基于支持向量机的人体姿态识别  

Human posture recognition method based on SVM

作　　者：郑莉莉[1] 黄鲜萍[1] 梁荣华[1] 

机构地区：[1]浙江工业大学计算机科学与技术学院,浙江杭州3100023

出　　处：《浙江工业大学学报》2012年第6期670-675,691,共7页Journal of Zhejiang University of Technology

基　　金：浙江省自然科学基金重点资助项目(Z1090630);浙江省自然科学基金资助项目(Y12F020180);浙江省科技计划项目(2012C23122)

摘　　要：人体姿态识别是当前自动视频理解技术的研究热点,提出的算法能够识别视频中的十二种人体姿态,包括走路、跳跃、爬行和弯腰等.算法首先提取前景图像中的星形轮廓位置、六星角度和离心率等多特征信息,组成人体姿态特征,结合该特征的向量表示利用基于径向基核支持向量机的分类器,实现各种姿态的识别.实验中,在公共数据库和部分自采集数据基础上构建特征库,对分类器进行训练,对其余自采集数据进行分类,结果表明:该算法对小样本下的人体姿态识别具有令人满意的结果.The human body posture recognition is a hot research area in current automatic video understanding technology. In this paper, an algorithm is proposed to recognize twelve kinds of human body postures in videos, such as walking, jumping, crawling, stooping and so on. In this algorithm, the multiple features of position of star skeleton, angles of six sticks, and eccentricity in the foreground image are extracted and are constructed as body posture vector. Combining with the feature vector, the classifier of RBF kernel-based support vector machine is used to recognize the different human postures. In the experiment, the feature library is constructed based on public databases and part self-collected data and the classifier is trained to classify the rest of the self-collected data. The results demonstrate that the proposed method has a satisfied result on the human body posture recognition under the small samples.

关 键 词：姿态识别 星形模型 六星角度 离心率 支持向量机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]