基于深度稠密时空兴趣点的人体动作描述.pdfVIP

第28 卷 第 10 期 模式识别与人工智能 Vol.28 No.10 2015年10月 PR AI Oct. 2015 基基于深度稠密时空兴趣点的 人体动作描述算法* 宋健明 张 桦 高 赞 张 燕 薛彦兵 徐光平 (天津理工大学 计算机视觉与系统教育部重点实验室 天津300384) (天津理工大学 天津市智能计算及软件新技术重点实验室 天津300384) 摘 要 目前基于深度数据的动作识别算法得到极大关注,至今仍无一种鲁棒、区分性好的基于深度数据的动作 描述算法.针对该问题,文中提出基于深度稠密时空兴趣点的人体动作描述算法.该算法选择多尺度深度稠密特征 时空兴趣点,跟踪兴趣点并保存对应轨迹,基于轨迹信息描述动作.通过在DHA、MSRAction3D和UTKinect深度动 作数据集上评估可知,与一些代表性算法相比,文中算法性能更优. 关键词 深度数据,稠密时空兴趣点,人体动作描述,轨迹跟踪 中图法分类号 TP 181 DOI 10.16451/ ki.issn1003-6059.201510009 Human Action Description Algorithm Based on Depth Dense Spatio-Temporal Interest Points SONGJian-Ming,ZHANG Hua,GAOZan,ZHANG Yan,XUE Yan-Bing,XU Guang-Ping (Key Laboratory of Computer Vision and System of Ministry of Education, Tianjin University of Technology,Tianjin300384) (Tianjin Key Laboratory of Intelligence Computing and Novel Software Technology, Tianjin University of Technology,Tianjin300384) ABSTRACT Much attention is paid to action description algorithm based on depth data now. However,there is no robust,efficientanddistinguishingfeaturerepresentationfordepthdata. Tosolvetheproblem,human action description algorithm based on depth dense spatio-temporal interest point is proposed. Multi-scale depth dense feature spatio-temporal interest points are selected and then tracked,and the trajectories of these points are saved. Finally,the trajectory information isutilized to represent human action. Through the evalu