基于姿态转换网络的行为识别  

Action recognition based on posture transition network

作　　者：裴利沈[1] 赵雪专[2] 李涛[1] 向涛[1] 包姣[1] 

机构地区：[1]电子科技大学计算机科学与工程学院,成都611731 [2]中国科学院成都计算机应用研究所,成都610041

出　　处：《计算机应用研究》2015年第8期2505-2509,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61375038);河南省科技厅科技攻关项目(142102210010);河南省教育厅重大专项项目(14A520028;14A520052)

摘　　要：为了有效地表征行为,提出了一种基于姿态转换网络的行为识别算法。首先对人体进行自动定位,并对人体区域进行形状与运动特征提取;然后对特征进行层次聚类,构建姿态二叉树,并将运动序列表示为姿态序列后,将其表征为姿态转换网络的权重;最后利用K近邻的方法对行为进行分类识别。实验结果表明,该算法对动态嘈杂背景、人体执行行为速度的快慢具有一定程度的鲁棒性。该算法在两个公用数据库上获得了较好的结果,验证了其有效性。In order to efficiently represent action, this paper proposed an action recognition algorithm based on posture transition network. First, it automatically localized the human body, and extracted the shape-motion features from the human regions of interest. Then, it hierarchically clustered the extracted features to build the posture binary tree, represented the motion sequence as posture sequence, and represented the action as the weight of the posture transition network. Finally, it recognized the action via the KNN algorithm. The experimental results demonstrate that this algorithm is robust to dynamitic clutter background and robust to the action performing speed. This algorithm gotten good performance on two public action datasets verifies its effectiveness.

关 键 词：行为识别 姿态转换 运动特征 形状特征 矩阵余弦相似度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]