A biologically plausible neural network for motion detection

Abstract

A neural network model is proposed that detects the actual velocity vector of an object by using local motion signals detected by local motion detectors. First, the computational theory to obtain the actual velocity vector of a rigid object is described. Then a neural network is shown that implements the theory. The neural network model is constructed by two layers: a local velocity vector extraction layer and an integration layer. Many velocity vectors of points on a moving object are detected by local detectors in the local velocity vector extraction layer. These local velocity vectors are integrated in the integration layer in order to obtain the actual velocity vector of the object. The computational processing pathway of the proposed model is similar to the motion processing pathway of visual stimulus in the actual nervous system.