A video stabilization algorithm is presented, which has the advantages of low computation cost, high accuracy and robustness. It consists of two parts: (1) global motion estimation is performed based on two-dimensional motion model of feature points set. The feature window with each selected feature point as the center is matched and then the corresponding points are obtained. The feature set has a stable structure, and the relative positions of all points are stable. So distance invariant criterion is proposed to check the feature matching, keep global coherence of all local motions and produce the global motion vector of the feature points; (2) a self-adaptive average filtering is applied to exclude the influence of the camera dithering. The average filter can smooth out the high frequency dithering effectively and adjust the window length automatically according to dithering frequency to prevent overstabilization or understabilization. Experimental results show that the algorithm can effectively reduce the influence of unwanted camera rotation and translation dithering.