作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

基于深度图像的人体动作识别方法

刘 飞a,郝矿荣a,b,丁永生a,b,刘 欢a   

  1. (东华大学 a.信息科学与技术学院;b.数字化纺织服装技术教育部工程研究中心,上海 201620)
  • 收稿日期:2013-06-28 出版日期:2014-08-15 发布日期:2014-08-15
  • 作者简介:刘 飞(1988-),男,硕士研究生,主研方向:模式识别,图像处理;郝矿荣,教授、博士后、博士生导师;丁永生,教授、博士、博士生导师;刘 欢,讲师、博士研究生。
  • 基金资助:
    国家自然科学基金资助重点项目(61134009);长江学者和创新团队发展计划基金资助项目(IRT1220);上海市优秀学术带头人计划基金资助项目(11XD1400100);上海领军人才专项基金资助项目;上海市科学技术委员会重点基础研究基金资助项目(13JC1407500,11JC1400200);中央高校基本科研业务费专项基金资助项目。

Human Action Recognition Method Based on Depth Images

LIU Feia,HAO Kuang-ronga,b,DING Yong-shenga,b,LIU Huana   

  1. (a.College of Information Sciences and Technology;b.Engineering Research Center of Digitized Textile and Fashion Technology,Ministry of Education,Donghua University,Shanghai 201620,China)
  • Received:2013-06-28 Online:2014-08-15 Published:2014-08-15

摘要: 为解决人体动作识别中的复杂背景和自遮挡问题,对深度图像进行研究,从深度图像中获取20个人体骨架关节点,在此基础上将动作时间序列的关节角度变化作为人体运动的特征模型。通过改进的动态时间规整算法计算不同动作之间关节角度变化序列的相似性,进行动作识别,以缓解传统DTW算法病态校准的问题。将识别方法在采集的动作数据库和MSR Action3D数据进行验证,实验结果表明,该方法能达到90%以上的识别率。

关键词: 深度图像, Kinect摄像头, 骨架关节点, 关节角度, 动态时间规整, 模板匹配

Abstract: Aiming at complex background and overlap problems in the human Action recognition task, depth images are used to obtain 20 joints in this paper. On the base of them, the joint-angle variation of action time series is proposed as human action feature model. Besides, this paper proposes an improved Dynamic Time Warping(DTW) algorithm to avoid the pathologic alignment that can be caused by original DTW algorithm. Then, the similarity between joint-angle variation series in different actions is calculated with improved DTW in order to recognize actions, i.e. template matching. The proposed method is carried out in self-collected action database and MSR Action3D database. Experimental results show the proposed method gains recognition of more than 90 percent accuracy.

Key words: depth image, Kinect camera, skeleton joint, joint angle, Dynamic Time Warping(DTW), template matching

中图分类号: