A novel motion parameters estimation approach for sub-pixel accuracy is presented which is based on Powell optimization and the pyramid decomposition method of the images. The non-linear least square formula for the difference of reference and sensed images is seen as the object function for Powell algorithm in which motion parameters are used as variations. Based on multi-resolution decomposition from fine to coarse representation of images, sub-pixel parameters can be accurately estimated. Numerical experiment results indicate that the proposed algorithm can be used to perform accuracy and robust sub-pixel parameters estimation. The drawback of this algorithm is the longer waiting time compared with that of other algorithms.