Probabilistic Approach with Three Hierarchies of Motion Estimation for Video Stabilization

This paper proposes a hierarchical and probabilistic approach for interframe motion estimation to eliminate unwanted camera motions. This approach adopts three hierarchies, local motion estimation, region motion estimation, and global motion estimation. The local motion is described as a probability distribution for randomly selected local patches. The region motion is defined as maximum likelihood solution using the local motions in certain area, and the region motion estimator can reduce the influence of outliers such as moving objects. The global motion is obtained from homography estimation by using the region motions. Through experiments, the effectiveness of the proposed method is shown in the sense of robustness and computation.