Visual robot guidance for an insertion task

Vision algorithms are proposed for an automated robot assembly task. The goal is to insert cylindrical objects with arbitrary crosssection shapes into narrowly matching holes. Scenes have variable compositions and the viewpoint of the camera is rather arbitrary. An affinely invariant shape description is used to compare inflection bounded segments of the image contours with similar segments describing a set of shapes in a model database. Once all matching image — model segment pairs have been found, they are grouped into complete model instances found in the image, by a consistency check algorithm. Besides recognition, the determination of object pose and position is described. A new method for determining camera focal length using image features is presented.