特征选择对FHMM性能影响研究  

Influence of feature selection on FHMM

作　　者：陈昌红[1,2] 赵恒[1] 梁继民[1] 焦李成[3] 

机构地区：[1]西安电子科技大学生命科学与技术学院生命科学研究中心,陕西西安710071 [2]南京邮电大学通信与信息工程学院,江苏南京210003 [3]西安电子科技大学智能感知与图像理解教育部重点实验室,陕西西安710071

出　　处：《西安电子科技大学学报》2010年第5期934-940,共7页Journal of Xidian University

基　　金：国家重点基础研究发展计划973资助项目(2006CB705700);教育部创新团队计划资助项目(IRT0645);国家自然科学基金资助项目(60902038;60872154);陕西省自然科学基础研究计划资助项目(SJ08F18)

摘　　要：在利用因子隐马尔可夫模型(Factorial Hidden Markov Model,FHMM)进行分类识别的过程中,特征选择是影响其性能的主要因素.通过研究特征选择对FHMM性能的影响,提出了一种性能分析的方案,得出了选择FHMM特征的准则.将FHMM引入到步态识别中,提取4种步态特征,得到使用不同特征组合的FHMM的实验结果.使用McNemar检验的方法将其与单个特征的识别性能做比较,结合由正则典型相关分析得到的维数不同的特征间的相关性,分析得到以下结论:基于FHMM的识别性能与特征间的相关性并没有必然联系,其性能更多地受到特征间的识别性能差异和单个特征的识别性能的影响.为发挥FHMM的优越性,应选择特征间识别性能差异小和单个特征识别性能好的特征组合,在此基础上特征间相关性越小越好.In the application of the factorial hidden Markov model（FHMM） to classification and identification,feature selection is the major factor impacting the performance of FHMM.An analytical scheme is proposed based on the impact of feature selection on the performance of FHMM and a feature selection criterion for FHMM is obtained.FHMM is introduced into gait recognition.Four gait features are extracted and the experimental results of FHMM with different feature combinations are gained.McNemar＇s test is employed to estimate the performance of FHMM and HMM with a single feature and regularized Canonical Correlation Analysis is used to calculate the relativity between features with different dimensions.Combining the experimental results and the relativities leads to the following conclusion： the recognition performance of FHMM does not have positive connection with the features＇ relativity,which is greatly influenced by the performance diversity between features and the performance of recognition with the single feature.In order to make good use of FHMM＇s advantage,the features should be selected according to the small performance diversity between features and the good performance of the single feature.Based on this,the lower the features＇ relativity,the better.

关 键 词：因子隐马尔可夫模型 特征选择 正则典型相关分析 MCNEMAR检验 步态识别 

分 类 号：TN911.7[电子电信—通信与信息系统]