基于混合特征的注视方向判别

摘要

提出了一种基于混合特征的注视方向判别方法.混合特征由模型特征和表观特征组成.模型特征是提取的特征点间的几何向量,表观特征是从眼睛图像提取的方向二值模式(DBP).将两种特征通过支持向量回归(SVR)算法融合起来,将组合特征一一对应于某一确定的头部姿态下的一个确定的注视方向.非特定人实验所用数据库有11个采集人,共计4089个样本.所用的样本在采集时保持头部正面面向摄像机,仅双眼注视预定的方向.实时测试时仅用一个单摄像机,输入是单帧人脸图像,输出为以摄像机坐标系计量的欧拉角度值.实验验证了混合特征的有效性,实验结果为3°的测试误差.%This paper proposes a method for gaze estimation based on the hybrid feature which comprises the model-based feature and the appearance-based feature. The Model-based feature is the vectors between the located marks of couple eyes.The appearance-based feature is the directional binary pattern (DBP), which treats an eye image as a high-dimension feature。 This article uses the support vector regression (SVR) to rank the hybrid feature, which maps a gaze direction with the frontal head pose. In the person-independent experiment, the dataset for training SVR included 4089 samples of 11 persons. In collecting the training data, the subject faced a camera within a frontal head pose and gazed the pre-defined points on the screen. In the real-time experiment, only one camera was used. The input was a captured picture and the output was the Euler angles with respect to the camera coordinate system. The experimental results showed that the proposed method was efficient and the accuracy is around 3°.