Research On High Precision Gaze Tracking Algorithms And Application In VR System

Gaze tracking technology is a technique for detecting the direction of the human eye or the point of fixation.The gaze tracking algorithm based on 3D model and using RGBD image and the head-mounted gaze tracking algorithm based on polynomial fitting are the two main types of gaze tracking algorithms.This paper has carried out in-depth research on the problems existing in these two types of mainstream gaze tracking algorithms.The gaze tracking algorithm based on the 3D model and using RGBD images has the following two problems: 1)The projection of the center of the three-dimensional iris plane on the imaging plane of the camera does not coincide with the center of the iris in the RGB image,which makes the existing three-dimensional iris localization algorithm difficult to correctly locate the position of the three-dimensional iris plane.2)Ignoring the torsional movement of the human eye,resulting in an error in the Kappa angle correction and the calculation of the parameters of the human eye model.In order to solve these two problems,this paper proposes a gaze tracking algorithm based on iris projection matching function.The main contributions of this algorithm are as follows:1)Propose an iris projection matching function.The function models the projection relationship between the 3D iris plane and the 2D iris image by the principle of projection geometry,and calculates the 3D iris plane and the iris on the image through the matching points of the pixel points and the boundary pixels in the iris projection.2)A three-dimensional iris plane localization algorithm based on iris projection matching function is proposed.In this paper,the three-dimensional iris plane positioning is performed by combining the human eye model with the iris projection matching function.Therefore,the algorithm can accurately represent the iris plane projection relationship by using the projection geometry principle,and improve the accuracy of the three-dimensional iris plane positioning.3)Propose a Kappa angle correction method based on polynomial fitting.In this paper,we establish a polynomial to model the change of Kappa angle rotation caused by human eye torsional motion.Therefore,the algorithm can calculate the degree of Kappa angular rotation according to the rotation direction of the human eyeball,and reduce the influence of the torsional motion on the gaze estimation algorithm.4)The calculation method of human eye model parameters based on iris projection matching function is designed.The purpose of this method is to calculate the human eye parameters required for the algorithms described in contributions 2,3,such as iris radius,iris rotation radius,human eye center,and Kappa angle.Because at this time,the algorithms in contributions 2 and 3 lack human eye model parameters,the algorithms in contributions 2 and 3 cannot be used to solve the problem of iris positioning and neglecting torsional motion when calculating human eye model parameters.Therefore,this method is designed to calculate the parameters of the human eye model.The method solves the problem of inaccurate positioning of the three-dimensional iris plane by introducing an iris projection matching function.And by using the Kappa angle rotation direction as the variable to be solved,the problem that the torsional motion reduces the calculation accuracy of the human eye model parameters is solved.The experimental results show that the proposed gaze tracking algorithm based on iris projection matching function has high accuracy(error is 1.07°-2.28°)and the speed reaches 52.6FPS.For the head-mounted gaze tracking algorithm based on polynomial fitting,there is a problem that the eye corner,the iris center,and the corneal reflection spot are not accurately extracted,which causes the accuracy of the gaze tracking algorithm to decrease.In order to solve this problem,this paper proposes a head-mounted gaze algorithm based on the pupil feature.The algorithm does not use the eye movement feature vector,but directly takes the pupil center feature into the polynomial mapping equation to calculate the fixation point.Moreover,in order to solve the problem that the pupil center feature is less robust to head shake.In this paper,a dither compensation algorithm based on pupil orientation is proposed.After the jitter occurs,the corresponding relationship between the pupil center coordinates before and after the jitter can be calculated through the pupil orientation,and the offset of the pupil center feature is compensated,thereby increasing the pupil center.The robustness of the feature to jitter.Finally,based on the proposed head-mounted gaze and algorithm based on the pupil feature,a VR gaze tracking system is built to test the performance of the algorithm and make the algorithm practical.The experimental results show that the error of the proposed head-mounted gaze tracking algorithm based on the pupil feature is 12.7 pixels,which is robust to jitter(the correction error after jitter is 33.0 pixels),and the algorithm runs at 47.17 FPS.