Improving model-based mobile gaze tracking

Abstract

Mobile,wearable gaze tracking provides flexible opportunities for extending gaze tracking research outside of laboratory environments. Wearable trackers are predominantly video based and fall in to two categories: pupil-corneal reflection methods and physical model-based methods. A number of error sources affect the feature extraction and gaze mapping and therefore the accuracy and precision of both systems. Here, we present two methods for improving tracking results: an advanced user calibration procedure for estimating gaze vectors applicable with any modelbased method and a Bayesian tracker for tracking any number of corneal reflections and the pupil center, applicable with both types of trackers. The results show clear improvements over the stability and robustness of recognizing and tracking features in the eye image and, ultimately, estimating the gaze vector.