Abstract

We have devised simple, local motion detection schemes providing information which can be integrated with other vision modalities to segment images. A simple scheme proposed for motion detection in the rabbit retina,¹ based on a logical comparison between image elements separated in time and space, can detect locally the direction of motion but not its velocity. To avoid aliasing (in time and space), the input is filtered by the Laplacian of a Gaussian and then thresholded at zero. The simple rabbit scheme is enhanced to suppress false motion detection at isolated intensity changes (e.g., changes in illumination and receptor noise). To compute velocity, this scheme has been extended to multiple scales, where each scale is selected by changing the size of the Gaussian mask. Also, the criteria have been narrowed to detect motion only at the selected scale. The fly motion detector (Reichardt-Correlator)² uses image grey values, multiplying neighboring receptor signals, after asymmetric temporal filtering. The rabbit implementation uses simple local binary operations to compute motion, permitting easy parallelization. On the Connection Machine, a mesh-connected parallel computer with nonlocal communication, motion detection on natural images can be performed in real time.

© 1986 Optical Society of America