Biologically based CAST-mechanism for visual motion analysis

Abstract

In this paper we present a working vision system for estimating size, location and motion of an object by using a set of randomly distributed receptive fields on a retina. The approach used here differs from more conventional ones in which the receptive fields are arranged in a geometric pattern. From the input level, computations are performed in parallel in two different channels: one for purely spatial properties, the other for time-space analysis, and are then used at a subsequent level to yield estimates of the size and center of gravity of an object and the speed and direction of motion. Movement analysis refining is implemented by a lateral interaction (spatial) and memory (temporal) schemes in which direction and speed are used to build a trajectory. The different parameters involved (receptive field size, memory weighting function, number of cells) are tested for different speeds and the results compared, yielding new insights on the functioning of the living retina and suggesting ideas for improving the artificial system.