G.7 - Markov random field image motion estimation using mean field theory

Abstract

The estimation of a dense displacement field from image sequences is an ill-posed problem because the data supply may have insufficient information; constraints are needed to obtain a unique solution in such cases. In the problems expressed by Markov random fields, the solution lies in computing the configuration of variables which maximizes the probability distribution. The classical algorithm for this kind of problems is the simulated annealing, a stochastic procedure which converges statistically to the optimal solution, but presents a great computational complexity. To overcome this problem other techniques have been developed. One of these is the iterative conditional mode and the other is mean field approximation. The main advantages of Markov random field modeling of the displacement field are its capacity to regularize the motion vector field, smoothing it while preserving motion discontinuities, and its power to easily integrate information derived from gradient based and feature based motion constraints, obtained by the introduction of other fields in the model.