Research On Pupil Center Location And Tracking Based On Near-Eye Infrared Video Images

Pupil center positioning and tracking technology has always been one of the important research topics in the field of computer vision and pattern recognition.Pupil center positioning and tracking technology is widely used,involving multiple disciplines such as medical detection,human-computer interaction,psychology,virtual reality,and so on.It is very meaningful to study how to quickly,stably,and accurately locate the pupil center.This paper focuses on the problem of pupil center localization and tracking in near-eye infrared video images.The main research contents include the following aspects:(1)Analyze pupil images and select appropriate cameras,lenses,and light sources for the pupil image acquisition hardware system based on the "dark pupil" characteristics of the pupils.We analyzed the Tenengrad image clarity evaluation algorithms for four different template kernels: 3X3(default size),5X5,7X7,and 9X9.The experimental results showed that as the template kernel increased,the Tenengrad algorithm’s sensitivity to noise decreased,but the algorithm took more time.Considering the algorithm’s noise resistance and real-time performance,we applied the Tenengrad algorithm of the 5X5 template kernel to focus when collecting pupil images,This ensures the clarity of the collected pupil images.(2)Aiming at the factors that affect the accurate positioning of the pupil center,such as eyelash occlusion,eyelid occlusion,stray light interference,and white spot interference caused by corneal reflection,a method for locating the pupil center from "coarse" to "fine" is proposed.The YOLOv5 n model is used to coarsely locate the pupil region.In order to improve the average accuracy of the YOLOv5 n model,the regression loss function CIOU is replaced by EIOU,and a k-means++algorithm is applied to calculate a priori anchor frame.The optimized model is applied to the pupil data set m AP@0.9 97.61%,1.47% higher than before optimization;In order to reduce the interference of white spots,eyelids,and eyelashes,a dynamic pupil image clipping method based on a circular region is proposed,leaving only the image at the edge of the pupil;Using the Canny operator to extract pupil subpixel edges,an adaptive Canny gradient threshold selection algorithm based on grayscale histogram and Otsu is proposed to address the differences in the grayscale gradient at the junction between the pupil and the background under different light intensities and the dependence of the Canny edge detection operator on the selection of artificial thresholds.At the same time,in order to improve operational efficiency,the image is subjected to two down-sampling processes before threshold selection;Finally,an ellipse fitting algorithm based on Tukey weight and L-BFGS-B is proposed to fit the pupil edge,thereby achieving accurate positioning of the pupil center.Experiments were conducted on the collected pupil data set,and the proposed pupil center positioning method had a positioning error of within 5 pixels,with an average positioning error of 0.696 pixels,meeting the requirements for pupil center positioning error,with an average positioning time of 31.3 ms.The experimental results show that the pupil center positioning method has good robustness,accuracy,and real-time performance.(3)Aiming at the shortcomings of traditional KCF tracking algorithm in pupil tracking,such as fixed scale and too frequent template updates,resulting in reduced template accuracy,the traditional KCF algorithm is improved.Using FHOG feature descriptors to describe pupil targets;Aiming at the scale changes of pupil targets during tracking,a pupil target scale estimation based on an improved seed region growth algorithm is proposed.Compared with the original algorithm,the improved seed region growth algorithm has a significant improvement in speed,and the larger the target region,the more obvious the advantages of the improved algorithm;Aiming at the situation where template updates are too frequent and lead to template degradation,a template update strategy based on APCE is proposed to effectively ensure the accuracy of the template.Experimental results show that the improved KCF algorithm has better pupil tracking performance and higher tracking accuracy compared to the traditional KCF algorithm.(4)Aiming at pupil center location and tracking based on near-eye infrared video images,a pupil center location and tracking method based on YOLOv5 n and KCF is proposed,and experimental verification is conducted.The experimental results show that the method has good robustness,accuracy,and real-time tracking performance.