Research On Head Free Remote Pupillometer Based On Deep Learning And Binocular Vision

Changes in pupil diameter are directly related to the activity of the nervous system.It can reflect many important human information states and is used in various fields such as fatigue driving,disease diagnosis and human-computer interaction.In recent years,with the acceleration of hardware processing speed and the rapid development of computer vision algorithms,pupillometers have become more and more widely used in entertainment,transportation and medical fields.The remote head free pupil diameter measurement system has the characteristics of non-contact,fast and so on,and has broad research and application prospects.However,current remote head free pupillometers are susceptible to ambient light interference and human eye tracking.Slow speeds and distortion errors are measured while measuring the diameter of the pupil and require a cumbersome calibration procedure before use.In this paper,a binocular vision measurement model based on deep learning is established for these problems,which improves the measurement speed and accuracy of the system and improves the system’s ability to resist ambient light interference without calibration before each use.The main research contents and results are as follows:First,an experimental platform was established for the pupil diameter measurement system.When the staff is working normally,we run the experimental platform and collect image data of the person.Hardware and software selection are based on working distance,working angle and processing speed.The hardware mainly includes industrial cameras,near-infrared light sources and workstations.The software mainly includes operating system,programming language and compiling software.Secondly,aiming at the problem of ambient light interference,an improved YOLO V2 deep learning near-infrared human eye detection algorithm is proposed.Based on the idea of SqueezeNet,a lightweight feature extraction network is designed,and the non-maximum suppression is replaced by non-absolute-maximum suppression.The experimental results of the actual data set show the effectiveness of the method,and the accuracy of near-infrared human eye detection can reach 95.8%.Then,by exploring the similarity of binocular camera images,we propose a master-slave structure acceleration algorithm for binocular camera human eye tracking.While maintaining fairly high accuracy,it surpasses the tracking speed of traditional parallel structures(<11 ms per frame).Finally,a pupil diameter estimation algorithm based on binocular vision was developed.Compared to a pupil meter that previously used a 2D projection image on a single camera to cause distortion errors in pupil diameter measurement,our system measures the pupil diameter in 3D space.There is no distortion and there is no need to calibrate before each use.The experimental results on the actual data set show the effectiveness of the proposed method.The average absolute error of the pupil diameter is estimated to be(0.0216±0.0168)mm,and the average absolute error percentage of the pupil diameter is(0.57±0.67)%,and the change of angle and distance is not sensitive.