Gaze Point Estimation Algorithm Based On Head Pose Compensation

In recent years,researchers have paid more attention to gaze location technology.Gaze points contain a large amount of information about human perception of the objective world,such as emotional changes,attention,etc..Accurate estimation of gaze point is the key to obtain this information.In common image-based two-dimensional gaze localization algorithms,the accuracy of gaze location is largely affected by head pose.In this paper,the complex head pose is reduced to three angles in order to improve the accuracy of the point of gaze estimation,and a gaze point estimation algorithm based on the convolution neural network is proposed.The main contents are as follows:Firstly,the estimation of the head pose is mainly composed of two parts: the active appearance model and the proportional orthogonal iterative transformation algorithm.Section one,based on the model-based method,the active appearance model algorithm is used to train and fit the face feature points on the 2D image set.Locating and tracking face feature points in the fitting process of active appearance model real-time.Section two,the three-dimensional rigid model of the human head body is used to select the equivalent 3D point corresponding to the position of the face feature points of the active appearance model.The pose estimation of human head model is carried out by proportional orthogonal projection iterative transform algorithm.Secondly,the deep learning technique is used to study the face image from low-level to high-level features by using convolution kernel,weight sharing and down-sampling of convolution neural network.The output of the head pose network is optimized by the activation function.Multiple feedback calculation reduce network error.Through the convolution neural network training results effectively reduce the pose angle error,and in the expression,light and other environments with strong robustness.Thirdly,after the head pose was obtained,the pupil center cornea reflex method was used to evaluate the gaze point in the infrared light source environment,and the head pose parameters were added to correct the model.Finally,compared the accuracy of the experiment before and after the compensation.Experiments show that the algorithm can estimate the head pose more accurately,and after adding the compensation algorithm,the accuracy of the gaze points estimation is improved obviously,and the positioning accuracy of gaze tracking system can be satisfied.