Research On Gaze Tracking Technology Based On Kalman Filter And Image Saliency Detection

In recent years,with the development of artificial intelligence needs,gaze tracking technology has received more and more attention.In the field of humancomputer interaction,the gaze tracking technique is often realized by the pupil location algorithm and the two-dimensional gaze point mapping algorithm.The tracking accuracy and robustness need to be improved due to the change of head pose,illumination and eye movement in the eye movement condition Therefore,this thesis proposes a kind of Kalman pupil tracking algorithm based on binocular synchronous motion constraint,and then obtains the position of the gaze point by the twodimensional polynomial mapping function,from the point of view of improving the tracking accuracy of dynamic fixation.Under the established eye movement route,it is compared with other commonly used filtering algorithms,and the superiority of the algorithm is verified by the experimental data.And the gaze point estimation information obtained by the algorithm is compared with the position of gaze point obtained by the saliency algorithm to verify the consistency in the case of simulation.The main contents are as follows:First,the tracking technique of gaze point is achieved by tracking the position of the pupil center and the cornea reflection with the mapping between the pupil position and the position of computer screens.The precise pupil localization is inseparable with the dynamic pupil tracking.In this thesis,the convex area voting and model constrained pupil location algorithm based on starburst algorithm are used to locate the corneal reflection and pupil center under infrared irradiation.Aiming at the localization of corneal reflection,this thesis uses an adaptive threshold segmentation of the spot detection algorithm to locate the largest spot in multiple bright spots in the cornea.In this thesis,the pupil edge candidate points are obtained by the starburst algorithm,and then the tangents perpendicular to the gradient are drawn at the candidate points used to divide the half plane to determine the pupil area to eliminate the noise point in the edge candidate points.Finally,random sampling consensus algorithm is used for ellipse fitting to locate the pupil center.Second,the Kalman filtering model is built to track the pupil under the moving state.Based on the law of binocular synchronous motion,the process equation and observation equation are constructed with the vector information of the distance between the left and right eyes as the state variables.The constant velocity Kalman filtering equation is simplified,and the dimension of the equation and the number of iterations is reduced.Third,the marked points on the computer screen are set up as the referenced gaze points.The mapping relationship between the vectors and the screen mark points is fitted with the quadratic polynomial model.After the mapping relationship is obtained,the pupil center-corneal reflection vector information filtered under the established gaze action is converted into the gaze point prediction information,and the robustness of the algorithm is verified in the case where there is a large disturbance noise point in the pupil position measurement.The results are compared with other commonly used filtering algorithms to verify the advantages of this algorithm for improving the tracking accuracy and shortening the running time of the program.Fourth,the non-task-driven bottom-up saliency algorithms of predictive gaze are studied.The experiment is carried out under the condition of gaze simulation,and the subject is tested with the real scene image as the target image.The probabilistically distribution information obtained from mapping transformation and the viewpoint information obtained by the saliency algorithm are compared to verify consistency.Experiments show that the proposed Kalman filter algorithm based on binocular synchrony constraint has strong robustness in gaze tracking.Compared with other filtering algorithms,the proposed algorithm has less running iterations and computation time,which proves that the proposed algorithm can meet the requirement of the gaze tracking system.When the static target of the simulation scene is gazed,it is consistent with the predicted point of view information which is estimated by the saliency algorithm.