Research On Gaze Tracking Method

Eye tracking is a technology that can track the movement trajectory of the human eye during the viewing process.It can be applied in multiple fields,such as psychology,human-computer interaction,and medicine.In eye tracking,special devices are used to record the movement trajectory of the eyes,including eye trackers and infrared trackers.The application of line of sight tracking technology is very extensive,one of which is in the field of psychology.In psychological research,eye tracking can be used to study people’s attention and attention bias,and can help researchers better understand people’s thinking processes.In addition,in the field of human-computer interaction,line of sight tracking can be used to evaluate the user’s effectiveness and ease of use of the interface,thereby helping designers improve product design.In the medical field,eye tracking can be used to help diagnose some eye diseases and help doctors better understand patients’ visual problems.The development of eye tracking technology is also very rapid,and many new types of eye tracking devices have emerged,such as wearable devices and virtual reality devices.The emergence of these devices has made the application of line of sight tracking technology more convenient and widespread.Eye tracking technology is a very important technology that can help us better understand human cognitive processes and behaviors,and can be applied in multiple fields,including psychology,human-computer interaction,and medicine.In the future,the application of line of sight tracking technology will be more widespread,and more new devices and technologies will also emerge,which will bring more opportunities and chalenges to the development of line of sight tracking technology.The main work of this article is as follows:(1)Construction of a fine-grained gaze estimation modelLine of sight direction is mainly determined by both eye sight direction and head posture.In selecting the network,this paper finds that the extraction of the line of sight estimation features is very difficult,especially for the eye images,where the difference in angle is small and the head pose has not changed,the difference in features is very small.Facing such a small feature difference,how to design an efficient real-time network becomes the primary problem solved in this paper.Deep learning is divided into two categories: classification tasks and regression tasks.The existing line of sight estimation directly uses regression to map the features directly.This approach appears to be too violent.In this paper,we design a fine-grained line-of-sight estimation model,which shifts the regression task to the form of classification followed by regression.The core point of this design is to improve the accuracy of classification,and the accuracy of regression will be improved accordingly.On the publicly available dataset,not only the computation time and cost are significantly reduced,but also the accuracy is improved to a certain extent compared with previous models,which is used as the benchmark model in this paper.(2)Design of fine-grained gaze estimation model based on one-dimensional Gaussian distributionIt is found that eye images are labeled with certain ambiguity due to small angle changes or unintentional eye closures.In order to suppress the blurriness of the line-ofsight labels in the unconstrained environment.Through the feature analysis and study of a large number of eye images in the experimental environment,it is found that the eye images between different labels imply certain similarity,and the similarity between each class of eye images and other images is found to follow a Gaussian distribution after a certain ranking.Based on the above research,this paper designs a one-dimensional Gaussian distribution instead of the original hard labels,and constructs a one-dimensional Gaussian distribution line of sight estimation model with fine-grained as the basic architecture,which is supervised and trained with a compound loss function.The results on three open datasets show that the one-dimensional line-of-sight estimation model proposed in this paper has a certain degree of experimental improvement with previous line-of-sight estimation models including the fine-grained line-of-sight estimation model.(3)Design of fine-grained gaze estimation model based on two-dimensional Gaussian distributionFurther analysis reveals that the one-dimensional Gaussian distribution line-of-sight estimation model is relatively single in terms of feature consideration,mainly in the selection of similarity features for calculation,considering only one dimension such as the change in pitch dimension,without restriction on the other yaw dimension,which also leads to a less accurate distribution construction.To solve this problem,this paper proposes a two-dimensional Gaussian distribution line of sight estimation model to construct a distribution from two dimensions jointly,instead of separating the two.This model continues to follow the design of the benchmark model architecture with an improved coarse classification module,and still supervises the training iterations of the whole network with a compound loss function.Experimental results on publicly available datasets commonly used in the field of line of sight estimation show that the twodimensional line of sight estimation model proposed in this paper is somewhat more effective than previous line of sight estimation models,including the one-dimensional line of sight estimation model.(4)Design of fine-grained gaze estimation model based on adaptive distributionRegardless of the one-dimensional Gaussian distribution line of sight estimation or two-dimensional Gaussian distribution line of sight estimation,the constructed distribution tries to fit the analyzed similarity distribution as much as possible,however,the distribution of data is very complex,sometimes it is a standard distribution,sometimes it is other uninvolved distribution,how to reasonably construct a prior distribution that compounds the real data distribution? In this paper,we propose adaptive distribution to solve this problem.The adaptive distribution uses deep learning to learn the features of the sample,and according to the output of the distribution of the corresponding image category,in order to strengthen this distribution,this paper adds the original hard labels on top of it.The model still follows the fine-grained view estimation model architecture,adding the design of adaptive distribution inside the coarse classification,and the fine classification module follows the original MSE as the supervision,and the whole model is supervised by the compound loss function for training iterations.Experimental results on public datasets commonly used in the field of line-of-sight estimation show that the adaptive distributed line-of-sight estimation model proposed in this paper has a certain degree of improvement over previous line-of-sight estimation models,including the 2D line-of-sight estimation model.