基于多尺度特征的双层隐马尔可夫模型及其在行为识别中的应用  

Multi-scale feature based double-layer HMM and its application in behavior recognition

作　　者：梅雪[1] 胡石[1] 许松松[1] 张继法[1] 

机构地区：[1]南京工业大学自动化与电气工程学院,江苏南京211816

出　　处：《智能系统学报》2012年第6期512-517,共6页CAAI Transactions on Intelligent Systems

基　　金：江苏省高校自然科学基金资助项目(09KJB510002);江苏省博士后科研资助计划资助项目(1001027B);南京工业大学青年学科基金资助项目(39710006)

摘　　要：借鉴人类视觉感知所具有的多尺度、多分辨性的特性,针对智能视频监控系统的人体运动行为识别,提出了一种基于多尺度特征的双层隐马尔可夫模型.根据人体行为关键姿态数确定HMM的状态数目,发掘人体运动行为隐藏的多尺度结构间的关系,将运动轨迹和人体姿态边缘小波矩2个不同尺度特征应用于2层HMM,提供更为丰富的行为尺度间的相关信息.分别用Weizmann人体行为数据库和自行拍摄的室内视频,对人体运动行为识别进行仿真实验,结果表明,五状态HMM模型更符合人体运动行为特点,基于多尺度特征的五状态双层隐马尔可夫模型具有较高的识别率.Learning from multi-scale and multi-distinguish attributes of human beings＇ visual perception and aiming at human movement behavior recognition in intelligent video surveillance system,a double-layer hidden markov model（DL-HMM） is developed based on multi-scale behavior features.Considering the human behavior characteristics,the number of HMM states is according to the number of key gestures selected.Discovering the relationship between the multi-scale structures hidden in the human movement behavior,two different scale features-human motion trajectory and wavelet moment of human gesture＇s edge,are applied respectively in two layers of DL-HMM,so as to provide more scale information about behavior.Experiments,using Israel Weizmann human behavior database and human actions indoor recorded by ourselves,show the five-state HMM more accords with the human motion behavior characteristics,and the five-state DL-HMM based on multi-scale feature has a higher recognition rate compared with traditional methods using one layer HMM.

关 键 词：双层隐马尔可夫模型 行为识别 多尺度特征 智能视频监控 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]