Most of the vision systems used for factory automation acquire straight pieces of two-dimensional information yielded from objects. By analyzing this information, the vision systems achieve tasks such as inspection, �pick and place� and assembling of the objects. To carry out these tasks, many kinds of image processing algorithms for binary, gray and color image data have been proposed and applied to practical tasks. However, it has been difficult for such systems to handle objects with unfixed aspects or complex three-dimensional objects. Therefore, a vision system suitable for 3D objects and applicable in unconstrained environment has been expected. We have developed a stereo vision system to inspect and recognize complex 3D objects which conventional vision systems could not cope with. This system is a binocular and feature-based vision system in consideration of both the environment to be used and the target objects. Especially, since it is easy to extract geometrical features such as line, circle and ellipse from images of automobile parts, the vision system, which recovers 3D geometrical features using edge primitives extracted from the left and right gray images, has been developed. This paper describes the developed stereo vision system and shows its performance using experimental results.
�@