Using real-time stereopsis for mobile robot control

Abstract

This paper describes on-going work in using range and motion data generated at video-frame rates as the basis for long-range perception in a mobile robot. A current approach in the artificial intelligence community to achieve timecritical perception for situated reasoning is to use low-level perception for motor reflex-like activity and higher-level but more computationally intense perception for path planning reconnaissance and retrieval activities. Typically inclinometers and a compass or an infra-red beacon system provide stability and orientation maintenance and ultrasonic or infra-red sensors serve as proximity detectors for obstacle avoidance. For distant ranging and area occupancy determination active imaging systems such as laser scanners can be prohibitivtly expensive and heretofore passive systems typically performed more slowly than the cycle time of the control system causing the robot to halt periodically along its way. However a recent stereo system developed by Nishihara known as PRISM (Practical Real-time Imaging Stereo Matcher) matches stereo pairs using a sign-correlation technique that gives range and motion at video frame rates. We are integrating this technique with constant-time control software for distant ranging and object detection at a speed that is comparable with the cycle-times of the low-level sensors. Possibilities for a variety of uses in a leader-follower mobile robot situation are discussed.