Objective Compared to the appearance-based method, Pupil Center Corneal Reflection (PCCR) method has higher accuracy, and has great significant application value in fields such as human-computer interaction and AR/VR. However, the existing researches mainly discuss the structural forms of eye tracking devices such as single camera single light, single camera multiple light, and multi camera multiple light, lacking quantitative hardware layout optimization methods. At the same time, in the PCCR method, the cornea is considered as a spherical surface, while human cornea is not ideal spherical, only the central part of it can be approximated as spherical. Especially when the distance between the light sources is too far, it can cause glints at the edge of the cornea, where the cornea is not spherical. The main solution is to make the glints as close as possible to the spherical area of the cornea, but if the distance between the light sources is too close, significant calculation errors will emerge. Therefore, optimizing the light layout of eye tracking devices has strong practical significance, when the distance between light sources is mutually constrained.

Methods The method of computer simulation is used to theoretically calculate the average error of gaze estimation under different light layout, thereby deriving the optimal light source position. The main simulation process is divided into three parts, including eye imaging, gaze estimation based on simulation images, and construction of loss functions. The non spherical surface of the human cornea is considered in the content of eye modeling and imaging, which is closer to the actual human cornea. A typical pinhole model with superimposed distortion is adopted in the camera imaging model. For the optical axis reconstruction in gaze estimation, a typical optical axis reconstruction method suitable for single camera multi light devices has been adopted, which first estimates the corneal center position and then calculates the gaze optical axis. In the loss function, the error statistics of uniformly distributed test points are used as comprehensive evaluation indicators to calculate the iteration step of the light source position.

Results and Discussions Different devices have different optimal light layouts, due to different camera focal lengths, installation positions, and user eye parameters. As an example, typical human eye parameters are used, and the optimal light layout for different forms of devices is discussed for both head mounted and desktop devices. In head mounted devices, the light source can be distributed in a circular pattern, and 7 discrete positions around the eye are taken as rough positions. After obtaining the optimal discrete position sets from 42 sets (Fig.14), as the initial value for optimization, the optimal light position can be iteratively optimized. Similarly, the light source of desktop devices is distributed in a linear pattern at the bottom of the screen. 8 discrete positions are taken and the optimal set of discrete positions is obtained from 64 sets (Fig.16). Then, the optimal light position can be iteratively obtained. The quantitative results indicate that there is an optimal position for the light source, and it can improve the accuracy of gaze estimation. Compared to existing studies, the results obtained by this method are more quantitative, and the ellipsoidal cornea is closer to the human cornea.

Conclusions The PCCR method is an effective method for gaze estimation, but the cornea is considered as a sphere, which limits the accuracy. A quantitative optimization method of light layout has been proposed, and a framework for light layout optimization has been established, which can optimize different optimal light layouts according to actual application settings. At the end, typical system parameters are simulated, and the results show that the distance between the light sources is not necessarily better as it is larger or smaller. There is an optimal light source position that can suppress the error increase caused by ellipsoidal corneal. The simulation results can also serve as a theoretical reference for corresponding device design.