Vision-based localization for mobile robots using a set of known views

Abstract

A robot localization scheme is presented in which a mobile robot finds its position within a known environment through image comparison. The images being compared are those taken by the robot throughout its reconnaissance trip and those stored in an image database that contains views taken from strategic positions within the environment, and that also contain position and orientation information. Image comparison is carried out using a scale- dependent keypoint-matching technique based on SIFT features, followed by a graph-based outlier elimination technique known as Graph Transformation Matching. Two techniques for position and orientation estimation are tested (epipolar geometry and clustering), followed by a probabilistic approach to position tracking (based on Monte Carlo localization). © Springer-Verlag Berlin Heidelberg 2008.