小 型 微 型 计 算 机 系 统 2018 年 月 第 期
Journal of Chinese Computer Systems Vol.30 No. 2018
收稿日期:2018-11-30 收修改稿日期:2019-03-01 基金项目:基于集成学习的个性化人体和服装建模及试衣仿真(61602324),浙江大学
CAD&CG 国家重点实验室开放课题(A1914)资助 作者简介:刘婷,女,1994 年生,硕士研究生,研究方向为虚拟现实;彭晓羽,女,1994 年
生,硕士研究生,研究方向为虚拟现实;谭小慧,女,1977 年生,博士,副教授,CCF 会员,研究方向为虚拟现实等.
单目深度摄像头下的人体尺寸的自动测量方法
刘 婷
1,2
,彭晓羽
1,3
,谭小慧
1,2,3
1
(首都师范大学 信息工程学院,北京 100048)
2
(首都师范大学 电子系统可靠性与数理交叉学科国家国际科技合作示范型基地,北京 100048)
3
(首都师范大学 北京成像理论与技术高精尖创新中心 首都师范大学信息工程学院,北京 100048)
E-mail :xiaohuitan@cnu.edu.cn
摘 要:为了解决在虚拟试衣等相关应用中个性化服装定制获取人体尺寸数据难以精确的问题.本文提出了一种自动测量
人体尺寸的方法.首先建立基于尺度不变的特征描述子和热传播的测地距离关系的随机森林回归模型,根据投票策略预测
预定义的人体关键特征点的位置;随后,将预测所得的关键特征点作为测量人体尺寸的源点,利用深度优先遍历的思想,获
取以源点为根节点的邻近点的最短路径;最后,自动计算最短路径的测地距离.实验结果表明本文的方法对不同体型和不同
姿态的三维人体模型能够准确地和鲁棒地进行提取关键特征点和测量尺寸.因此本文的特征点提取对人体模型的噪声和孔
洞不敏感.从测量结果可知本文测量人体尺寸的相对误差的范围在 0.075~0.001.
关键词:热核特征描述符;随机森林回归;人体关键特征点; 人体尺寸测量
中图分类号:TP301 文献标识码:A 文章编号:1000-1220(2019)02--
A Method of Automatic Measurement of Human Size Based on Depth Camera
LIU Ting
1,2
,PENG Xiao-yu
1,4
,TAN Xiao-hui
1,2,4
1
(School of Information Science and Technology, Capital Normal University, Beijing 10048)
2
(Beijing Key Laboratory of Electronic System Reliability Technology, School of Information Science and Technology, Capital Normal University, Beijing 10048)
3
(Beijing Key Laboratory of Light Industrial Robots and Safety Verification, School of Information Science and Technology, Capital Normal University, Beijing
10048)
4
(Electronic System Reliability and Mathematical Interdisciplinary National International Science and Technology Cooperation Demonstration Base, Beijing
Imaging Technology High-tech Innovation Center , School of Information Science and Technology, Capital Normal University, Beijing 10048)
Abstract: It is difficult to accurately acquire the personal body size by non-contact measurement methods, especially for clothing
customization virtual fitting and other related applications. To meet the challenge, in this paper, automatic body size measurement
method based on geodesic distance is proposed. Firstly, a random forest regression model is established describing the relationship
between scale-invariant feature descriptors and geodesic distances. According to the voting strategy, the location of predefined human
focal feature points can be predicted accurately. Subsequently, the predicted focal feature points are used as the key points for the
measurement of the 3D human body. With the idea of depth-first traversal, a method for calculating the geodesic distance is
constructed, and finally the size of the human body is automatically obtained according to the accurate position of the feature points.
The experimental results show that the proposed method can accurately and robustly extract focal feature points and measure different
3D human body shapes and styles. Therefore, the feature point extraction in this paper is not sensitive to the noise and holes. From the
measurement results, the relative error of measuring 3D human body size is in the range of 0.075~0.001.
Key words: heat kernel feature descriptor; random forest regression; focal feature points; human body size measurement
1 引 言
人体尺寸的测量在服装业、医疗保健、体育和三维虚拟
游戏等领域发挥着重要的作用,利用 3D 扫描技术以非
接触的方式自动测量人体尺寸,不仅节约人力和物力,而且
可以促进服装定制行业的发展。随着虚拟试衣等新型购物方
式逐渐兴起,通过相关的技术手段就可以直接观看到试穿衣