Global motion estimation for video stabilization

Video sequences captured by handset are generally shaky. A video stabilization scheme developed to stabilize the video sequence and improve its visual quality uses circular block motion search to obtain the local motion vectors of pixels in the current frame. Then a linear system is constructed from these local motion vectors and solved to determine the global motion parameters which are refined by the motion iteration process. Experimental results show that the algorithm can estimate translational motions larger than 30 pixels, rotational motions larger than 30 degrees, and scaling motions between 0.85 and 1.14. The estimation precisions of these three motions are 0.05 pixels, 0.01 degree, and 0.0001, respectively. The algorithm not only deals with larger translational, rotational, and scaling motions than optical flow-based methods, but also achieves higher estimation precisions.