Learning-Based Gaze Estimation And Its Application

The typical human-computer interaction(HCI)realizes the information transmission between the user and the machine through the direct touch operations.With the continuous development of multimedia and artificial intelligence,non-contact HCI brings user new experiences both in sensory and psychology.As an important branch of non-contact HCI,gaze interaction has gradually gained people's attention with its important values in real applications.What we called gaze interaction is the technique that estimates the gaze direction and then achieves the touching operations by eyeball movements.Therefore,it's crucial to estimate the gaze direction easily and accurately.The traditional gaze estimation methods estimate gaze direction through the positions of the pupil and cornea center.This method usually requires high-resolution cameras,complicated external equipment such as infrared light sources,and various sensor devices,etc.They're usually expensive,and inconvenience to the users.The deep learning has been greatly developed in the area of artificial intelligence by training neural networks through a large number of images and has gradually become the focus of the current gaze estimation researches.However,most of the current learning-based gaze estimation methods use regression methods to estimate the specific positions of the fixation points.Since the existence of saccadic eye movements,it is difficult to estimate the specific positions of the fixation points.The low accuracy and large errors can't meet the requirements of touch operations.In addition,different with the traditional model-based gaze estimation methods,learning-based gaze estimation methods require a large amount of data.So,building a suitable eye images database is crucial for relevant researches.In view of the above problems,taking into account the basic characteristics of touch operations,the learning-based gaze estimation method proposed in this paper uses classification as the main research point,and simulates the touch keys by building an appropriate screen block structure,and designs a corresponding human eye images database.Then,the experiment uses Convolutional Neural Network(CNN)to achieve eye image feature extraction and then to achieve the mapping from the image to the gaze direction.Based on the introduction above,this paper has made a series of improvements to existing methods in system architecture and experimental methods.The main innovations are mainly reflected in the following aspects:(1)Use binocular images as training data.In view of the fact that most of the learning-based gaze estimation methods use monocular images(only left-eye images or only right-eye images)as input,fully taking into account the differences in the dominant eyes for different individuals and the influence of the mutual position of the both eyes,this paper adopts binocular images as input.(2)Estimate the gaze blocks to simulate the touch buttons.Most of the currently learning-based methods use regression to estimate the specific fixation point locations,but this method exits large errors due to the presence of saccadic eye movements.The method of quantizing the images to different regions relaxes the estimation from the points to the blocks.This paper explores the accuracy of the gaze estimation from the two scales based on the computer screen size.(3)Gaze estimation based on classification task.In the touching operation,any point within the block can trigger the corresponding operation.It is assumed that the different gaze directions correspond to different blocks,which means that there is a mapping between the eye image and the block position.Any eye images falling into the same gaze block can be classified into the same category to trigger the corresponding button.At the same time,considering the superior classification performance of CNN,this paper uses the classification method to estimate the gaze direction.