摘要
为了有效识别视频监控领域中的打斗和抢劫等异常交互行为,提出一种基于耦合隐马尔可夫模型(CHMM)的异常交互行为识别方法.首先对人与人之间异常交互行为与正常交互行为的特征差别进行分析,然后提取了包括速度、面积变化率、目标外接矩形长宽比变化率、目标间距、目标运动方向角度差以及方向梯度直方图6类人体目标的运动特征和形态特征,并组成训练数据集,在此基础上使用耦合隐马尔可夫方法构建异常交互行为模型.实验中引入一些典型的行为数据库,如CASIA和CAVIAR数据集,通过和传统的基于隐马尔可夫模型(HMM)的识别方法进行对比,表明CHMM方法更适合于识别少数人的异常交互行为,且识别率更高.
To effectively recognize the abnormal interactions such as fighting and robbing in an intel-ligent video surveillance area,a recognition method for abnormal interactions based on coupled hid-den Markov models （CHMM）is presented.First,the difference between the features of abnormal interactions and that of normal interactions is analyzed.Then the motion features and shape features of the object are extracted to construct the training data set,which are the speed,area change rate, change rate of the bounding rectangle aspect ratio,distance,angle difference of motion direction and the histogram of oriented gradients.Based on them,the CHMM is exploited to construct the abnor-mal interactions model.In the experiments,some classical test cases such as CASIA and CAVIAR are used,and the traditional recognition based on hidden Markov models （HMM）is adopted for comparison.By these experiments,it is proved that the CHMM is more suitable for recognizing the abnormal interactions between fewer people than the HMM,and the recognition rate of the CHMM is higher than that of the HMM.
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2013年第6期1217-1221,共5页
Journal of Southeast University：Natural Science Edition
基金
国家自然科学基金资助项目(60972001)
苏州市科技计划资助项目(SG201076)
