团队简介   研究论文   研究人员   发明专利   研究资助   联系我们
基于局部姿态先验的深度图像3D人体运动捕获方法

苏 乐 1,2  柴金祥 3  夏时洪 1

1 (中国科学院计算技术研究所 前瞻研究实验室 移动计算与新型终端北京市重点实验室 北京 100190) 
2 (中国科学院大学 北京 100049) 
3 (美国德克萨斯州农工大学 计算机科学与工程系 美国 TX 77843-3112)

摘 要
        本文提出一种基于局部姿态先验的从深度图像中实时在线捕获3D人体运动的方法。 关键思路是根据从捕获的深度图像中自动提取具有语义信息虚拟稀疏3D标记点,从事先建立的异构3D人体姿 态数据库中快速检索K个姿态近邻并构建局部先验模型,通过迭代优化求解最大后验概率,实时在线重建3D人 体姿态序列。实验结果表明,本文方法能够实时跟踪重建出稳定、准确的3D人体运动姿态序列;并且只需经过 个体化人参数自动标定过程,能跟踪身材尺寸差异较大的不同表演者;帧率约25fps。因此,本文方法可应用于 3D游戏/电影制作、人机交互控制等领域。
 
Local pose prior based 3D Human motion capture from depth camera

SU Le 1,2  CHAI Jin-Xiang 3  XIA Shi-Hong 1

1 (Beijing Key Lab of Mobile Computing and Pervasive Devices, Advanced Computing Research Laboratory, Institute Of Computing Technology Chinese Academy Of Sciences, Beijing, 100190, China) 
2 (University of Chinese Academy of Sciences, Beijing, 100049, China) 
3 (Texas A&M University, Computer Science and Engineering, Texas 77843-3112, USA))

Abstract
This paper introduces a local pose prior based real-time online approach to capture 3D human animation from a single depth camera. The key idea is to learn a series of local pose prior models with K motion capture examples from a pre-established large and heterogeneous human motion database, according to automatically extracted labelled virtual sparse 3D markers from captured depth image. Then, by solving a Maximum A Posterior (MAP) problem via an iteratively optimization process, the system automatically tracks the 3D human motion sequence. The experiments show that, the proposed approach robustly captures accurate 3 D human motions at 25fps. The proposed tracking system can easily applied to different actors with large different body sizes via an automatically indi vidual body parameters calibration process. The proposed system can widely apply to 3D game/movie produce, human-machine interaction.
Video
Downloads

基于局部姿态先验的深度图像3D人体运动捕获方法. 计软件学报,2016.(中国计算机图形学大会 2016,最佳论文奖 [Paper 1,735KB] [Slides 1,577KB] [Video 189,220KB]