基于Grassmann流形距离的动作识别
Action recognition based on the Grassmann manifold distance
云南民族大学学报:自然科学版,2016,25(6):572-578

杨江峰 YJF

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 purpose 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 function 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 proposed system outperforms the classical ones published in recent years.

参考



全文: PDF      下载: 186      浏览: 119


counter for myspace
云南民族大学学报(自然科学版) 1991—2016 Copyright
地址:云南省昆明市一二.一大街134号 邮编:650031 全国邮发代号:64-47
电话:0871-65132114 传真:0871-65137493 Email:ynmzxyxb@163.com