A biologically inspired neural network model for 3-D motion detection

Abstract

Proposes a biologically inspired neural network model that computes three-dimensional motion based on monocular cues. In the approach, instead of computing a 2D optical flow field and extracting motion information from it, the 3D motion is computed directly. Motion in the z-axis is detected and localized by a network of dilation-sensitive neurons, and the z-motion is parsed with an x-y component. The correspondence problem is resolved by the inherent neuronal characteristic of temporal and spatial locality. Temporal locality refers to the smooth decay of neuronal activity within a small time interval after a stimulus is removed. This property provides a temporal signal bridging consecutive image frames. Spatial locality refers to the localized receptive field of a neuron. This property ensures that the correspondence between consecutive frames is restricted to a small neighborhood. Together, they provide the temporal and spatial continuity in the sequence of time-varying frames as the basis for computing 3D motion. >