Abstract

Autonomous navigation based on optical flow can be significantly simplified by coordinate transformation or retinotopic mapping prior to motion detection. For detection and avoidance of obstacles in the plane, the appropriate map is the inverse perspective. This mapping transforms the egomotion field into a parallel vector field where the velocity vector of points in the ground plane has equal length. Obstacles will stand out by the length of their flow vectors, the deviation scales with their height above the plane.¹ We tested this mapping paradigm with moving synthetic images. Motion detection was performed in the transformed images using a parallel motion algorithm which is not subject to the aperture problem.² This area-based voting is computationally expensive when performed in a 2-D neighborhood (4 s for 256² images on the Connection Machine) and less suited for natural and artificial real-time vision systems. The inverse perspective mapping speeds up the voting step considerably by restricting the 2-D search to a 1-D search along the egomotion vector. The map thus implements the ecological constraint that movement occurs in a plane while all derivations from that assumption are considered obstacles.

© 1988 Optical Society of America