Region Selective Stereo Vision

Abstract

With the world moving towards an era of self-driving cars and human-assisting robots in home and warehouse environment, the need for reliable, accurate measurement of the proximity of objects has become essential. One of the most well-studied methods of object localization in a 3D environment is the stereo-vision system. Stereopsis is responsible for depth perception capabilities in human beings. But unlike human eyes, the conventional stereo cameras used in computer vision do not have monocular degree of freedom. This thesis studies the monocular motions in stereo cameras. We further design an algorithm for a region-selective stereo vision system which is capable of monocular motions with better depth accuracy. This thesis discusses the various advantages and challenges of using region selective motion rather than a single fixation point. This work explores the idea of using stereo vision for applications where the cameras in the stereo rig are freely moving.