Abstract: This paper presents a robust video stabilization system with feature points tracking based global motion estimation and Kalman filtering based motion compensation. First, global motion is estimated from the local motions of selected feature points, each of which is tracked using feature window matching. Considering the stable relative positions between points’ sets, the distance criterion is proposed to make matching validation, which would delete moving points or mismatched points (i.e. , local features). Then, the global motion vector is decided to sub-pixel accuracy by substituting all validated pairs (i.e., global features) into a motion model and refined by the Levenberg-Marquardt method. Finally, the global motion parameters are accumulated to be Kalman filtered for correction. The experimental result illustrates that the proposed system is effective to stabilizing translational, rotational and zooming jitter and robust to local motions.

Key words: electronic image stabilization, global motion estimation, Kalman filter, feature point