Foreground stabilization of image sequences

Hand jitters result in unintentional fluctuation of image sequences taken by hand-held video cameras. Stabilization of the foreground object of interest in pictures is essential for good visual quality. In this paper, foreground stabilization algorithms of image sequences are proposed and evaluated. After performing the block-based motion estimation, grouping techniques are used to identify the foreground motion. The tested grouping techniques include the iterative centroid of foreground, the k-means clustering, and the LMedS (least median of squares) algorithms. An adaptive IIR filtering technique is subsequently applied for motion correction. Compared to conventional stabilization techniques without foreground separation, the proposed methods substantially improve the visual quality of image sequences both subjectively and objectively.