Image Motion Analysis Made Simple and Fast, One Component at a Time

Real-time vision can be organised as a sequence of discrete observations, or focal probes, that gather scene information critical to the vision task. Such selective analysis simplifies computations by isolating signal components, and reduces the data load by avoiding unimportant image detail. I outline a sequential approach to the analysis of image motion. The approach uses selection mechanisms analogous to foveation and eye tracking in human vision to isolate motion components one at a time. Each observation estimates motion of a single patch undergoing simple translation. But a sequence of observations can interpret complex patterns containing discontinuities and transparency. Computations are implemented within an image pyramid to provide direct selection of signal components in space, time, resolution, and velocity.