Multiscale MSE-Minimizing Filters for Gradient-based Motion Estimation

Abstract

Gradient-based algorithm play a vital role in motion estimation. In this paper, a motion estimation algorithm based on gradient methods for low signal-to-noise (SNR) scenarios was presented by using statistical performance of the estimator. The cost function model of mean square error (MSE) was developed based on Cramer-Rao low bound, which the noises were taken into account. The motion estimation MSE was minimized to find the gradient optimal filters. In combination with multiscale pyramid approach, the estimator accuracy of such an algorithm is further improved. Compared to other methods, the estimator performance is performed better for low SNR situations using this optimal filters technique. Experimental simulations show that the estimator bias is less than 0.01 pixels for large motion estimation of low SNR scenarios.