Autonomous mobile robot navigation in urban environment

Abstract

Unmanned ground vehicles (UGV) is one of the highly versatile carriers for transportation, surveillance, search, and rescue task. For the service type mobile robot that ability to travel through indoor and outdoor environment may encounter complex challenges different than that of street vehicles. The urban pedestrian environment is typically GPS-denied which demands a further integrated approach of perception, estimation, planning and motion control to surmount. In this thesis, we present the design and implementation of Beobot 2.0?an autonomous mobile robot that operates in unconstrained urban environments. We developed a distributed architecture to satisfy the requirement for computationally intensive algorithms. Furthermore, we propose several biological-inspired visual recognition methodologies for indoor and outdoor navigation. We describe novel vision algorithms base on saliency, gist, image contour, and region segment to construct several perception modules such as place recognition, landmark recognition, and road lane detection. To conquer the latencies and update frequencies of each perception module while performing real-time navigation task. We further investigate hierarchical map representation to fuse the quick, yet limited state information while time-consuming but higher discriminating data remains in processing. We validated our system using a ground vehicle that autonomously traversed several times in multiple outdoor routes, each 400m or longer, in a university campus. The routes featured different road types, environmental hazards, moving pedestrians, and service vehicles. In total, the robot logged over 10km of successfully recorded experiments, driving within a median of 1.37m laterally of the center of the road, and localizing within 0.97m (median) longitudinally of its true location along the route.