Building, Registrating, and Fusing Noisy Visual Maps

Abstract

This paper deals with the problem of building three-dimen sional descriptions (we call them visual maps) of the environ ment of a mobile robot using passive vision. These maps are local (i.e., attached to specific frames of reference). Since noise is present, they incorporate information about the ge ometry of the environment and about the uncertainty of the parameters defining the geometry. This geometric uncertainty is directly related to its source (i.e., sensor uncertainty). We show how visual maps corresponding to different positions of the robot can be registered to compute a better estimate of its displacement between the various viewpoint positions, as suming an otherwise static environment. We use these esti mates to fuse the different visual maps and reduce locally the uncertainty of the geometric primitives which have found correspondents in other maps. We propose to perform these three tasks (building, registrating, and fusing visual maps) within the general framework of extended Kalman filtering, which allows efficient combination of measurements in the presence of noise.