Redundant tracking of a robotic arm

Abstract

This paper describes the development of a practical stereo vision system which is intended for use in visual tracking of man-made targets from a hovering subsea vehicle. Blob tracking, centroid computation, camera calibration, and stereo vision algorithms are combined to form a robust real-time implementation requiring minimal human intervention. The cameras are calibrated using images of a manipulator-mounted target. Using a less-than-state-of-the-art workstation (Sun IPX) without any special-purpose hardware, targets within a 400/spl times/150/spl times/400 millimeter viewing volume are tracked at rates exceeding 1 Hz (including the acquisition of 3 512/spl times/512 images) with a worst-case accuracy of better than 3 mm for both two- and three-camera configurations.