Abstract

When humans (or robots) move through a scene, the scene can be represented as an optic flow (optical flow) field that contains vectors representing all of the movement within the scene projected onto a two-dimensional sensor. The movement may be caused by the agent moving through the scene, or by independent objects moving within the scene. A simultaneous sample of these resulting vectors contain a good deal of information in addition to object movement, e.g. relative depth; shape of objects; or signatures of specific biological motion. A general model of motion estimation of this field would therefore be valuable. Previous reported attempts at computing motion estimation models have been dominated by the machine vision community. However these attempts are not specifically concerned with biologically plausibility. Here, the author presents a model of motion estimation that computes motion based on filtering the moving image into sinusoidal responses varying in spatial frequency and orientation similar to the early visual responses found in human vision. Unlike similar spatio-temporal energy models ``motion energy’’ is not computed. The model is mathematically explicit and simulated in MATLAB . It has been tested using over 7000 synthetic moving images with known veridical velocity (ground truth). These images range from sparse translating patterns containing 1 to 10000 random pixels, to dense narrow band sinusoidal patterns. Simulation results show that the model correctly estimates motion trajectories between 90\ – 100\ angular direction error, and displacement error (within +/- 1 pixel). The results remain robust at different contrasts. In addition, explanations for a number of psychophysical and physiological results that emerge from the model are presented.