提高光流估计性能的渐进性高斯多维预滤波方法的研究  

The Improving Characteristics of the Gradual Gaussian Multidimensional Pre-filter for Optical Flow Estimation

作　　者：付昀[1] 徐维朴[1] 

机构地区：[1]西安交通大学人工智能与机器人研究所,西安710049

出　　处：《中国工程科学》2004年第12期56-61,共6页Strategic Study of CAE

基　　金：国家自然科学基金资助项目 (60 2 0 5 0 0 1) ;国家自然科学创新研究群体科学基金资助项目 (60 0 2 43 0 1)

摘　　要：基于光流计算方法统一框架理论 ,研究了一种利用高斯多维滤波器的渐进性和时空性提高光流估计性能的有效方法。在保持现有光流计算方法的前提下 ,通过调节时间维和空间维的方差参数 ,改变时空预滤波和光滑效果 ,突出时间混叠和光流主信息 ,从而提高重构视频序列的信噪比。试验中以标准的FlowerGarden和Football序列的前 5 0帧作为参考图像序列 ,以LK算法为参考光流算法。结果显示 ,滤波窗口为 5× 5时的最佳时间方差参数为 0 4 ,最佳空间方差参数为 [1 6 ,2 0 ];加入高斯多维预滤波前后利用光流场重构图像的平均峰值信噪比PSNR提高 2 5 72dB ,提高幅度为 13 6 %。Based on the unified estimation-theoretic framework, an effective method of using the gradual Gaussian multidimensional pre-filter to improve the optical flow estimation is presented. The pre-filtering and smoothing effect, which attenuate the temporal aliasing and the interesting signal structure of the optical flow field, are altered with adjusting the spatiotemporal standard deviation parameters. The first 50 frames of the standard Flower Garden and Football video sequence are tested as the reference image sequences, and the LK algorithm as the reference optical flow computing method. Experimental results in objective evaluation show that the optimum temporal standard deviation parameter is 0.4,the optimum spatial standard deviation parameter is in a range of 1.6～2.0 under the condition that the pre-filtering window size is 5×5 pixels. After pre-filtering the image sequence by the Gaussian multidimensional filter, the average PSNR of the reconstructed frames enhance 2.572 dB, higher than that using the standard optical flow computing method by nearly 13.6 %.

关 键 词：光流计算 高斯多维滤波 峰值信噪比 运动估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN713.1[自动化与计算机技术—计算机科学与技术]