A Neural Network for Motion Processing

Abstract A locally connected artificial neural network beised on physiological and anatomical findings in the visual system is presented for motion processing. A set of velocity selective binary neurons is used for each point in the image. Motion processing is carried out by neuron evaluation using a parallel updating scheme. A deterministic decision rule is used to ensure quick convergence of network to probably a local minimum. In view of high parallelism and local connectivity, this network is suitable for VLSI implementation. Both batch and recursive algorithms based on this network are presented for computing optical flow and recovering depth from a sequence of monocular images. The batch algorithm simultaneously integrates information from all images by embedding them into the bias inputs of the network, while the recursive algorithm uses a recursive least squares (RLS) method to update the bias inputs of the network. Detection rules are also used to find the ocdiscontinueluding elements. Based on information on the detected occluding elements, the network can automatically locate motion and depth discontinuities. Both these algorithms need to compute optical flow at most twice and depth only once. Hence, less computations are needed and the recursive algorithm is amenable for real time applications.