Video Stabilization Based On Optimization

With the popularization of hand-held devices, like digital cameras, smart phones, and the development of wearable devices such as Google Glass, it is much more convenient for people to obtain videos today. However, these videos are usually shaky, requiring to be stabilized in order to enhance the visual quality. This process is called video stabilization. Besides, video stabilization can be also treated as the pre-process of other video-related tasks, such as object tracking, object detection and video compression, to improve the precision and robustness. This paper focuses on video stabilization algorithms based on optimization, and performance assessment on some video stabilization algorithms. The research work includes:(1) We proposed a video stabilization algorithm with l1-l2 optimization. With both the L1 norm and L2 norm constraints on camera path, we can not only obtain smooth video, but also ensure that the stabilized video is as close to the original video as possible, thus retaining more information of the shaky video. Besides, users can adjust the degree of stabilization according to their needs.(2) We proposed an online video stabilization scheme. For long videos, it can greatly reduce the time of stabilization by processing the video segment by segment with overlapped frames. And for a real-time recording video, the scheme can almost achieve online stabilization by buffering several frames with several seconds delay.(3) We proposed a shaking video synthesis algorithm to produce different kinds of jitters. The synthetic videos are very similar to the real shaky ones, and can be stabilized by different algorithms. Then we can compare the stabilized videos with ground-truth ones, in order to evaluate the performance of different video stabilization algorithms.(4) We proposed a video shakiness detection method. Two kinds of shakiness features are extracted from the input video, then we train a SVM classifier with many videos, with which we can detect the degree of shakiness of input video. The method can be used to decide whether to perform stabilization or not, and if yes, which algorithm is suitable. Besides, it can be also applied in no-reference video stabilization performance assessment by detecting the shakiness of stabilized videos.