Real-time digital video stabilization for multi-media applications

Amateur digital video footage, shot with a hand-held camera, is characterized by visually annoying, jerky movement between successive frames. In this paper a fast and robust image stabilization technique is presented to remove the effects of unwanted camera motion while not effecting true pans or zooms. Unlike other image stabilization techniques, which are typically used as a pre-processing step to more complicated image processing operations, the proposed technique is specifically aimed at providing "visually jerk-free video" for consumer viewing. Empirical evidence suggested that a simple global translational model for each pair of successive images followed by local integration of cumulative motion is sufficient for this purpose. We use a novel technique to estimate the translational motion model which reduces the total computational burden to less than 1 computation/pixel, while providing robust jerk elimination, thus making real-time operation possible on most PC/UNIX platforms. Tests with a number of video sequences have shown the efficacy of the proposed method.