基于多特征融合的运动人体行为识别  

Human action recognition based on multi-features fusion

作　　者：赵海勇[1] 李成友[2] 

机构地区：[1]聊城大学计算机学院,山东聊城252000 [2]聊城大学网络信息中心,山东聊城252000

出　　处：《计算机应用研究》2012年第8期3169-3172,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61104179)

摘　　要：提出了一种基于特征级融合的运动人体行为识别方法。应用背景差分法和阴影消除技术获得运动人体区域和人体轮廓;采用R变换提取人体区域特征,采用小波描述子提取人体轮廓特征;然后将这两种具有一定互补性的特征采用K-L变换进行融合,得到一个分类能力更强的特征;最后,在传统支持向量机的基础上,结合模糊聚类技术和决策树构建多级二叉树分类器,从而实现行为多类分类。该方法在Weizmann行为数据库上进行了实验,实验结果表明所提出的识别方法具有较高的识别性能。This paper presented a recognition algorithm of human action based on multi-features fusion. First, through the background subtraction and shadow elimination, obtained human motion silhouettes and contours. It described human silhouette features by R transform. Then fused these features with human contour features which were extracted using wavelet descriptor by K-L transform to get a new feature which had higher discriminating power. At last, through integrating fuzzy clustering and decision tree, constructed a multilevel binary tree classifier. It evaluated the method on the Weizmann action dataset. The experimental results show that the approach has efficient recognition performance.

关 键 词：行为识别 特征融合 小波描述子 支持向量机 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]