Spatially-adaptive regularized pel-recursive motion estimation based on cross-validation

Pel-recursive motion estimation is a well established method for finding the displacement vector-field (DVF) between adjacent image frames. The motion due to the optical flow in image sequences is estimated recursively. In this paper, we improve the Wiener-based pel-recursive algorithm by using spatially-adaptive regularization. The outcome of the regularized solution is dependent upon the value of the regularization parameter. This work employs a data-driven approach called generalized cross-validation (GCV) to determine the optimal value of the regularization parameter for each pixel. Experimental results are presented and the linear minimum mean-squared (LMMSE) solution (also known as the Wiener solution) is compared to the proposed approach.