基于表观的归一化坐标系分类视线估计方法

视线估计能够反映人的关注焦点，对理解人类的情感、兴趣等主观意识有重要作用。但目前用于视线估计的单目眼睛图像容易因头部姿态的变化而失真，导致视线估计的准确性下降。提出一种新型分类视线估计方法，利用三维人脸模型与单目相机的内在参数，通过人脸的眼睛与嘴巴中心的三维坐标形成头部姿态坐标系，从而合成相机坐标系与头部姿态坐标系，并建立归一化坐标系，实现相机坐标系的校正。复原并放大归一化得到的灰度眼部图像，建立基于表观的卷积神经网络模型分类方法以估计视线方向，并利用黄金分割法优化搜索，进一步降低误差。在MPIIGaze数据集上的实验结果表明，相比已公开的同类算法，该方法能降低约7.4%的平均角度误差。

Gaze Estimation Method Using Normalized Coordinate System Classification Based on Apparent

Gaze estimation can naturally reflect people's focus of attention, and plays an important role in understanding human emotions, interests and other subjective consciousness.However, monocular images used for gaze estimation tend to be distorted due to head pose changes, which reduces the accuracy of gaze estimation.This paper proposes a new classification-based gaze estimation method.A three-dimensional face model and the inherent parameters of monocular camera are used to form a head pose coordinate system through the three-dimensional coordinates of the center of eye and mouth.The camera coordinate system and the head pose coordinate system are combined to establish a normalized coordinate system, and the camera coordinate system is corrected.The gray eye image is obtained by restoration, magnification and normalization.Finally a classification method using an appearance-based convolution neural network model is established to estimate gaze direction, and the golden section method is used to optimize the search process and further reduce errors.The experimental results show that compared with other similar methods, the proposed method can reduce the average angle errors by about 7.4% on the commonly used MPIIGaze test dataset.