Dual-infrared glint geometry-constrained gaze estimation for low-cost head-mounted monocular eye tracking

Abstract

Head-mounted eye tracking has significant applications in areas such as human–computer interaction and virtual reality. To address the underutilization of dual-infrared glint information and calibration feature jitter in low-cost monocular near-eye systems, this paper proposes Dual-Glint Geometry-Constrained Gaze Estimation (DG²C). Specifically, a YOLO-based detector is employed to localize the pupil center and dual corneal glints, from which a six-dimensional hybrid geometric feature (hybrid-crx) is constructed. A ridge-regularized affine regression model is then used to map eye features to screen gaze points, while a calibration quality screening mechanism based on intra-frame feature dispersion is introduced to improve calibration robustness. Results from four repeated experiments demonstrate that DG²C consistently outperforms baseline methods across all evaluation metrics, achieving a pixel error of 31.49 px, an angular error of 0.76°, and an intersession coefficient of variation of 19.8%. Furthermore, ablation studies confirm the effectiveness of the proposed quality screening mechanism.