结合MACH滤波最大池化及多类SVM的行为识别  

Activity recognition combined with MACH filter,max-pooling and multi-class SVM

作　　者：何俊林 赵晓亮 孙连海 甘胜江 

机构地区：[1]成都师范学院计算机科学学院,四川成都611130 [2]中原工学院信息商务学院信息技术系,河南郑州450000

出　　处：《计算机工程与设计》2017年第12期3431-3435,共5页Computer Engineering and Design

基　　金：中国博士后科学基金面上基金项目(2014M560730);四川省科技厅应用基础基金项目(2015JY0071);成都师范学院高层次引进人才专项科研基金项目(YJRC2014-9);成都师范学院自然科学类培育基金项目(CS14ZD02)

摘　　要：提出一种人体行为识别方法。构建MACH滤波器组,对视频片段的三维时空体进行滤波,得到时空相关体;采用三层最大池化方法提取时空相关体的特征向量,采用高斯隶属函数对池化特征向量进行扩展;构建多类SVM分类器并进行特征分类,识别行为类别。在ADL和UCF Sports两个国际上通用的人体行为数据集上进行人体行为识别实验,实验结果表明,该方法的识别率高于现有的人体行为识别方法,对不同人体行为的区分能力更强。A human activity recognition method was proposed.A set of MACH filters was built to filter 3 D space-time volume of video clips,for obtaining a space-time correlation volume.Three-layer max-pooling method was used to extract feature vector of the space-time correlation volume,and Gauss membership function was used to extend the pooling feature vector.Multi-class SVM classifier was built to classify features and recognize activities.Experimental results on two internationally common datasets of human activity including ADL and UCF Sports show that,the proposed method has higher recognition rate than the existing methods for human activity recognition,and it has strong ability to distinguish different human activities.

关 键 词：行为识别 MACH滤波 最大池化 多类支持向量机 高斯隶属函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]