Outdoor vehicle navigation using passive 3D vision

A prototype 3D vision system which uses the structure-from-motion principle is described. The system is based on the extraction and matching of image tokens such as edges and corners. The 3D locations of the causative features in the viewed scene are deduced by triangulation and are updated with each observation by application of Kalman filters. The operation of the vision system is illustrated with results obtained from sequences of natural outdoor imagery. The output of the vision-system kernel is a rich 3D representation of feature points and edges in the scene. A method is described by which surfaces can be simply constructed from this information. Results indicate that the inferred surfaces can be readily segmented into navigable and nonnavigable regions. This segmentation can be used by a path planner to select a safe and economical route for subsequent vehicle travel. A hardware realization of the vision system is discussed