Dynamic accommodation measurement using Purkinje reflections and machine learning

Quantifying eye movement is important for diagnosing various neurological and ocular diseases as well as AR/VR displays. We developed a simple setup for real-time dynamic gaze tracking and accommodation measurements based on Purkinje reflections, which are the reflections from front and back surfaces of the cornea and the eye lens. We used an accurate eye model in ZEMAX to simulate the Purkinje reflection positions at different focus distances of the eye, which matched the experimental data. A neural network was trained to simultaneously predict vergence and accommodation using data collected from 9 subjects as well as synthetic data generated using the eye model. We demonstrated that the use of Purkinje reflection coordinates together with pupil center and shape in machine learning improves the accuracy of the prediction. The proposed system accurately predicted the accommodation with an accuracy better than 0.21 D using two-point calibration data in tests performed with 9 subjects in our setup.