To realize fast and robust digital image stabilization, a new optimum motion estimation method is proposed. Firstly, an approach for block judgment is presented. Before calculating motion vectors, gradient information is analyzed to judge the blocks, and only useful reference blocks that are indispensable for accurate motion estimation are selected with its reliability and consistency on motion estimation, by which the number of blocks for subsequent calculation is reduced. Secondly, in the block matching, an improved Sequence Similarity Detection Algorithm (SSDA) is used to reduce computing cost. Finally, the LMedS method is used to eliminate wrong motion vectors and the parameters of whole motion model are obtained by least squares algorithm. Experimental results show that the method gets 1/4 pixel accuracy, and one process time is less than 7 ms.