基于Latent SVM的多视角行为识别方法.pdfVIP

第34 卷第6 期 Vol. 34 No.6 吉林大学学报(信息科学版) 2016 年 11 月 Joumal of Jilin University (Information Science Edition) Nov. 2016 文章编号: 1671-5896(2016)06-0747，，(后 基于 Latent SVM 的多视角行为识别方法 王丹1 气喊雪柏2 ，陈奋君2 (1.北华大学信息技术与传媒学院，吉林吉林 132011; 2. 吉林大学计算机科学与技术学院，长春 13∞12 ) 摘要:为了在静态图像中获取有效信息，构建行为模型，提出了行为覆盖区 ACA( Action Coverage Area) 和行为 核心 AC( Action Core) 的概念，基于Latent SVM (Support Vector Machine) 目标识别方法，设计了一种多视角行为 模型 MVAM (Multiple Viewpoint Action Model) 。建立了独立的用于行为模型训练和测试的行为数据库。实验表 明，该表示法对静态图像中的人体行为能有效地进行分类和检测。 关键词:行为识别;隐变量支持向量机;行为覆盖区:行为核心;多视角行为模型 中圄分类号: TP182 文献标tR码:A Multiple-Viewpoint Human Action Recognition ßased on Latent SVM WANG Dan 1 2 , ZANG Xueba? , CHEN Fenjun2 • (1. College of Infonnation Technology , Beihua University , Jilin 132011 , China; 2. Department of Computer Science and Technology , Jilin University , Changchun 13∞12 ， China) Abstract: We study the concepts of ACA (Action Coverage Area) and AC (Action Core) , and design a multiple-viewpoint action model based on the advanced object recogniton method Latent SVM ( Support Vector Machine). Multi-viewpoint ACA can succinctly describe the non-rigid change and appearance variation due to di旺erent viewpoints and multi-viewpoint AC , because the auxiliary of ACA can help to improve the efficiency and robustness of action recognition. We created an independent dataset of human action for model training and testing. Experiments showed that our model had high performance in classifying and detecting human action in still images. Key words: action recognition; latent suppo民 vector machine ( SVM) ; a