Structure-From-Motion Based On Information At Surface Boundaries

点击翻译按钮获取中文摘要

Existing computational models of structure-from-motion are based on variations of optical flow or feature point correspondences within the interior of single objects. This project investigates a new, boundary-based cue for determining motion and shape. An approach is outlined for determining that an object is rotating in depth. Only simple, qualitative computations are required, avoiding the need to accurately determine how flow patterns are varying over the surface of interest. Generalizations involving integration of boundary and surface cues will be explored. Ultimately, the approach will be extended to other types of motion such as translation of a wide field-of-view cameras.To support this boundary-based analysis improved methods will be evaluated for estimating optical flow in the presence of flow discontinuities. The proposed algorithm integrates flow/surface analysis with the structure-from-motion determination while addressing certain deficiencies in previously proposed techniques. Results from this study should provide more general reliable, and economical methods for dealing with the important problems involved in analyzing visual motion.