CONTEXTUAL POINT MATCHING FOR VIDEO STABILIZATION

Abstract

We explore the potential of applying a contextual shape matching algorithm to the domain of video stabilization.This method is a natural fit for finding the point correspondences between subsequent frames in a video.By using global contextual information, this method outperforms methods which only consider local features in cases where the shapes involved have high degrees of self-similarity, or change in appearance significantly between frames while maintaining a similar overall shape.Furthermore, this method can also be modified to account for rotationally invariant data and low frame rate videos.Though computationally-intensive, we found it to provide better results than existing methods without significantly increasing computational costs.