Research On Gaze Estimation In Free Posture

The main task of gaze estimation is to estimate the direction of people's attention or the point of regard.As one of the important research of computer vision,gaze estimation has important application value in human-computer interaction,medical diagnosis and psychological research.Recently,the appearance-based gaze estimation has gradually become the focus of research due to the advantages of simple system structure and operation.However,there will be large-scale head posture change in the application of gaze estimation.Lighting conditions and occlusion can also affect the quality of the captured image.These factors can lead to incorrect gaze estimation results.In order to achieve high-precision gaze estimation,three new gaze estimation algorithms based on deep convolutional networks are studied.The main research contents of this thesis are as follows:1.The position of the pupil center and eye center can determine the direction of the gaze.Learning the position of the pupil center can help improve the performance of the gaze estimation algorithm.Aiming at this point of view,a gaze estimation algorithm based on pupil positioning information is proposed.By adding the pupil center positioning auxiliary task to the gaze estimation network and adding the learned pupil position information as an additional feature to the feature vector,the network's accuracy is improved.2.Aiming at the problem that the existing gaze estimation algorithms are not robust to the head pose,a gaze estimation method based on head pose clustering is proposed.By head pose clustering,the gaze estimation task is divided into several simple subtasks.At the same time,in order to solve the problem of training data reduction after clustering,a loss function with class resolving ability is designed to train the common feature extraction network and the gaze estimation branch belonging to different head pose subclasses.This gaze estimation algorithm based on the head pose clustering idea can guarantee the feature extraction network's generalization ability and solve the large-scale head pose problem in gaze estimation.3.A gaze estimation algorithm based on face and eyeball context information is proposed.A gaze estimation network including three feature extraction branches is designed.The face and binocular features are extracted to directly estimate the gaze angle.Estimating the angle of gaze directly from the face image avoids the head pose step,thereby reducing the head pose estimation error.At the same time,the eye feature extraction branch is added to make full use of the eyeball deflection angle information contained in the eye part.In addition,for the inconsistency of picture quality caused by illumination and occlusion of both eyes,the binocular feature spectrum weight learning module is constructed to weight the binocular feature spectrum to further improve the accuracy of gaze estimation.The experimental results show that the proposed algorithm improves the accuracy of gaze estimation by adding pupil positioning task,head pose clustering and extracting face and binocular feature spectra.