基于图像变形融合时空滤波的视频细微运动增强算法

Subtle Video Motion Magnification by Spatial-temporal Filtering and Image Warping

提出一种基于图像几何变形的视频细微运动增强算法,该算法可在不放大图像噪声的前提下,揭示视频中人眼无法察觉的细微运动信息。其融合了Eulerian和Lagrangian对运动目标描述的形式,以Eulerian视频增强算法(Eulerian Video Magnification,EVM)作为时-空滤波器,通过逐帧检测视频中像素级运动信息建立运动映射图,再根据该运动映射图以Lagrangian的形式计算几何变形网格。最后,使用变形网格对原始输入视频的每一帧图像进行几何变形,放大视频中细微运动目标的运动幅度。实验结果表明,提出的视频运动增强算法能显著降低图像噪声对输出视频画面质量的影响,其视频数据处理管线具备较高的可扩展性,适合于引入先进图像预处理和网格以进一步提高输出视频画面的质量。

An image warping based video motion magnification method was introduced to reveal subtle motion in the input video that are difficult or impossible to see with the naked eye.The main advantage of the presented method is that it amplifies the video motion without increasing the frame noise.The proposed method fuses the approaches of Eulerian and Lagrangian to calculate the motion.The Eulerian video magnification method is used as a spatial-temporal motion analyzer to get pixel-level motion mapping of each frames in the input video.Then each frame of the input video is warped based on this mapping to amplify the input video motion.Experiments show that the presented method is significantly less sensitive to noise and its data processing pipeline is high scalable for introducing advanced image preprocessing filters or mesh post-processing algorithms to further improve the visual quality of the output video.