A neural network computing observer's heading from optical flow

Abstract

In this paper a neural network model is proposed for the computation of the instantaneous direction of translation of an observer moving relative to a static environment. This direction is given by the focus of expansion (FOE) associated with the radial optical flow pattern arising as a consequence of the translational component of motion. The network is characterized by a feedforward architecture and is trained through the standard supervised backpropagation algorithm. Its input signals are the directions of the optical flow vectors, while its output nodes represent the image coordinates of the FOE associated with the input optical flow field. A number of experiments have been performed both for theoretical optical flow fields and for flow signals actually computed from a TV image sequence by an Hopfield network implementing a gradient-based algorithm. The network is able to recover the FOE position of testing flow fields with a mean error of 0.1 pixels and is resistant to noise. >