Fast and Accurate Robot Vision for Vision Based Motion

This paper describes the vision module from the soccer playing robots of the Dutch Team. Fast vision is necessary to get a close coupling with the motion software in order to allow fast turning and dribbling with the ball without loosing it. Accurate vision is necessary for the determination of the robot's position in the field and the accurate estimation of the ball position. Both fast and accurate are necessary for the goalkeeper, but also when one robot passes the ball to another. While the Dutch team has pneumatic kicking devices that allows catching a ball smoothly, fast an accurate vision is mandatory. We use lens undistortion, a new color segmentation scheme and a shape classification scheme based on linear and circular Hough transforms in regions of Interest. We use a severe calibration procedure to get very good distance and angle measurements of the known objects in the field of view of the robot. For the keeper robot we use a Linear Processor Array in SIMD mode, that is able to execute the entire robust vision algorithm within 30ms. However the same software was programmed for the other robots with a WinTV framegrabber on the on-board Pentium of the robot. With optimizing for speed we also remained within 25ms, however, omitting the circular Hough transform for the ball and processing in a separate thread the Linear Hough transforms for self-localization on lower rate of about 50msec. The angular errors at 0 °, 20 ° and 30° heading are about 0.6 °, 0.0° and 0.4°. The distance error at 0 ° heading is 2cm at 1.5m and 3 cm at 2m.