| Video Stabilization is one of those important video processing techniques.It has been widely applied in target tracking,surveillance system and personal video processing software,etc.With the popularity of video equipments,Video Satbilization has become a hot research topic.The goal of video stabilization is to remove the unwanted camera vibration in a video sequence for improving the video's quality and using it as the input of other applications.It includes three modules:motion estimation,motion compensation and video completion.In this paper,we propose a new video stabilization technology based on SIFT features.Since they are proved to be invariant to image scale and rotation,and robust to affine distortion and change of illumination,motion estimation based on them is more robust and accurate.Then,the unwanted vibrations are separated from the intentional camera motion with the combination of Gaussian kernel filtering method and parabolic fitting method,which is demonstrated to avoid over-smoothing effectively and reduce the complexity of video completion because of keeping more video information.Finally,we adopt the mosaicing method with Dynamic Programming to reconstruct the undefined areas, resulting from motion compensation.Since the resource image patchs used to repair the video are from a 'single' candidate frame,the blur caused by the motion object or blending is not yielded.The proposed methods have been confirmed to be effective over a widely variety of videos.They could effectively remove camera's shaky motion and reconstruct videos with high quality which satisfy the need of watching and related applications. |
