A structure-from-motion algorithm for robot vehicle guidance

Reports a new structure-from-motion algorithm designed for autonomous robot vehicle guidance. This algorithm takes input from a binocular image sequence obtained from cameras mounted on the robot, and extracts feature points using a corner detection algorithm developed at Oxford. 3D information is instantiated from feature disparities given by feature-point matching using normalised correlation and disparity gradient techniques. Vehicle motion is estimated from the tracked feature points in the image sequence using a Kalman filter. The algorithm is partly implemented on a heterogeneous parallel machine PARADOX commissioned at Oxford