基于Grassmann流形距离的动作识别  

Action recognition based on the Grassmann manifold distance

作　　者：杨江峰[1] 邢传玺[1] 陈跃斌[1] 

机构地区：[1]云南民族大学电气信息工程学院,云南昆明650031

出　　处：《云南民族大学学报（自然科学版）》2016年第6期572-578,584,共8页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：国家自然科学基金(61271288);云南民族大学创新团队项目(35723455)

摘　　要：利用时空兴趣点间的位置及局部特征相似度信息,提出一种局部邻域特征以描述局部数据结构,然后引入核主角度及Grassmann流形距离以度量2个邻域特征距离,通过构造随机流形森林学习数据样本在Grassman流形上的类条件概率密度函数,最后使用多核学习算法实现对随机流形树的剪辑及动作分类.在KTH及UCF-CIL动作数据库的实验证明:所提动作表示方法能有效描述人体运动信息,且该动作识别算法的识别率优于近年提出的典型动作识别系统.Making use of both the position information of spatio - temporal interest points and the similarity between corresponding local feature descriptors, a novel local neighborhood feature is proposed in the paper. For the pur- pose of measuring the distance of the neighborhood features, the kernel principal angles and Grassmann manifold distance are employed. Then, the random manifold forest is constructed to get the class probabilistic density func- tion on the Grassmann manifold. Finally, the multiple kernel learning is utilized for pruning the forest and doing the classification. Experimental results from the KTH and UCF - CIL human action dataset show that the proposed neighborhood feature can effectively depict the human motion information, and the recognition accuracy of the pro- posed system outperforms the classical ones published in recent years.

关 键 词：局部领域特征 Grassmann距离 核主角度 动作识别 动作表示 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]