基于动作串的人体行为识别  

Human Action Recognition Using Action String

作　　者：赵海勇[1] 李俊青[1] 

机构地区：[1]聊城大学计算机学院,聊城252000

出　　处：《计算机科学》2013年第10期296-300,共5页Computer Science

基　　金：国家自然科学基金(61104179);山东省高校智能信息处理与网络安全重点实验室资助

摘　　要：提出了一种以运动人体侧影为特征的基于模板匹配的人体行为识别方法。首先,利用背景差分法和阴影消除技术提取运动人体侧影。利用R变换对人体侧影进行特征提取,将时变的2D区域形状转换为对应的1D距离向量。然后,利用谱系聚类方法提取动作序列的关键姿态,将关键姿态编码为称为动作串的模板。最后,利用动态时间规整算法度量测试序列与标准模板之间的相似性。实验结果表明,本方法对人的6种日常行为进行识别的正确识别率达到85%以上,具有简单实用的特点。A template match-based method was presented by using the silhouettes of image sequence as representative descriptors of human posture to achieve human actions recognition. First, human silhouette extraction was obtained by background subtraction and shadow elimination. A new silhouette descriptor based on R transform was defined, which is employed to transform the silhouette into a 1D distance signal. Hierarchical clustering method was proposed to extract key postures of human action. Then, key postures were coded as a template called action string. At last, dynamic time warping was proposed to measure the similarity of the templates and the test sequences. This algorithm was evaluated using six human daily actions. The experimental results show that this method can achieve the correct recognition rate above 85 %.

关 键 词：计算机视觉 人体侧影 关键姿态 聚类 行为识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]