Abstract

Recently it was shown that linear spatiotemporal filters can be tuned to relatively narrow ranges of normal velocity [1, 7, 16]. Despite this tuning, a subsequent stage is still necessary to compute velocity estimates because 1) the filters are linear and therefore do not yield quantitative estimates of speed and orientation, 2) more accuracy is required than that given by the filter tunings alone, and 3) if a unique 2-d flow field is desired, then the aperture problem must be overcome as the filters are tuned only to normal velocities. Adelson and Bergen [2] suggested the computation of velocity from the ratios of energy in the outputs of differently tuned filters. Heeger [9] based the computation of 2-d flow on a least-squares fit of local energy to a model of 2-d translation.

© 1989 Optical Society of America