基于张量的KFLD-SIFT与RVM模糊积分融合的人体行为识别方法

一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一一

?国家自然科学基金项目(No.51205185,61308066)二江苏省博士后科研资助计划项目(No.1001027B)资助收稿日期:2013-05-20;修回日期:2013-07-23

作者简介一肖迪(通讯作者),女,1975年生,博士,副教授,主要研究方向为模式识别二数据挖掘二系统优化等.E-mail:xiaodi_12@https://www.wendangku.net/doc/b44764661.html,.南雷光,男,1988年生,硕士研究生,主要研究方向为图像处理二模式识别.基于张量的KFLD-SIFT 与RVM 模糊积分融合的

人体行为识别方法?

肖一迪一一南雷光

(南京工业大学自动化与电气工程学院一南京211816)

摘一要一针对人体行为识别领域中视频序列的大样本及多特征问题,提出一种基于张量的核Fisher 非线性鉴别(KFLD)-尺度不变特征变换(SIFT)与相关向量机(RVM)模糊积分融合的人体行为识别方法.该方法首先通过预处理视频序列得到二值视频,并采用三阶张量表示.然后针对大样本特征提出KFLD-SIFT 局部特征提取算法,对不同初始尺度下的关键点周围的多特征降维,同时提出RVM 模糊积分融合算法进行行为分类.最后应用4种经典评价指标及计算得到的平均识别率对比分析文中方法与其他相关方法的识别效果,数据采用KTH 人体行为数据库中的视频,并采用三重交叉方法验证和测试.实验表明文中方法对多种行为取得较好的识别效果,平均识别率比其他主流方法至少提高2.3%.

关键词一张量,核Fisher 非线性鉴别(KFLD),尺度不变特征变换(SIFT),相关向量机(RVM),模糊积分融合

中图法分类号一TP 391

KFLD-SIFT with RVM Fuzzy Integral Fusion Recognition of Human Action Based on Tensor

XIAO Di,NAN Lei-Guang

(College of Automation and Electrical Engineering ,Nanjing Tech University ,Nanjing 211816)ABSTRACT

Due to the large sample and multiple characteristics of video sequence in the field of human action recognition,a method of kernel Fisher nonlinear discriminant (KFLD)-scale invariant feature transform

(SIFT)and relevance vector machine (RVM)fuzzy integral fusion recognition based on tensor is proposed.Firstly,video sequence is pre-processed into binary video sequence,and then it is described as third-order tensor.Furthermore,as for large sample characteristics,a local feature extraction method of KFLD-SIFT is proposed to reduce the dimension around the key points under different initial scales.Meanwhile,RVM fuzzy integral fusion algorithm for behavior classification is presented.Finally,the proposed method and other relevant methods are compared through four kinds of evolution indexes and average recognition rates.The video sequence of KTH human action database and triple-cross verification method are used to test the recognition methods.Experimental results show that the proposed method achieves good recognition effect,and its average recognition rate rises by at least 2.3%compared to other 第27卷一第8期一一一一一一一一一一一一一模式识别与人工智能一一一一一一一一一一一一一一一Vol.27一No.8一2014年8月一一一一一一一一一一一一一一一一PR &AI一一一一一一一一一一一一一一一一一一Aug.一2014