摘要： 针对目前大部分视频稳像算法实时性及适用性较差的问题,基于AKAZE特征,提出一种复杂的抖动数字视频稳像算法。通过AKAZE特征提取视频图像的特征点,采用快速近似最邻近库算法匹配视频邻帧间的特征点。利用特征点筛选机制剔除误匹配的特征点,以提高运动估计的准确性,并采用随机抽样一致性算法迭代求解视频图像间的运动参数。为得到去抖动的视频运动参数,使用高斯滤波器将运动参数进行滤波,对视频序列进行运动补偿得到稳像视频。实验结果表明,该算法在特征点数较多的情况下,其运动估计速度接近于加速健壮特征(SURF)算法的3倍,比基于SURF特征点的运动估计算法具有更强的实时性及健壮性,比三维内容保持变换算法、子空间视频稳像算法效果更稳定。

关键词: 加速KAZE特征, 视频稳像, 复杂抖动, 快速近似最邻近库, 特征匹配, 运动估计, 高斯运动滤波

Abstract: To solve the problem of weak real-time processing ability and applicability of most current video stabilization algorithms,a digital video image stabilization algorithm based on AKAZE features for complicated shakiness is proposed.Feature points of video images are extracted by AKAZE features,and feature points between video adjacent frames are matched by Fast Library for Approximate Nearest Neighbor(FLANN) algorithm.In order to improve the accuracy of motion estimation,a feature point screening mechanism is proposed to eliminate mismatched points.And random sample consensus algorithm is used to solve the motion parameters between video images.To obtain the stabilized video motion parameters,a Gaussian filter is applied to smooth out the motion parameters,and the stabilized video can be obtained from motion compensation process for video sequences.Experimental results show that in the case of many feature points,the motion estimation speed of the proposed algorithm is approximately triple as fast as Speed Up Robust Feature(SURF) algorithm.It has stronger robustness and higher real-time performance than the motion estimation algorithm with SURF points,and can achieve more reasonable stabilization results than the algorithms like Content Preserving Warps for 3D(CPW-3D) and Subspace Video Stabilization(SVS).

Key words: Accelerated KAZE(AKAZE) feature, video image stabilization, complicated shakiness, Fast Library for Approximate Nearest Neighbor(FLANN), feature matching, motion estimation, Gaussian motion filtering

中图分类号: 

• TP301.6