身体姿态和面部捕捉及虚拟角色动作迁移研究

Motion and Facial Capture and Virtual Character Motion Migration

针对虚拟人角色动作捕捉设备成本过高、识别准确率较低和操作过于繁琐的问题,提出了一种基于单目摄像头的运动捕捉系统及其实现方法。方法基于OpenPose算法获取每个人的2D实时运动关节点,进而通过一个线性回归神经网络将2D关节转换为3D关节点,可以实时采集多人运动的图像并计算各关节位置姿态,从而实现人体的运动捕捉以及向虚拟角色的迁移。还研究了一种面部捕捉算法,使用径向基函数的插值方式,将面部控制点信息转换为三维网格的实时变形动画效果。实验表明,上述方法操作简单、价格低廉、识别准确率高、运行稳定,可以用于影视特效和游戏娱乐素材制作,以及各种交互系统当中。

This paper presents a motion capture system based on monocular cameras, with which users can capture images of multiple people’s movements in a simple way. The method is based on the OpenPose algorithm to obtain 2 D real-time motion joints of each person, and then convert the 2 D joints into 3 D joints through a linear regression neural network, which captures images of multiple people’s movements and calculates the pose of each joint in real-time. This paper also investigates a simple facial capture algorithm that uses radial basis function interpolation to convert facial landmark points into 3 D mesh points for real-time deformation animation. The solution is simple to implement, inexpensive and stable, and can be used in film and entertainment production, as well as in various interactive systems.