A Resource-EÆcient Approach to Obstacle Avoidance via Optical Flow

This paper describes an active and purposive approach to vision in order to implement a fundamental ability required for robotic navigation: obstacle avoidance. It employs properties of a sparse optical flow field that is produced if the robot moves on a locally planar ground plane. By separating obstacle flow from ground plane flow, obstacles can be detected and measurements of their distance, direction, and width can be extracted. In contrast to other methods, the required resources are reduced to a minimum by avoiding three-dimensional scene or motion reconstruction, resulting in an inexpensive and energyefficient monocular vision system. The results show that static obstacles, protrusions as well as depressions, can be reliably detected. Furthermore, the obtained low computing times allow real-time implementations on the embedded processors available today. Thus, the system is suitable for a broad area of application, even in miniature robotics.