Inferring spatial structure from feature correspondences (machine, computer vision, robotics, mobile, artificial intelligence)

Abstract

Building a three-dimensional map of the local environment is a critical task for many computer vision systems. This dissertation discusses two methods for building such a map: model-based vision and image sequence analysis. A model-based vision system uses previously stored models of objects to recognize objects in images and to estimate their locations relative to the camera. In the form of image sequence analysis discussed here, a mobile robot's vision system compares two or more images formed at different locations to infer the structure of the scene and the robot's path through it. The aspect of model-based vision considered is the estimate of an object's location once the object has been recognized and correspondences between image and model features established. New algorithms are introduced for models consisting of simple geometric features: a set of coplanar points, a set of coplanar lines, a circle, and a set of general lines in space. Next a simple technique for nonlinear, sequential estimation is applied to a problem in image sequence analysis. A camera moves through an planar scene consisting of a set of points; the problem is to estimate the camera locations and the scene points' positions from the images. Processing images sequentially saves computation and provides information as the camera traverses its path. Nonlinear estimation theory enables transforming prior distributions of image noise into posterior distributions on scene points' positions and camera locations that are used to interpret subsequent images. Finally a technique for image sequence analysis to exploit knowledge of the camera path is presented. Bolles and Baker's Epipolar-Plane Image Analysis, originally applicable to linear camera paths and fixed viewing directions, is extended to unrestricted viewing directions. Next duality in the projective plane is used to extend the analysis to general paths in the plane and to curved and moving objects; this technique is applied to an image sequence collected by the SRI mobile robot to reconstruct a planar slice of the scene. Duality in projective space is used to derive the extension to general paths in space.