| Eye movement trajectory usually contains information about a person’s attention and intention,so eye tracking has a wide range of applications in various fields,because the traditional method of measuring eye movement trajectories have high requirements for the measurement equipment and the environment,eye tracking is difficult to develop in many fields.In recent years,with the development of neural networks,it has become possible to realize eye tracking methods based on deep learning.In this paper,based on the deep learning eye tracking method,the object detection method is used to extract the face and eyes,and the gaze estimation network is designed to obtain gaze information according to the relevant information of the face and eyes.In order to improve the accuracy of the gaze estimation network,combined with the mainstream lightweight network and the generative adversarial network(GAN)further improve the model.The main contributions of this paper are as follows:(1)In order to improve the accuracy and efficiency of eye tracking,this paper builds a face and eye extraction model based on the YOLOv5 network.The model adds a residual structure,Focus structure and CSP structure to improve accuracy and reduce the amount of calculation.Experimental results show that the average precision of the network reaches 85.04%(m AP@0.5),and the recall rate reaches 85.08%.(2)Combining the Mobile Netv3 network and the Itacker network model,a new gaze estimation model,Mobile Tracker,is proposed.This model uses the bneck structure as the main structure,and builds the network according to the Itacker connection method.The weight of the input network of the two eyes is shared during feature extraction.The experimental results show that on the gazecapture dataset,the accuracy of the model is not significantly different from the Itacker model,but the amount of calculation is reduced.(3)In order to reduce the dependence of the model on the original data and improve the accuracy of gaze estimation,a new generative adversarial network model Mobile Deep Convolution Generative Adversarial Network(MDCGAN)was constructed based on the Mobile Net V3 network and incorporating gaze information into the loss function of the traditional generative adversarial network enables the training of the gaze estimation model.Experiments have shown that the new gaze estimation network based on GAN can achieve higher accurate,the gaze error is 5.89 degrees,which is 0.63 degrees less than the original method,and 0.12 degrees less than the same model that does not rely on the GAN network. |
