AUTONOMOUS VEHICLE LOCALISATION WITHOUT PRIOR SENSOR MAPPING

Abstract

This thesis presents alternative maps as substitutes to metric maps for autonomous localisation and will discuss three aspects of alternative non-metric maps usage - The challenge of inconsistent map scale, the advantage of annotated landmarks, and problem of dynamic obstacles. Firstly, a novel localisation algorithm using stochastic gradient descent learns the immediate map scale and most likely pose of the robot in a map. This achieved average translational and rotational errors of 0.598 m and 6.88o on metric maps and prevented loss of localisation in alternative maps. Secondly, Hot-NetVLAD, a visual place recognition algorithm, is developed to recognise unique landmarks labelled on alternative maps. The algorithm was extended to perform global localisation, showing excellent results for indoor global localisation of about 3 m and is independent to the metricity of the map. Finally, an algorithm utilising LiDAR-vision fusion is used to track and utilise dynamic objects to aid localisation, helping prevent localisation loss when heavily occluded by dynamic obstacles, improving localisation error from 0.590 m and 5.08o to 0.467 m and 4.78o in a crowded environment. By analysing and proposing methods to perform localisation with alternative non-metric maps, this thesis advances knowledge to perform localisation without prior sensor mapping.