Motion Compensated Recursive Filtering of Noisy Displacement Fields*1This material is based on work supported in part by NATO under Grant No. 0103/88 and by Siemens.

In this paper, the recursive filtering of noisy displacement, vector fields (DVFs) is presented. Each component of the original DVF is assumed uncorrelated with the other and is modeled by a 3-D Markov field. The temporal direction of the 3-D AR prediction model is taken along the motion trajectory using motion compensated (MC) prediction based on the DVF in the previous frame. The noisy DVF is estimated from the intensity field using a recursive gradient-based algorithm and is assumed to contain an additive Gaussian noise vector component. A recursive MC 3-D estimator is applied to the noisy DVF with steady-state gain computed a priori. Experimental results demonstrate a good performance of the proposed algorithm with respect to the smoothness of DVF when applied to typical video-conferencing image sequences.