Research On Human Eye Gaze Tracking Technology In Complex Light Environment

Gaze tracking,also known as gaze estimation or eye tracking,is a technique that uses various detection means,such as electronics and optics,to obtain the current gaze direction or fixation point of the subject.The early gaze tracking device mainly attaches the mechanical rotating device,electrode,lens or coil directly to the eyeball surface of the subjects,and only can be used in the controlled environment.The gaze tracking technology based on optical video uses the optical charateristics of eyeball,records eye movement through the external camera,which has little interference to the subject and has a wider application range.After the rapid development in recent years,eye tracking technique has achieved high accuracy in the indoor environment with controlled illumination.However,in the daliy light environment,the prediction of gaze direction is severely disturbed by the complex illumination conditions such as low lighting,side lighting,high lighting and occlusions,as well as the large rotation of head pose.In view of the difficulties of eye tracking technique under the complex illumination conditions and based on the characteristics of the human eye gaze,a series of remote gaze estimation methods are proposed.This thesis studies the key technical issues such as robust image appearance representation,insufficient training dataset,flexible eye gaze estimation method,explicit calibration in real environment,head pose tracking under large head rotation and etc.On this basis,driver's gaze tracking method and automatic calibration method in real driving environment are also proposed.The effectiveness of the proposed gaze estimation methods are validated under typical complex light environment.The main research contents and results of this thesis are as follows:(1)Appearance-based eye gaze estimation method using deep features is studied,which solves the problem of image appearance under complex illumination conditions.Based on the combination of feature extraction method of deep feature and feature regression method by cluster-to-classify Random Forest,the method builds a feature space with sparse activation characteristics and uses it to learn how to predict the gaze direction.The Random Forest reduces the error within the nodes by prior clustering and classification,and then builds regression model for sampled data in a completely splitting way.The experiment results show that the gaze estimation method has robustness and certain anti-occlusion ability under different illumination conditions.(2)Eye images augmentation method combining Convolutional Neural Network(CNN)and Generative Adversarial Nets(GANs)is studied,which is used to refine the realism of massive synthetic eye inages and solves the problem of insufficient training dataset under different illumination conditions.First,this method uses CNN to increase the real background information with high-level semantics,and then uses GANs to continue iterative learning to improve the realism of the image.The experimental results show that the gaze estimation using refined eye images has a lower gaze estimation error.(3)A gaze estimation based on neighbor selection and neighbor regression is proposed,which can more effectively and extensively utilize massive eye images data and solve the problem of extensive generalization of the gaze estimation model under massive data.Based on the characteristics features of head pose,pupil center and eye images,this method structure a cascaded three-space neighbor selection framework to select the nearest proximity data that is more similar to the test eye images.The experimental results show that the neighbor-based gaze estimation has a better gaze estimation accuracy on datasets with different light conditions.(4)An novel automatic calibration method for driver's gaze zone based on the monocular camera is proposed,which solves the problem of driver's gaze calibration in real driving environment.Based on the vehicle environment and driver's gaze behavior characteristics,the proposed method estimate head pose based on the three-dimensional geometry of the facial face landmarks.In addition,auxiliary sampling particle filter method is used to track and learn the driver's head pose in the given state space to achieve automatic calibration.The experimental results show that the proposed method has an error less than 2 degree on head pose estimation,which can meet the needs of attention-related driving behavior research.(5)Driver's gaze zone estimation method combining gaze estimation and head pose tracking with a depth camera is studied,which verifies the effect of the eye gaze estimation method and solves the problem of driver's gaze estimation in real complex environments.The studied method use depth information for background segmentation to quickly extract 3D point cloud of face region,and iterative calculate the matching between face region point cloud and the pre-extracted point cloud for different gaze zone.Thus,the rotation matrix of head movement is obtained,and the head pose estimation value is given by head pose tracking.The experimental results show that the gaze estimation error of the proposed method is less than 8 degree,and a stable driver's gaze zone estimation in real driving environment can be realized.