基于优选特征轨迹的全分辨率视频稳定

该文提出一种基于优选特征轨迹的视频稳定算法。首先，采用改进的Harris角点检测算子提取特征点，通过K-Means聚类算法剔除前景特征点。然后，利用帧间特征点的空间运动一致性减少错误匹配和时间运动相似性实现长时间跟踪，从而获取有效特征轨迹。最后，建立同时包含特征轨迹平滑度与视频质量退化程度的目标函数计算视频序列的几何变换集以平滑特征轨迹获取稳定视频。针对图像扭曲产生的空白区，由当前帧定义区与参考帧的光流作引导来腐蚀，并通过图像拼接填充仍属于空白区的像素。经仿真验证，该文方法稳定的视频，空白区面积仅为Matsushita方法的33%左右，对动态复杂场景和多个大运动前景均具有较高的有效性并可生成内容完整的视频，既提高了视频的视觉效果，又减轻了费时的边界修复任务。

Full-frame Video Stabilization Based on Preferred Feature Trajectories

A novel video stabilization algorithm based on preferred feature trajectories is presented. Firstly, Harris feature points are extracted from frames, and foreground feature points are eliminated via K-Means clustering algorithm. Then, the effective feature trajectories are obtained via spatial motion consistency to reduce false matches and temporal motion similarity for long-time tracking. Finally, an objective function is established, which contains both smoothness of feature trajectories and degradation of video qualities to find a set of transformations to smooth out the feature trajectories and obtain stabilized video. As for the blank areas of image warping, optical flow between the defined area of current frame and the reference frame is used as a guide to erode them, mosaicing based on the reference frame is used to get a full-frame video. The simulation experiments show that the blank area of the stabilized video with the proposed method is only about 33% of that with Matsushita method, it is effective to dynamic complex scenes and multiple large moving objects, and can obtain content complete video, the proposed method can not only improve the visual effect of video, but also reduce the motion inpainting.