基于对称KL距离的相似性度量方法  

Similarity measure method using symmetric KL divergence

作　　者：姚志均[1] 刘俊涛[1,2] 周瑜[1] 刘文予[1] 

机构地区：[1]华中科技大学电子与信息工程系,湖北武汉430074 [2]军械工程学院计算机工程系,河北石家庄050003

出　　处：《华中科技大学学报（自然科学版）》2011年第11期1-4,38,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(60873127;60903172);河北省自然科学基金资助项目(F2009001435)

摘　　要：提出了一种基于对称KL距离的空间直方图相似性度量方法.将空间直方图中的每个区间的空间分布看作一个带权重的高斯分布,其权重为该区间的概率值,均值和协方差矩阵为该区间内所有像素坐标的均值和协方差矩阵;然后计算2个空间直方图对应区间之间的相似度,即计算2个带权重的高斯分布之间的对称KL距离.理论和实验证明:提出的相似性度量方法的区分能力优于已有度量方法,视频跟踪结果也比已有方法更稳定、更精确.A method to measure the spatiogram similarity based on symmetric Kullback-Leibler（KL） divergence was presented.In this method,the spatial distribution of each bin was regarded as a weighted Gaussian distribution,where the weight was the probability of the corresponding bin,and the mean vector and covariance matrix were computed by all pixels belonging to the corresponding bin.Then,the similarity of corresponding bin of two spatiograms was computed by the symmetric KL divergence with two weighted Gaussian distributions.Theoretical and experimental tests show that the proposed measure is superior to the existing methods in the discriminative power and achieves promising performance in tracking object from single or sequence of images.

关 键 词：相似性度量 空间直方图 KL距离 高斯分布 目标跟踪 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]