research presented in this thesis are (i) A systematic odometry error model for a synchronous drive robot; and (ii) A novel method to synchronise two independent sensor data streams, and (iii) A proposal for an exteroceptive truly odometric sensor system - 'Visiodometry'. Cyclops is a synchronous drive mobile robot. The kinematics causes the path of the robot to curve, with the degree of curvature affected by the orientation of the wheels. A systematic odometry error model is proposed to correct for this. The proposed model is supported both experimentally, and theoretically from modelling the kinematics. Combining sensor data from different sensor data streams is commonly done to improve the accuracy of estimated variables. However, in some cases the sensors are not networked making it impossible to synchronise the data streams. The second item of research proposes a novel method to estimate the time difference in the local clocks of the discrete sensor data from their time-stamps alone. A proposed enhancement to the method improves both the rate of convergence and the precision of the estimate. Results show that the method is more optimum and robust than one based on known methods, including those based on Gaussian assumptions. Wheel odometry is a common method for mobile robot localisation. However, wheel odometry is unreliable if there is wheel slip. In these environments visual odometry has been used. However, the method does not work well on planar surfaces or surfaces with fine texture. It is also unable to accurately detect small motions less than a few centimetres. The third area of research proposes an exteroceptive odometric sensor called 'visiodometry' which is independent of the kinematics and therefore robust to wheel odometry errors. Two methods are proposed (i) a dual camera 'shift vector' method and (ii) a monocular 'roto-translation' method. The results demonstrate that the proposed system can provide odometric localisation data in planar environments to a high precision. The method is based upon extracting global motion estimates of affine transformed images of the ground using the phase correlation method. Experimental results demonstrate that, as a proof-of-concept, this type of sensor input is an alternative genuinely odometric input which has the potential to be comparable in accuracy and precision to wheel odometry in environments where wheel odometry and visual odometry methods are unreliable.