Research On Eye-tracking Based On Deep Learning

With the continuous development of virtual reality,many research scholars have studied virtual reality as a hot topic.Based on the current research status of virtual reality,eye tracking has become an important technical field of virtual reality,and eye tracking is applied in virtual reality scenes.To ensure that the accuracy is not reduced,the image processing speed in the scene and the reality in the scene are improved,the performance consumption is reduced,and the user's sense of vertigo and fatigue are reduced.In the development of eye-tracking,the performance of deep learning has achieved excellent results.Deep learning divides eye-tracking into detection in single-frame images and tracking in multiple-frame images.In the deep-learning single-frame image detection algorithm,the end-to-end YOLO detection algorithm achieves an accuracy of 50,but from the speed point of view,the YOLO algorithm has a processing speed of only 45 FPS in the virtual reality,and the user will appear to be stuck in the picture.The situation occurred.This thesis addresses the above issues:(1)Improved the detection algorithm YOLO for single-frame images by merging the feature maps corresponding to each convolutional layer to obtain more obvious feature information,and improved the YOLO application network in the improved network.All convolutional layer feature maps are predicted and the final eyeball position information is obtained by using border regression and other training methods.(2)The combination of the YOLO algorithm and the recursive neural network of multi-frame images,in the processing of multi-frame images,the spatial correlation of before and after the information is larger,through the combination of YOLO improved algorithm and recursive neural network,before and after the frame image The feature information is used for spatial association learning.When the eyeball is occluded by external factors,the eyeball position information is predicted using the confidence maps of the five frames of image information before and after.(3)Application of eye-tracking in virtual reality,associating eye-tracking with radial blur rendering,and applying it to virtual reality scenes,aiming at precise real-time rendering of users with low hardware conditions of PC or mobile electronic devices Functions(accuracy 60 or more,FPS60 or more),reduce performance consumption,improve the authenticity of the scene.In the simulation experiment in this thesis,the improvement of the YOLO detection algorithm of single-frame image compared with the traditional YOLO algorithm,the precision increased by 10%,the speed can reach 87 FPS.The combination of YOLO and recursive neural networks for multi-frame images achieves a precision of 78.2% and a speed of 17 ms.In the experiment of eye-tracking based on radial blur scene rendering,the accuracy reaches 63,FPS can reach 66,there will be no card frame and frame dropping situation,and the expected effect is satisfied.