基于归一化R变换分层模型的人体行为识别  

Hierarchical human action recognition based on normalized R-transform

作　　者：欧阳寒[1] 范勇[1] 高琳[1] 王宗祺[1] 

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621010

出　　处：《计算机工程与设计》2013年第6期2170-2174,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(10776028);四川省科技创新苗子工程基金项目(2011-050);四川省教育厅基金项目(10ZC010);西南科技大学网络工程实验室开放基金项目(11zxwk05)

摘　　要：为保留多特征的鉴别能力和区分能力,同时不增加特征维度,提出了一种基于归一化R变换分层模型的人体行为识别方法。第一层选取描述了运动发生区域的运动能量图像(MEI)作为特征,并依据其归一化R变换曲线对行为进行大类划分;第二层利用细节特征更丰富的关键姿态的星状模型以及宽比和高比特征对各大类进行细分,实现行为的分治识别。对Weizmann和ViHaSi人体行为数据库进行实验,识别率分别为92.47%和96.67%。实验结果表明。In order to reserve the resolving ability of multiple features without increasing the dimension of features, a hierarchical model for human action recognition based on normalized R-transform is proposed. The first layer is to classify the query action into a category according to the normalized R-transform on MEI which suggests where the movement occurs. The second layer is to determine which class the action belongs to. Based on the divide and conquer principle, different classes use different strategies. Star skeleton and the width ratio as well as the height ratio of key poses are chosen to recognize actions in the Weizmann and ViHaSi which are publicly available databases. The proposed method has obtained 92. 47% and 96.67% accuracy respectively. The experiments highlight the advantages of this approach.

关 键 词：R变换 人体行为识别 分层模型 运动能量图像 星状模型 关键姿态 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]