DETECTING COLLISION FROM GRAY-LEVEL EXPANSION BY A NEURAL-NETWORK

A neural network is used for detecting an imminent collision from the gray-level map generated by a textured surface. The network maximizes the output entropy in learning the probability density functions of th e data, corresponding to ''safe'' and ''dangerous'' categories. First- order temporal and spatial derivatives of the optical flow, which are related to the time to collision through the local divergence, are use d as inputs to the network. Detection is based on the relative sizes o f the two densities corresponding ro a given input. In contrast to a p revious design, the one presented here does not require thresholding t he input data, and the network size is equal to the input dimension.