Experience with optical flow algorithms based on the geometry of points in two frames

摘要

Abstract: A number of optical flow techniques exist for recovering the relative sensor-scene motion from image sequences, for subsequent use in vehicle guidance and navigation tasks. However, the problems of finding depths and motion, as well as various differential and kinematic approximations within the algorithms themselves, often lead to difficulties in the presence of noisy data. A re-examination of the basic geometry of the problem when posed as a set of matched points in two or more frames, each separated by an arbitrary motion, has produced a revised error metric for use with iterative solution methods. In experiments with small motions the method has been demonstrated to be less susceptible to noise than the standard linear formulation from which it is derived. Improved performance has also been observed in the recovery of rotational motion. !33