电子学报 ›› 2017, Vol. 45 ›› Issue (12): 3060-3069.DOI: 10.3969/j.issn.0372-2112.2017.12.032

• 科研通信 • 上一篇    下一篇

基于高斯增量降维与流形Boltzmann优化的人体运动形态估计

李万益1,2, 孙季丰1   

  1. 1. 华南理工大学电子与信息学院, 广东广州 510641;
    2. 广东第二师范学院计算机科学系, 广东广州 510303
  • 收稿日期:2015-01-28 修回日期:2016-11-30 出版日期:2017-12-25
    • 通讯作者:
    • 李万益
    • 作者简介:
    • 孙季丰,男,1962年生于广东广州,现为华南理工大学电子与信息学院教授,博士生导师.主要研究方向为数字视频处理,生物信息处理和DNA数据建模和压缩,人体形态分析及建模,自组织无线网络等.E-mail:ecjfsun@scut.edu.cn
    • 基金资助:
    • 国家自然科学基金 (No.61202292); 广东省自然科学基金 (No.9151064101000037); 广东省普通高校青年创新人才项目 (No.2016KQNCX111)

Human Motion Estimation Based on Gaussion Incremental Dimension Reduction and Manifold Boltzmann Optimization

LI Wan-yi1,2, SUN Ji-feng1   

  1. 1. School of Electronic and Information Engineering, South China University of Technology, Guangzhou, Guangdong 510641, China;
    2. Department of Computer Science, Guangdong University of Education, Guangzhou, Guangdong 510303, China
  • Received:2015-01-28 Revised:2016-11-30 Online:2017-12-25 Published:2017-12-25
    • Supported by:
    • National Natural Science Foundation of China (No.61202292); National Natural Science Foundation of Guangdong Province,  China (No.9151064101000037); Youth Innovative Talent Program of Universities in Guangdong Province (No.2016KQNCX111)

摘要: 为了从多视角轮廓图像估计出含空间位置信息的三维人体运动形态,该文提出高斯增量降维与流形Boltzmann优化(GIDRMBO)算法.该算法把表示三维人体运动形态的高维数据分成表示空间位置信息和姿态信息两段子向量后,用高斯增量降维模型(GIDRM)分别对其样本进行降维,建立相应的低维空间及映射关系,然后在相应的低维空间使用流形Boltzmann优化算法来对轮廓匹配目标函数进行优化,从而实现估计.其中,所提算法分别利用了两段子向量样本的低维数据作为先验信息,可较好的避免陷入局部最优区域进行搜索,最终生成与各视角原始运动图像匹配且含空间位置信息的三维人体运动形态.经仿真实验验证,所提算法与常用粒子滤波算法相比,其估计误差小,并且还能起到消除轮廓数据歧义和克服短时遮挡的作用.

关键词: 高斯增量降维模型, 流形Boltzmann优化, 人体运动形态, 轮廓图像, 子向量

Abstract: In the interest of estimating the human motion in 3D (3 Dimensions,3D) and its spatial position from multi-view silhouettes,an algorithm called Gaussion incremental dimension reduction and manifold Boltzmann optimization (GIDRMBO) is proposed.After the high dimensional data denoting human motion in 3D is divided into two subvectors,the two subvectors denote information of spatial position and pose respectively.The proposed algorithm takes the advantage of Gaussion incremental dimension reduction model (GIDRM) to reduce the dimension of samples of the two subvectors respectively,so that the low dimensional spaces and relevant mappings can be built,and the optimization of objective function matching silhouettes can be carried out to achieve the estimation through the manifold Boltzmann optimization in the low dimensional space.The proposed algorithm can utilize the prior informations from low dimensional data of the samples of the two subvectors respectively.It can skip the area of local optimization better during searching,and generate human motion in 3D and its spatial position which can match the original motion images of multi-view.By the comparison of some traditional particle filters in the experiments,the proposed algorithm has better performance in lowering estimation error,disambiguating the silhouettes and overcoming the transient occlusion.

Key words: Gaussion incremental dimension reduction model (GIDRM), manifold Boltzmann optimization, human motion, silhouettes, subvector

中图分类号: