摘要
基于虚拟现实(VR)头戴式显示器(HMD)设备的远程会议或远程呈现是一个非常有趣和有前景的应用,因为HMD可以为用户提供身临其境的感受,但是面对面通信时,无法实时获取面部表情,阻碍了正常的交流。针对这一问题,本文提出了一种基于卷积神经网络(CNN)的解决方案,用于HMD用户的实时3D面部重建,首先通过据合成生成大量的HMD面部标签数据集,然后使用该数据训练一个面部神经网络,以获得3D人脸模型参数。大量的实验结果表明,该系统可以高效、有效地生成具有面部姿态、面部表情的3D头像。
Remote conferencing or telepresence based on virtual reality(VR)head-mounted display(HMD)devices is a very interesting and promising application,because HMD can provide users with an immersive experience,but can not be real-time in face-to-face communication.Get facial expressions that hinder normal communication.In response to this problem,this paper proposes a convolutional neural network(CNN)-based solution for real-time 3D facial reconstruction of HMD users.First,a large number of HMD facial tag data sets are generated by data synthesis,and then the data is trained using the data.A facial neural network to obtain 3D face model parameters.A large number of experimental results show that the system can efficiently and efficiently generate 3D avatars with facial gestures and facial expressions.
作者
张雁鹏
陈世均
高建勇
ZHANG Yan-peng;CHEN Shi-jun;GAO Jian-yong(Suzhou Thermal Power Research Institute Co.,Ltd.Equipment Management Department,Shenzhen 518028,China)
出处
《传感器世界》
2020年第4期28-32,5,共6页
Sensor World
基金
国家科技重大专项一高温堆示范工程可靠运行技术研究课题运行可靠性关键设备保障技术研究子课题资助(NO.2018ZX06906012)
