Our paper, “Real-Time Low-Complexity Digital Video Stabilization in the Compressed Domain” has been accepted for publication. More specifically, it will be published in the proceedings of the International Conference on Consumer Electronics in Berlin, Germany (ICCE Berlin 2018).
Due to undesired vibration in videos, video stabilization is an extremely active field of research. Motion estimation is the most computationally expensive step of the video stabilization process. Our goal was to circumvent this expensive step in order to achieve real-time performance. We did so by only using the already available motion vectors from the encoded video streams. Thus, we operate exclusively in the compressed domain.
These motion vectors already contain an approximation of the motion present in a video scene. A low-level motion model is used for mitigating complexity, and a low-pass filter performs motion smoothing. Finally, a motion compensation step corrects the video accordingly. In many real-time applications where the video vibration is moderate, the proposed framework can reach online video stabilization. Results showed video correction at more than 30 frames per second for HD video and more than 60 frames per second for SD video. Subjectively, the algorithm delivers acceptable results similar to state-of-the-art pixel-based approaches.