Purchase this article with an account.
Kamran Binaee, Rakshit Kothari, Flip Phillips, Gabriel Diaz; Characterization and Calibration of Eye Tracking Data from Head Mounted Displays. Journal of Vision 2016;16(12):846. doi: https://doi.org/10.1167/16.12.846.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Eye tracking data can suffer from multiple sources of error. This is especially true for trackers integrated within virtual reality (VR) head mounted displays (HMDs). During an immersive VR experience, the quality of the data diminishes due to static and dynamic error sources. Static tracking errors arise owing to factors such as optical aberrations and display optical characteristics. These are further exacerbated by other non-linear, in-engine distortions - typically introduced to correct the projected imagery for the aforementioned distortions. While these corrections make the scene perceptually more veridical, they frequently lead to distortion in the tracker data. For an active subject in VR, the quality of the tracking data is continuously affected by physical shifts of the HMD on the observer's head. These shifts can vary from slow, slippage-related drift over the course of an experiment to paroxysmal bumps and jolts caused by contact or the inertial characteristics of the HMD/tracker combination. It is common practice for calibration procedures to establish eye-to-screen mappings, typically at the beginning of a session and fixed over the timecourse of an experiment. However, in an HMD configuration, even with periodic recalibration, data quality can quickly degrade due to the dynamic issues mentioned. Here we present a novel calibration method that corrects for static as well as dynamic distortions. We combine a static, non-linear correction with a dynamic, temporally adjusted linear correction. By incrementally sampling a small subset of ground-truth space throughout the experiment, we can continuously determine data quality as well as perform temporally specific calibration. Over the timecourse of the experiment, our static method reduces angular error by more than 50% compared to the manufacturer-provided default static calibration. Adding dynamic compensation reduces error by as much as 80%, overall. We outline our method as well as several demonstrations of real-world correction in experimental conditions.
Meeting abstract presented at VSS 2016
This PDF is available to Subscribers Only