摘要
提出了一种低精度视觉设备布置环境下的单目视觉人体动态定位方法,通过构建的人体定位多元回归预测模型,建立人体从2D图像空间到3D位置空间的映射关系,无需对视觉相机进行内外参数标定。首先,采集人体的2D图像序列和3D坐标位置序列,构建时间同步的视觉人体定位样本集;其次,通过人体骨架检测算法提取样本集中2D图像的人体骨架信息,进而选取并构建适用于人体定位的关键特征;然后,训练基于多元回归的人体定位模型,采用基于L1惩罚项的嵌入型特征选择方法和K-fold交叉验证方法,在模型训练过程中实现特征选择和超参数估计;最后,对获取的人体定位模型进行定位误差评估,实验结果表明模型的均方根定位误差为2.95 cm,验证了所提方法的可行性及有效性。
This paper presents a monocular vision localization approach for dynamic human body in low-precision visual equipment layout environment. By formulating the human body localization multiple regression prediction model, the human body mapping from 2D image space to 3D position space is established. This approach does not need to calibrate the intrinsic and extrinsic parameters of the camera. Firstly, the 2D image sequences and 3D coordinate position sequences of human body are both collected to construct the visual human body localization sample set with time synchronization. Secondly, the human body skeleton information of 2D images in the sample set is extracted by human skeleton detection algorithm. The key features suitable for human localization are selected and constructed. Thirdly, the human body localization model based on multiple regression is trained. The embedded feature selection method based on L1 penalty term and K-fold cross-validation method is employed to implement the feature selection and super-parameter estimation during model training. Finally, the localization error of the obtained human body localization model is evaluated. Experimental results show that the root-mean-square localization error of the model is 2.95 cm, which verifies the feasibility and effectiveness of the proposed method.
作者
杨傲雷
任海燕
费敏锐
徐昱琳
陈灵
Yang Aolei;Ren Haiyan;Fei Minrui;Xu Yulin;Chen Ling(School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200444,China;Shanghai Key Laboratory of Power Station Automation Technology,Shanghai University,Shanghai 200444,China;College of Engineering and Design,Hunan Normal University,Changsha 410081,China)
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2020年第7期252-260,共9页
Chinese Journal of Scientific Instrument
基金
上海市自然科学基金(18ZR1415100)
国家重点研发计划项目(019YFB1405500)
国防基础科研计划项目(JCKY2017413C002)
上海市科委重点项目(16010500300)
国家自然科学基金(61703262)
111引智基地(D18003)项目资助
关键词
视觉人体定位
多元回归
特征构造
特征选择
vision body localization
multiple regression
feature construction
feature selection
