Abstract

A new parallel and fast algorithm for computing the optical flow and its implementation on the Connection Machine has been described.^1,2 This algorithm is based on a regularization technique that exploits a simple assumption, that the optical flow is locally uniform. It can be easily translated into the following description. Consider a network of elementary motion detectors holding the results of multiplying (or logical “anding”) image features (intensity or edges) for different displacements. Each detector collects a vote indicating support that a patch of surface exists at a certain displacement in the second image. The final step is to choose the velocity v(x,y) out of a finite set of allowed values that has maximum vote (nonmaximum-suppression or winner-take-all scheme). The corresponding v(x,y) is taken as the velocity of the point (x,y). While the true velocity vectors for a rotating barber pole are strictly horizontal, our algorithm computes a vertical velocity field which is consistent with the well-known barber pole illusion. If a moving sine-wave grating is superimposed on a pair of alternating and uncorrelated random-dot patterns, most of the dots in the display appear to move as a uniform sheet in synchrony with the sine-wave grating (motion capture³). The voting scheme described computes a similar result without using any trick as suggested by others.³

© 1987 Optical Society of America