A neural network model for the velocity vector of an object and its consistency with psychological phenomena

Abstract

A biologically plausible neural network model is proposed that can detect the actual velocity vector of an object by using local motion signals detected by many local motion detectors. First, the computational theory to find the actual velocity vector of a rigid object correctly 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. Then these local velocity vectors are integrated in integration layer in order to obtain the actual velocity vector of the object. The computational processing pathway of the proposed model is well fit to the motion processing path of visual stimulus in the actual nervous system (from LGN to MT via VI). We also show the model can explain motion perceptual phenomena in the case where a moving stimulus is observed through an aperture.