Research On Methods Of Human Eye Gaze Estimation With Head-Mounted Eye Trackers

The head-mounted eye-tracking system is a natural tool for visual attention analysis and human-computer interaction,which has become a hot research area in the international arena.Based on the analysis of the current status of domestic and international research,this thesis conducts systematic and in-depth research on the challenging problem of how to estimate the human eye gaze point with high accuracy,from the aspects of human eye pupil detection in real environments,3D gaze estimation,and head-mounted eye-tracking system design.The main innovations are as follows:1.A pupil detection method driven by human eye image characteristics is constructed to solve the poor stability of pupil detection in real environments.Through human eye images'observation and experimental study,it is found that there are characteristics of high local contrast and large global darkness in the pupil image region.By combining the two characteristics,a horizontal weighted Haar-like feature to represent the pupil image region is established.We propose a compact pupil region detection method(CPRD)and a refined method to improve pupil region detection using image sequence correlation(CPRD-init).On a dataset of more than 100,000 images,the detection success rate and the 5-pixel error rate of the CPRD-init method are improved by 7.1%and 3.9%,respectively,compared with the state-of-the-art methods.2.To address the problem that the eye pseudo-center degrades the accuracy of 3D gaze estimation,a mapping-surface-based gaze direction estimation method is proposed.Through the simulation and sensitivity analysis of the eye pseudo-center,the effect of the eye pseudocenter error on the gaze estimation accuracy is discussed.It is found that one eye's gaze directions approximately intersect the eye pseudo-center when the eye pseudo-center error is in the millimeter range and the distance between the calibrated mapping surface and the gaze point is within 0.5 m.A high-quality calibration method for the eye pseudo-center is established,and a single mapping surface-based gaze estimation method is constructed.A double mapping surface-based gaze estimation method is proposed by replacing the eye pseudo-center with another mapping surface.This method improves the gaze estimation accuracy by 20%and 17%,respectively,compared with existing methods on two datasets with gaze depth ranges of 1 m and 2 m.3.To address the slippage problem of long-time wearing eye trackers,a robust and accurate gaze estimation method based on the real pupillary axis is proposed.By establishing the corneal optical system model,the law of pupil refraction through the cornea is revealed,and the function from the pupil image contour captured by the eye camera to the real pupil normal vector is derived.A 2D gaze estimation method using the real pupil normal vector and its calibration data filtering method are established.A 3D gaze estimation method that converts the real pupillary axis to the gaze direction using the analytic method is proposed,which is slippage-robust as the real pupillary axis is insensitive to slip.Experimental results show that the method can achieve comparable accuracy to the state-of-the-art pupil centerbased gaze estimation method,but the latter has the slippage problem.4.An experimental validation platform of a head-mounted eye-tracking system for real environments is constructed to verify the theories and methods proposed in this thesis.