基于时空兴趣点的单人行为及交互行为识别  

Single and interactive human behavior recognition algorithm based on spatio-temporal interest point

作　　者：王世刚[1] 孙爱朦 赵文婷[1] 惠祥龙 

机构地区：[1]吉林大学通信工程学院,长春130012

出　　处：《吉林大学学报（工学版）》2015年第1期304-308,共5页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金广东联合基金重点项目(U0935001);'863'国家高技术研究发展计划项目(2012AA011505);教育部博士学科点专项科研基金项目(20120061110091)

摘　　要：本文方法首先从视频中提取出代表足够运动信息的时空兴趣点,并通过人体前景剪影连通性分析判别时空兴趣点的点集范围。然后对每个视频的兴趣点样本进行高斯混合聚类生成时空单词。最后对时空单词进行训练得到每个行为的高斯混合模型用于人体行为的识别。该方法既可用于单人行为识别也可用于双人行为识别。在行为库上的实验结果证明了该方法有较高的正确率。; First, spatio-temporal interest points containing enough human motion information are detected, and a set of spatio-temporal interest points are selected based on the information of connectivity of human silhouettes. Then, the GMM clustering is performed on the points in the training set and the spatial-temporal words are generated. Finally, these spatial-temporal words are trained to obtain the GMM of each behavior for human behavior recognition. This algorithm can be applied both to single behavior recognition and to interactive behavior recognition. Experiment results on activity database show that this approach has a satisfactory identification rate of human activities.

关 键 词：通信技术 人体行为识别 时空特征点 混合高斯模型 

分 类 号：TN911[电子电信—通信与信息系统]