Abstract: There are inevitable shakes in videos taken by handheld devices and UAVs, which reduce the effect of video viewing and post-processing. In the video stabilization technology, similar matrix transformation is often used to perform inter-frame correction to have a real time processing. But the compensation effect for three-dimensional motion is poor to use similar matrix transformation. On the other hand, the homography matrix transformation that can represent three-dimensional motion, can not meet the real time requirements. To solve this problem, a video stabilization technique based on homography revised P-KLT is proposed. This algorithm is based on the optical flow tracking theory and realizes real-time parameter tracking calculation for the homography matrix transformation. A distortion correction algorithm based on vertex filtering is proposed to meet the requirements of 3D motion video real-time image stabilization. The experimental results show that the video images processed by this algorithm are stable and natural, especially for videos with motion blur. 

Key words: Video stabilization, optical flow tracking, Kalman filter, distortion correction

摘要： 手持设备和无人机拍摄的视频存在不可避免的抖动，降低视频观看及后期处理的效果。视频稳像技术中考虑到实时性，常用相似矩阵变换来进行帧间校正，对于三维运动的补偿效果较差，但是直接采用可以表示三维运动关系的单应性矩阵变换，实时性无法满足要求。对此，提出了一种基于单应性校正P-KLT的视频稳像算法，此算法是以光流跟踪理论为基础，对单应性矩阵变换实现实时参数跟踪计算，针对单应性变换存在图像局部畸变的问题，提出了基于顶点滤波的畸变校正算法，满足了三维运动视频实时稳像要求。实验结果表明，经过本算法处理后的视频画面稳定自然，尤其是对存在运动模糊的视频也有较好的处理结果。

关键词: 视频稳像, 光流跟踪, 卡尔曼滤波, 畸变校正