Translation and rotation invariant video stabilization for real time applications

Undesired vibration in videos is more and more common due to the rise of hand-held or vehicle mounted cameras. Non-indented motion of a video capturing device causes an unpleasant vibrating effect for the consumer. Therefore, video stabilization has been an extremely active field of research in the past years. Motion estimation is the most computationally expensive step of the video stabilization process. Our goal is to circumvent this expensive step in order to achieve real-time performance. We do so by using the precomputed motion vectors from the encoded video streams and thus operate in the compressed domain. These vectors already contain an approximation of the needed motion information. A low-level motion model is used for mitigating complexity, and a low-pass filter performs motion smoothing before the final motion compensation step which is used for correcting the video. In many real-time applications where the video vibration is moderate, the proposed framework can reach online video stabilization at 30 frames per second for high definition video and 60 frames per second for lower resolutions, while retaining satisfactory performance in video correction, comparable to pixel-based equivalent algorithms. Index Terms-Digital video stabilization, compressed domain processing, motion vectors, real-time systems, low-complexity algorithms.

show abstract