Use of eigenspace techniques for position estimation

Abstract

Presents a method for solving the problem of mobile robot localisation in an indoor environment. It is based on an approach in which the scene is "learned" by taking images from a large number of positions (x, y) and orientations (/spl theta/) along a 2D grid. The positioning problem is then reduced to a problem of associating an unknown input image with a "learned" image. A brute force solution requires too many image correlations, and would be computationaly too expensive to provide a system running in real-time. To overcome this problem, the image set is compressed using principal components analysis, converting the search problem into an addressing problem. The aim is to obtain a low dimensional subspace, in which the visual workspace is represented as a continuous appearance manifold. This method allows a process to determine the current robot pose by projecting an unknown input image into the eigenspace, and comparing its exact position to the appearance manifold.