Probabilistic approach to robust wearable gaze tracking

Abstract

This paper presents a method for computing the gaze point using camera data captured with a wearable gaze tracking device. The method utilizes a physical model of the human eye, advanced Bayesian computer vision algorithms, and Kalman filtering, resulting in high accuracy and low noise. Our C++ implementation can process camera streams with 30 frames per second in realtime. The performance of the system is validated in an exhaustive experimental setup with 19 participants, using a self-made device. Due to the used eye model and binocular cameras, the system is accurate for all distances and invariant to device movement. We also test our system against a best-in-class commercial device which is outperformed for spatial accuracy and precision. The software and hardware instructions as well as the experimental data are published as open source.

Keywords

• Wearable gaze tracking
• Human eye modeling
• Bayesian modeling
• Kalman filtering