Continuity properties of the appearance manifold for mobile robot position estimation

摘要

This paper describes the results of experiments using an appearance manifold of a scene to estimate the position and orientation of a mobile robot. These experiments use principal components of images of the environment as an orthogonal basis for representing appearance from different positions and orientations. A regularly sampled grid of images is projected into a set of N principal component images to produce a 3D manifold within an N dimensional appearance space.The orthogonal basis serves as an index for generating hypotheses of position and orientation from a newly acquired image. The basis provides a similarity metric for matching images. Distance from the projection of an observed image to the manifold gives an estimate of the confidence of a position estimate. Reconstructing the scene from the manifold allows obstacle detection.In these experiments, we examine the effects of variations in image resolution, number of basis images, and density of samples used for constructing the manifold. We investigate the use of interpolation over the appearance manifold in order to improve the precision of the estimated position and orientation. We demonstrate position estimation with 1D, 2D and 3D manifolds. Our results indicate that such an approach can provide a fast and reliable method for visual navigation.