Position estimation for a mobile robot using monocular vision and odometry

We describe a localisation system for a robot moving in a known environment. Unlike the currently used methods for industrial robots, our approach does not require any beacons to be installed: the system uses odometry to estimate the vehicle position continuously, and corrects this estimation when necessary by identifying some objects of the environment through vision. These objects, used as landmarks, were previously recorded in a data base. The different parts of the system are presented, particularly the way the uncertainty on odometry is updated and how prior knowledge (position estimation and data base) is employed to facilitate landmark identification. 7 cm on xy and 1 deg on the heading is the typical precision obtained in term of localisation. (authors). 36 refs., 26 figs