Representations and Algorithms for Recognizing and Locating Three Dimensional Curved Objects from Monocular Images

点击翻译按钮获取中文摘要

This award supports research on ways for computer vision systems to exploit the Computer Aided Design models available for many manufactured objects. The goal is to learn how to build a vision system capable of recognizing and locating instances of curved 3D objects modeled by parametric surface patches and their intersection curves in the imperfect line drawing obtained by edge detection from a single image. Ongoing implementation of research system building blocks will be completed, including specialized components of a solid modeler, algorithms for constructing exact aspect graphs of object models, and techniques for locating model instances from image contours. Novel recognition strategies will be investigated, based on both object- and viewer-centered representations. New data structures and algorithms will be developed that efficiently implement these strategies; these will be integrated into a working recognition system. A notable aspect of this research is the systematic characterization of geometric constraints by polynomial equations, which may then be solved by application of available computational and mathematical tools.