Robust multiscale algorithms for gradient-based motion estimation

Gradient-based techniques represent a very popular class of approaches to estimate motions. A robust multiscale algorithm of hierarchical estimation for gradient-based motion estimation is proposed in this article using a combination of robust statistical method and multiscale technique. In such a multiscale approach of hierarchical estimation, motion at each level of the pyramid is estimated using different gradient filters. The iterative multiscale estimation begins by using five-tap central filter, and it is switched to nine-tap Timoner filter after a few iterations. In addition, robust M-estimators are applied at each level of the pyramid to overcome the problem of the outliers caused by illumination variations and motion discontinuities in motion estimation. Experimental simulations show that the new algorithm not only provides an improvement in estimator accuracy, but also achieves computational speedups. (c) 2008 Wiley Periodicals, Inc.