Mobile Robot Localisation using Active Visual Sensing

In autonomous navigation in intially unknown surroundings, localisation over extended periods requires the discovery and use of enduring landmarks which can be repeatedly referred to. Active vision is well-suited to (re-)finding such landmarks, its wide field of view permitting the same features to be viewed during large robot motions. This paper describes the implementation of a robot system, comprising mobile base and stereo active camera platform, which is able to detect, store and track suitable landmarks during goal-directed navigation. We present an active strategy for selecting which available set of landmarks to track at a certain time to best improve position and map knowledge, and suggest rules for adding to or deleting from the map to maintain its usefulness. Maintenance of a map of landmark features requires the propagation of quite complete information about their positional uncertainties and how they relate to ego-motion estimates. This mitigates the problem of motion drift often encountered in sequential navigation or structure from motion approaches, where systematic errors build up in both feature and robot location estimates. The high-quality map of our system is necessarily sparse due to the detailed uncertainty information, but the paper shows how useful maps can be processed in real-time.