Abstract

We discuss the formal relationships among several techniques for computing optical flow fields from sequences of images, and we extend those techniques to compute distributed representations of optical flow. Many computational techniques for deriving flow fields are based on the spatial and temporal derivatives of the intensity signal (i.e., the gradient). By writingspatial and temporal derivatives as quadratic combinations of spatiotemporally oriented filter outputs, traditional gradient techniques can be related to spatiotemporal energy models. Furthermore, these models can be viewed as performing a linear-regression analysis in the spatiotemporal frequency domain, thus fitting a planar surface to the signal spectrum. We discuss extensions of these techniques that compute probability distributions over the space of flow vectors. Such extensions allow the analysis of multiple motions at each spatial position (e.g., transparency) and provide a plausible model for physiological systems.

© 1990 Optical Society of America