Neural Geometric Engine Feasibility Study.
Abstract
The goal of this project is to research and develop a neural geometric engine for rapidly determining geometric relations between parts of a scene from sensor images. The subject of building a spatio-geometric and kinetic model of file scene from images was considered "image understanding" or "early vision" in artificial intelligence research. The approach we have taken to spatio-geometric modeling of the scene is a smart sensor approach. It is fundamentally different from the current art. The novel neural computing system is based on Lie group model of neural processing in primate's visual cortex. Termed "information processing" approach to vision by David Marr, the pioneer of computational vision research, the current art of early vision is build upon the concept that the spatio-geometric information can be extracted by processing the image data, and the process can be formed as a computer algorithm. While the term "information processing approach" sounds very general, it does lead to a specific method of algorithm design. Particularly, it was suggested that in order to determine the changes (motion, binocular disparity, geometric distortion) in images and to further infer the scene geometry and motion, or register images, the first step should be to determine how a point on the image plane is moved to another place. It was further suggested that a process of feature detection followed by feature matching will do the job. All the spatio-geometric information are considered directly or indirectly derived from feature matching. (KAR)
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 25, 1995
- Accession Number
- ADA301093
Entities
People
- Thomas Tsao