视线追踪是一种通过电子、光学设备与软件算法等获取眼动信息以此估计参与者注视方向的技术。视线追踪技术不仅应用于人机交互、辅助驾驶与虚拟现实等多个领域，而且作为重要的工具广泛的应用于研究人类的心理与情感等。随着普通摄像头变得无处不在，自然光源下单目摄像头视线追踪方法成为研究的热点。然而，普通摄像头采集的图像分辨率较低，实现准确的视线追踪算法具有很大的困难。因此，针对自然光源下基于回归的视线追踪方法的关键技术，论文研究了视线追踪预处理、视线眼睛中心的定位、眼动向量参考点的计算、眼动向量的计算、回归函数的构造与多阈值平均二值连通域质心特征方面的内容。具体研究内容如下： 1.视线追踪预处理眼睛状态识别方法研究睁眼与闭眼的眼睛状态对眼睛中心定位的准确性具有很大的影响。论文在眼睛中心定位之前，建立了基于二值图像眼睛状态识别的方法，依据二值图像睁眼与闭眼虹膜区域的不同特性，通过计算二值图像虹膜连通域垂直像素标准差进行眼睛状态的识别。在公共视频数据库Talking Face Video进行眼睛状态识别的验证，眨眼的检测精确度为98.3％。 2.多判决虹膜中心递进定位方法研究 由于眼睑遮挡、眼镜反射与图像噪声等因素的影响，中低分辨眼睛图像中准确的虹膜中心定位具有极大的挑战。论文在上述分析的基础上，提出了基于眼睛图像的多判决虹膜中心递进定位方法，在虹膜中心粗略定位后，根据睁眼或闭眼的眼睛状态，通过不同判决确定是否进一步进行虹膜中心圆拟合精确定位，或者虹膜中心区域质心精确定位。论文将虹膜中心递进定位与多判决条件结合在一起。在虹膜中心粗略定位时，利用眼睛解剖学常数改进并消除了主动轮廓模型的迭代，依据眼睛图像虹膜与巩膜之间颜色差异的特性获取虹膜中心的粗略位置，并建立了有效区域面积比值的判决条件；在虹膜中心圆拟合精确定位时，设计了双圆模型改进了左右虹膜边缘的提取，并对左右虹膜边缘进行最小二乘圆拟合，获取虹膜中心的精确位置，并建立了虹膜边缘数量、半径大小的判决条件；在区域质心精确定位时，参考虹膜颜色较暗的特性，根据睁眼或闭眼的眼睛状态，提取符合虹膜生理尺度比例的二值连通域质心或最大面积的二值连通域质心，获取虹膜中心的精确位置。论文在满足实时性的基本要求下，有效地提升中低分辨率眼睛图像虹膜中心定位的精度，并已经实验验证。在正常的无人工标注情况下，在最具挑战之一的BioID数据库的关键归一化误差e≤0.05范围内，其精度超过了以往最好结果。 3.多因子平稳眼动向量视线追踪方法研究 基于眼动向量的自然光源下视线追踪方法，通常需要事先获得参考点(定点)与动点(虹膜中心)后才能计算眼动向量。由于大角度头部转动造成的遮挡、眼睛附近的光照阴影的影响、视线方向变化而导致眼角参考点位置的不稳定，以及双眼所导致不同视觉的差异，是中低分辨率下视线追踪准确度提升所面临的重要问题。论文提出了多因子平稳眼动向量的视线追踪方法，将相对稳定的脸廓、鼻梁、鼻翼、眼廓特征点(因子)虚拟成参考点，与双眼动点(因子)形成双眼动向量后，通过不同权重(因子)将双眼动向量合成最终的平稳眼动向量；增加了头部姿态角用以改进回归函数，最终得到了视线追踪结果。论文在满足实时性的基本要求下，有效地提升了自然光源下中低分辨率视线追踪准确度，并已在公共数据库与自建数据库上得到了验证。在公共数据库正常无人工标注的情况下，其准确性超过了以往最好结果，并为虚拟参考点与双眼视觉优化提供了思路。 4.多阈值平均质心视线追踪方法研究 以往的虹膜中心作为眼动向量的动点，无法体现因视线变化对眼睑及眼睑与虹膜交界处产生的影响。为了克服上述问题，论文提出了多阈值平均质心的视线追踪方法，通过选取一系列的灰度阈值对归一化的眼睛图像进行二值化，得到一系列的二值连通域来求取平均质心；在平均质心与眼睑、眼角特征点分别形成垂直、水平方向的双眼眼动向量后，通过不同权重将双眼眼动向量合成最终的眼动向量。多阈值平均质心降低了视线变化带来图像误差，提高了视线追踪的准确性，并已在公共数据库与自建数据库上得到了验证。在公共EYEDIAP数据库，其水平方向的准确性超过了以往最好结果，同时，此方法在非个体定标方式中同样取得了较优的结果。 关键词：自然光源，视线追踪，虹膜中心，参考点，多阈值平均二值连通域质心

 Gaze tracking is a technique for estimating the gaze directions of participants by extracting eye movement information through electronic, optical devices and software algorithms. Gaze tracking is not only used in many fields such as human-computer interaction, driving assistance and virtual reality, but also widely used as an important tool to study human psychology and emotions. With the popularity of ordinary cameras, gaze tracking with a single camera under natural light has become a research hot spot. However, it is difficult to implement accurate gaze tracking algorithms because the low resolution images are captured by ordinary cameras. Therefore, in view of the key techniques of the regression-based gaze tracking method under natural light, this paper studies the eye state recognition for preprocessing, eye center localization, the calculation of the anchor point of eye vectors, the calculation of eye vectors, the construction of mapping functions and the feature of multi-threshold average-binary-connected-component-centroid. The specific research contents are as follows: 1.Research on eye state recognition method of gaze tracking preprocessing The accuracy of the eye center localization is greatly affected by the state of the open and closed eyes. Before locating the center of the eye, an eye state recognition method based on the binary image is proposed in this paper. According to the different characteristics of the iris regions in binary images of the open and closed eyes, the vertical pixel standard deviation of the iris connected component of the binary image is calculated for eye state recognition. The eye state recognition method is performed in the public video database Talking Face Video, and achieves an accuracy of 98.3% for blink recognition. 2.A multi-decision iris center progressive localization method Due to eyelid occlusion, spectacle reflection and image noise in moderate and low resolution eye images, accurate iris center localization is an extremely challenging task. Based on the analysis mentioned above, this paper proposes a multi-decision iris center progressive localization method. After the iris center is roughly located, according to the eye states of open or closed, the accurate circle fitting or regional centroid method for iris center localization is determined by different schedulable conditions. The paper combines the progressive localization of the iris center with multiple decision conditions. The paper combines the progressive positioning of the iris center with multiple schedulable conditions. When the iris center is roughly located, the anatomical constant of the eye is used to improve and eliminate the iteration of the active contour model. According to the color difference between the iris and the sclera of the eye image, the rough position of the iris center is obtained, and the schedulable condition of the ratio of the effective area is established.when the iris center is accurately located by circle fitting, the double circle model is designed to improve the extraction of the left and right iris edges, and then the precise position of the iris center is obtained by the least squares circle fitting. Meanwhile, the schedulable conditions of the number of the iris edge and the size of the radius are established. According to the dark iris in the eye image and the eye states of open or closed eye, the regional centroid is used for accurately locating the iris center. The centroid of the binary connected component in proportion to the physiological scale of the iris or the maximum area of the binary connected component is extracted for accurately locating the iris center. Under the basic requirements of real-time, the paper effectively improves the accuracy of iris center localizaiton in moderate and low resolution eye images, and the proposed method has been experimentally verified. In the normal non-manual labeling, the accuracy exceeds the best results of previous work on the most challenging BioID database of the key normalized error e ≤ 0.05. 3.Research on gaze tracking method based on multi-factor stable eye vectors Gaze tracking method under natural light usually necessary to obtain reference point (anchor point, fixed point) and moving point (iris center) before calculating eye vector. The occlusion caused by the rotation of the large angle head, the influence of the shadow near the eyes, the unstable eye corner reference point caused by the changes of gaze direction, and the difference in vision caused by the two eyes, are important issues for improving the accuracy of gaze tracking in moderate and low resolution images. The paper proposes a gaze tracking method based on multi-factor stable eye vectors. The method virtualizes the relatively stable feature points such as facial contour, nose bridge and alar, and eye contour (factor) into reference points. The reference point and the moving points of both eyes (factor) are used to computed the left and right eye vectors, which they are synthesized into the final stable eye vectors by different weights (factor). The head pose angles are used to improve the regression function, which it improves the accuracy of the gaze tracking. Under the basic requirements of real-time, the paper effectively improves the accuracy of gaze tracking under natural light in moderate and low resolution images, and has been verified in public and self-built databases. Its accuracy exceeds the best results of previous work on the public database without the manual labeling, and the paper provides ideas for virtual reference points and binocular vision optimization. 4.Research on gaze tracking method based on a multi-threshold average centroid The iris center is used as the moving point of the eye vector, which it cannot characterize the influence of gaze changes of the eyelid and iris junction. To overcome the above problems, the paper proposes a multi-threshold average centroid gaze tracking method. By selecting a series of gray thresholds to perform binarization on the normalized eye images, the average centroid is obtained by a series of binary connected components. The vertical and horizontal of the left and right eye vectors are computed by the average centroid and the eyelid, the eye corner feature points, respectively. The final eye vectors are computed by implementing different weights in the left and right eye vectors. The multi-threshold average centroid reduces the error caused by the changes of gaze direction, improves the accuracy of the gaze tracking, and the proposed method has been verified in public and self-built databases. The accuracy of the horizontal gaze direction exceeds the best result of previous work on the public database. At the same time, this method also achieves better results for non-specific individuals. Keywords: Natural light, gaze tracking, iris center, anchor point, multi-threshold-average-binary-connected-component-centroid