面向电子稳像应用的模糊聚类运动估计算法研究

Study on motion estimation based on fuzzy classification in electronic image stabilization application

电子稳像(electronic image stabilization,EIS)技术主要解决视频抖动问题,是视频增强的重要技术。实现技术内容包括运动估计、运动滤波和运动补偿,而目前运动估计中分割出前景的运动是难点。提出了一种模糊聚类运动估计方法,能够分割出前景和背景运动分量,提取出全局最优运动矢量。通过理论推导和实测数据分析,结果表明,所提算法能够有效分割前景运动,保证所检测的特征点来自背景,减少了前景运动对于稳像算法的干扰,稳像效果更佳。与流行的RANSAC(random sample consensus)相比,此算法提取的特征点分布在背景上的概率由原来的最低百分比21%提高到最低百分比93%,六段测试视频的ITF(interframe transformation fidelity)分别提高了15.42%、17.41%、13.15%、12.09%、12.42%、8.67%,运动路径也得到了明显优化。稳像后,视频质量明显提高。

EIS mainly solve the problem of video jittering and video enhancement. It includes motion estimation, motion compensation and motion filtering. Especially, now segmenting the foreground motion is difficult. This paper presented a motion estimation method based on fuzzy clustering. It could segment the foreground and background motion vector, extracted the globally optical motion vector. By theoretical analysis and data analysis, the experimental results show that, this algorithm can segment the foreground motion effectively, ensure that the detected feature points are from background, to reduce the interference from foreground. Compared with RANSAC, the probability of feature points extracted distribution on the background is improved from minimum 21% computed by RANSAC to minimum 93% by this algorithm, the ITF of six test videos are increased by 15.42%, 17.41%, 13.15 %, 12.09%, 12.42%, 8.67 %. Motion path has also been significantly optimized. After stabilization, the video quality has improved greatly.