一种基于非线性优化的人脸姿态高稳定实时估计方法

High-Stability Real-Time Method for Headpose Estimation Based on Nonlinear Optimization

人脸姿态求解是人脸识别系统中的关键技术之一。为了解决少量人脸特征点求解姿态不稳定的问题,提出了一种基于非线性优化的姿态求解方法,该方法首先根据相机成像原理求解出3D点重投影坐标,并根据投影点坐标与观测点坐标关系构建关于重投影误差的最小二乘问题,再结合李代数相关知识求解出最小二乘方程的雅克比矩阵,利用高斯牛顿法在梯度方向上迭代求解最小重投影误差,最终求解出人脸姿态角度。仿真和实验结果表明,在5个特征点情况下能稳定求解出人脸姿态信息,准确性优于其他算法,旋转角度与实际值误差降低至1.9%,平移降低至1.5%。本算法已应用到实际产品人脸合规检测流程中,人脸姿态识别更精确,比传统算法有了8%左右的提升。该算法在不同噪声和不同特征点对的情况下均具有较高的稳定性,满足了人脸识别的实时性要求,其在识别前的合规判断中有重要的应用意义。

Headpose solving is one of the key technologies in face recognition systems.In order to solve the problem of unstable attitude of a few face feature points,this paper proposes a method of attitude solving based on nonlinear optimization.Firstly,the 3D point re-projection coordinates are solved according to the camera imaging principle,and the coordinates and observations are based on the projection points.The point coordinate relationship is used to construct the least squares problem of reprojection error.Then the knowledge of Lie algebra is used to solve the Jacobian matrix of the least squares equation.The Gauss-Newton method is used to iteratively solve the minimum reprojection error in the gradient direction.Face attitude angle.The simulation and real experiments prove that the headpose information can be solved stably under the condition of 5 feature points,the accuracy is better than other algorithms,the error between the rotation angle and the actual value is reduced to 1.9%,and the translation is reduced to 1.5%.This algorithm has been applied to the face compliance detection process of actual products.Face gesture recognition is more accurate,which is about 8%better than the traditional algorithm.