3-D Vision

Abstract

Machine vision research has centered on the problem of obtaining depth and surface orientation from an image, creating what has been called by some authors the ‘2½-D’ sketch [39]. Currently, there are several sensing systems that can derive depth or surface orientation from a scene. Among these are laser range finders[1,38,64], photometric stereo [29] and binocular stereo [5]. Laser imaging is potentially hazardous and has difficulty with shiny metal reflective surfaces. At present, it is a more expensive depth sensing technology than the other methods mentioned above. Photometric stereo puts great demands on the illumination in the scene and on properly understanding the reflectance properties of the objects to be viewed. We have chosen to use binocular stereo in this work because it has the advantage of low cost and the ability to perform over a wide range of illuminations and object domains. It is also a well understood and simple ranging method, which motivates its use in a generalized robotics environment where many different task and object domains may be in effect. Used as a single robotics sensing system, however, stereo has clear deficiencies. If there is a lack of detail on the object, only sparse measurements are possible. If too much detail is present, the matching process can easily become confused. Detail also causes a marked degradation in performance as the potential match space increases.