Ohta : Uncertainty Models of the Gradient Constraint 959

The uncertainty involved in the gradient constraint for optical ow detection is often modeled as constant Gaussian noise added to the time-derivative of the image intensity. In this paper, we examine this modeling closely and investigate the error behavior by doing simulation. Our result indicates that the error depends on both the spatial derivatives and the image motion. We propose alternative uncertainty models based on our experiment. It is shown that the optical ow computation algorithms based on them can detect more accurate optical ow than the conventional least-squares method. key words: optical ow, gradient constraint, noise model, probability model, least squares, maximum likelihood criterion.