一种新的人脸姿态估计算法

Novel Approach for Head Pose Estimation

提出基于正射投影模型的人脸姿态参数估计算法。从图像获取特征点,计算这些点到正面姿态参考模型特征点的最佳姿态变换参数,算法只需获取4个特征点就能确定图像人脸姿态。同时基于该模型提出新的人脸图像姿态拟合方法,以最小化SSD为目标函数,迭代法求姿态参数。该方法的优势是人脸姿态估计并不依赖于特征点检测,给定初始姿态就能找到模型到图像的最佳姿态拟合。模拟图像和真实图像的姿态估计实验结果显示,拟合算法的迭代过程能迅速收敛,并且该算法可有效用于姿态跟踪。

A head pose estimation algorithm based on orthographical projection is presented. Firstly, the feature points are detected on a novel input image. The optimum pose transformation parameters among these points and their counterpart from a reference frontal head model are calculated. The head pose can be determined by only using four detected feature points. Then a novel head pose fitting algorithm is proposed aimed at minimizing the sum of the squared intensity difference between the fitted model face image and the novel input face image by an iterative approach. Its advantage is that the head pose estimation is not dependent on feature points. The given initial pose parameter to fit from the model to the image automatically. The experiment on the pose estimation of the simulated face and the real face image shows that the algorithm converges quickly and can be used for pose tracking.