基于面部特征点的单幅图像人脸姿态估计方法

针对目前基于学习的姿态估计方法对训练样本及设备要求较高的问题，提出一种基于面部特征点定位的无需训练即能估计单幅图像中人脸姿态的方法。通过Adrian Bulat人脸特征点定位器和Candide-3构建稀疏通用人脸模型并获得五官特征点，确定模型绕Z轴的旋转范围及搜索步长，在指定Z轴旋转角度下，使用修正牛顿法通过模型的旋转、平移及缩放变换对齐模型和图像中人脸五官角点，得到该角度下模型绕X轴、Y轴的旋转角度及绕Z轴候选角度下的损失函数值，根据最小损失函数值确定人脸绕3个轴旋转的最佳值。实验结果表明，该方法能够快速估计自遮挡的大姿态角度人脸，在公共人脸库Multi-PIE、BIWI和AFLW上的平均误差分别为3.79°、4.37°和6.04°，明显高于同类人脸姿态估计算法，具有较好的实用性能。

Facial Pose Estimation Method on Single Image Based on Facial Feature Points

To deal with the high requirements of learning-based pose estimation methods on training samples and devices,this paper proposes a method based on facial feature point positioning that can estimate the facial pose in a single image without training.By Using the Adrian Bulat face feature point locator and Candide-3,a sparse general face model is constructed and the facial features are obtained.The rotation range of the model around the Z axis and the search step length are determined.Then within the rotation range around the Z axis,the improved Newton iterative algorithm is used to align the key facial feature points of the 3D model and the image through translating,rotating,and scaling the 3D model.As a result,the rotation angles of the model around the X and Y axes,and the loss function value under the candidate angle around the Z-axis are obtained.Finally,the method selects the optimal facial angles around three axes based on the minimal loss function value.Experimental results show that this method can quickly estimate selfoccluded faces with large pose angles and has good practical performance.Its average errors on the public face databases including Multi-PIE,BIWI and AFLW are 3.79°,4.37°and 6.04°respectively,which are significantly higher than similar facial pose estimation algorithms.