Study On The Methods Of Video Stabilization Based On Block Matching And Local Subspace

With the development of electronic technology, all kinds of camera equipment are widely used. Due to outside interference or amateur-level capture, different kinds of undesired motion occur in the camera trajectory causing jitter in the captured video. The shaking in the video will badly affect the convenience of watching and cause further difficulties in afterwards video processing and analysis such as face detection and object tracking. With the rapid development of digital video stabilization, this technology is widely used in military and consumer aspects.In this paper, current video stabilization techniques are summarized based on a lot of domestic and foreign literatures in this field, and on the basis of existing stabilization algorithm, one block matching based method and one local subspace based method are descripted in detail. The main work is as followed:(1) The existing domestic and international image stabilization algorithm are studied and summarized in this paper. Firstly, the basic principle and mathematical model of image transformation are introduced briefly. Secondly, the existing classic 2D video stabilization algorithm is described in detail, including motion estimation and motion compensation. Finally, the basic principles and algorithms of 3D stabilization are presented and specifically describing Content-Preserving Warps (CPW).(2) A real-time video stabilization method is proposed based on hierarchical block matching for motion estimation and improved Kalman filter for motion compensation. In this method, key blocks having obvious characteristics are picked firstly in the image and using hierarchical method for block matching to ensure the operation speed and accuracy. Additionally, M-estimator Sample Consensus (MSAC) is used to remove outliners in the matched blocks and motion vectors are computed using least squares algorithm. In the compensation section, an improved Kalman filter is applied to achieve smoothness of the output motion path and track the original curve properly at the same time. This method works in real time and can achieve good video stabilization result.(3) Another video stabilization method is proposed based on AKAZE feature extraction and local subspace algorithm. In this method, AKAZE feature is firstly extracted to calculate the trajectory of the video ensuring the accuracy of trajectory. Secondly, for the phenomenon of image distortion and incoherent in the output video caused by smoothing all the trajectories directly, local subspace method is applied to generate eigen-trajectory by decomposing the original trajectory matrix then filtering using Gaussian kernel. Finally, CPW algorithm is used for image reconstruction and output video generated afterwards. This method can achieve high quality stabilization result.The two different video stabilization method are tested by lots of simulation experiments for videos captured in different scenes and compared with other state-of-the-art algorithms, real-time, robustness and precision are all tested and the experimental results of our methods show that the proposed methods both perform well.