In this paper, the algorithm of global motion estimation is investigated, and a new algorithm for fast global motion estimation is introduced. Global motion estimation is the key problem in sprite coding, which is adopted in MPEG-4 compression standard. The proposed new algorithm is much faster than the traditional Levenberg Marquadet algorithm, but the estimated result of the new algorithm is more accurate. The new algorithm is based on nonlinear dense estimation and applies the affine model. In order to accelerate the algorithm, three-level pyramid is utilized in the calculation. It uses those important feature pixels in the calculation of every pyramid level, and combines Gauss-Newton algorithm. In order to guarantee the accuracy of the result, noise is effectively eliminated by histogram-based and block-based methods. Robust statistics is used in the calculation to reduce the influence of noise. The experimental results show that the new algorithm is superior to Levenberg-Marquadet algorithm.