High-resolution oculomotor measurements via a digital Dual Purkinje Image eye-tracker

Ruei-Jr Wu, Paul Jolly, Soma Mizobuchi, Ashley Clark, Zhetuo Zhao, Bin Yang, Janis Intoy, Michele A. Cox, Michele Rucci · Journal of Vision · 2022

Precisely measuring eye movements and accurately localizing gaze are vital to understanding how humans acquire and process visual information. Here we present recent advances using the Dual Purkinje Image (DPI) approach, an eye-tracking method originally developed by Cornsweet and Crane in 1973. This method enhances resolution by measuring relative, rather than absolute, motion of reflections generated by surfaces at different distances from the eye rotation axis, specifically the cornea and the back of the lens (first and fourth Purkinje images, P1 and P4, respectively). These images move similarly under translation but differentially under rotations, facilitating rejection of translation artifacts. DPI methods have traditionally been implemented in cumbersome and fragile analog devices, which have remained restricted to specialized laboratories. However, recent advances in imaging technology and computational power now allow the DPI approach to be enriched with the flexibility and intelligence of digital systems and implemented in more compact and practical devices. We describe oculomotor measurements obtained with a digital DPI, which directly images the P1 and P4 reflections of an infrared beam by means of a high-speed camera. A real-time algorithm running on a dedicated GPU extracts the Purkinje images and uses them as a metric for eye rotation. Experimental results with both artificial and real eyes demonstrate that this system can resolve sub-arcminute eye rotations at 1Khz sampling rate. We show that the main DPI artifact, the post-saccadic movement of the lens, is highly predictable. The lens rebound lasts less than 20 ms for saccades of all sizes and has magnitude equal to ~14% of the saccade amplitude for saccades larger than 2o. We further show that the DPI approach can lead to arcminute-accuracy localization of the line of sight and can be implemented in a head-mounted device with minimal loss of resolution.

Read the paper · More papers on PaperTik