A Computational Paradigm for Three Dimensional Scene Analysis

Abstract

This paper presents a computational paradigm for a system which will dynamically model the contents of a three diinensional scene. The dynamic scene model may be made available to processes which analyze and interpret the scene as a composition of objects, and processes which plan and execute actions based on the composition of the surfaces or objects in the scene. This computational paradigm is presented as a collection of processes and data structures, many of which are currently areas of active research. The systern receives information in the form of a time sequence of stereo images. These images are immediately converted into an "initial representation" which facilittes the processing of later stages. The "initial representation" is then passed to a number of independent processes called "shape experts". Each shape expert extracts information about three dimensional surfaces from a different source. Surface information is integrated with the information obtained over time to maintain a "Cornposite Surface Model". The Composite Surface Model is then made available to processes for planning, analysis, or object recognition. The framework is introduced, and then each of the components are examined. The problems associated with each component are discussed, and a brief description is given of current research in that area.