Research On Deep Learning Based Eye Tracking Technology

Eye-movement based HCI is commonly used in behavioral analysis,medical assistance,and intelligent systems.The key technologies for Eye-movement based HCI are eye tracking and eye-movement behavior recognition.With the advent of artificial intelligence,eye-tracking and eye-recognition technologies have applied deep learning to enable the appropriate tasks to be performed without an eye tracker.In this thesis,we focus on the study of convolutional neural network-based eye tracking techniques and eye-movement behavior recognition.The main tasks are as follows.1.Research on convolutional Neural Network(CNN)based eye tracking technology.CNN-based eye tracking includes person-independent eye tracking and person-specific eye tracking.Person-specific eye tracking usually fine-tunes the model to correct bias and requires individual samples.When the sample size of individuals is limited,limited samples needs to be utilized to effectively reduce eye tracking errors.We addressed the question of how to select valid samples.First,we use Res Net as a generic eye tracking model,and then we use the model as a tool for extracting features from individual samples and regressing the features using SVR to obtain the average error of different samples.We use Unity3 D to make a sample acquisition program to capture images data as user gazes at calibration targets,which change the position distribution of calibration targets to obtain different samples.We set up two modes of distribution,each with a different number of targets,with the number of targets corresponding to the number of individual samples.The experiment shows that eye tracking errors decreased as the number of individual samples increased;for the same number of calibration targets,the average error for samples where calibration targets are distributed according to a regular pattern is 0.2 to 0.4 cm lower than for samples where calibration targets are distributed according to a random pattern.2.Gaze-gesture recognition research.Gaze-gesture recognition belongs to eye-movement behavior recognition.The purpose of eye-movement behavior recognition is to distinguish different types of eye movements based on gaze information.Eye movement behaviors in eye-movement interaction are categorized into fixation,blinking,saccade,and smooth pursuit.Interaction using saccades is gaze-gesture interaction.Usually gaze data is collected by an eye tracker,and then the gaze data is analyzed and classified.In this thesis,we use an appearance-based gaze gesture recognition method using convolutional neural networks and long short-term memory networks(LSTM).LSTM was good at recognizing single-stroke gaze gestures,but not so good at recognizing multi-stroke gaze gestures.To address this problem,we used Bi LSTM instead of the LSTM in the model to identify the multi-stroke gaze gestures and increases the classification accuracy by 18.42%.