Video stabilization based on adaptive local subspace of feature point classification

Abstract

Video stabilization removes jitters from shaking videos, which enhances videos quality to achieve stable and comfortable ones. In this paper, we propose a novel method for video stabilization. First, we classify feature points into inliers and outliers based on the global motion estimation to exclude the feature points on moving objects to stabilize camera movements without the interference of outliers. Second, we assemble the trajectory matrix with inlier trajectories across adaptive frames to guarantee sufficient complete trajectories for factorization. Then every frame is smoothed in separate local subspace. This model is more flexible than a global subspace. In addition, to make the inter-frame transition consistent, we exploit homography consistency to alleviate the abrupt transition of inter-frame segments. Experiments demonstrate that our results are comparable with the state-of-the-art methods.