The problem of estimation of visual motion from sequences of images has been considered within a framework consisting of three stages of processing. First the extraction of motion invariants, secondly a local measurement of visual motion, and third integration of local measurements in conjunction with a priori knowledge. We have surveyed a series of attempts to extract motion invariants. Specifically we have illustrate the use of local Fourier phase. The Fourier phase is shown to define the local shape of the signal, thus accurately localizing an event. Different strategies for local measurement of motion under an assumption of translatory motion has been considered. Furthermore we have described methods for quantifying the directional certainty with which we have measured the local displacement. A method based on local estimation of the spatio temporal orientation is generalized to give a continuous description of certainty of the estimated motion. Examples on application of the techniques are given. With respect to integration of local measurements, we have surveyed different techniques within a Bayesian framework. Generalization of prior distributions using 2-D Markov random fields are given. In particular we have investigated the use of smoothness of the second order derivatives, and the use of edge model and prior destributions for the field that favor discontinuities to characterize the motion field. A succesful implementation of a temporal interpolation in a sequence of weather satellite images based on the estimated motion field is shown.
fluid flow; optical flow; visual motion; markov random field