A Differential 2D Gaussian Ellipse-Based Eye Movement Analysis

Zhang, Jiahao; Liu, Yang; Wen, Xu; He, Dengbo · 2025 · Proceedings of the Human Factors and Ergonomics Society Annual Meeting

DOI: 10.1177/10711813251358252

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of noise and limited information in raw gaze-point data from eye-tracking devices, which hinders accurate modeling of driver visual attention. While visual behavior is critical for driving safety and the development of adaptive human-machine interfaces, discrete gaze points often fail to capture the full scope of visual perception due to device inaccuracies and rapid, unconscious eye movements. To resolve this, the authors propose the Differential 2D Gaussian Ellipse (D2DGE) representation. This method captures gaze distribution within a sliding time window by calculating the k-sigma range of a 2D Gaussian distribution fitted to gaze points, resulting in an ellipse defined by five parameters: mean coordinates, area, major axis length, and rotation angle. This approach aims to reduce noise while preserving the spatial distribution of visual attention. To validate D2DGE, the researchers conducted a driving simulator experiment with 12 participants, divided equally into novice and experienced drivers. Participants viewed 87 driving scenarios derived from the Deepaccident dataset, with eye-tracking data recorded at 60Hz. The study employed two primary analytical methods. First, a Generative Adversarial Imitation Learning (GAIL) model was trained on both raw gaze data and D2DGE data to generate visual attention patterns. The similarity between generated and original data was quantified using Kullback-Leibler (KL) divergence. Second, linear mixed-effects models and one-way ANOVA tests were used to compare the five D2DGE parameters and raw gaze coordinates between novice and experienced drivers, assessing whether the new representation revealed differences in visual scanning behaviors that raw data missed. The results demonstrated that D2DGE data provided a superior approximation of raw gaze distributions compared to raw data alone. KL divergence values were significantly lower for D2DGE-generated data than for raw gaze-generated data for both driver groups, indicating that D2DGE preserves spatial distribution characteristics more effectively. Furthermore, statistical analysis revealed that while raw gaze coordinates and most D2DGE parameters (mean coordinates, major axis length, rotation angle) showed no significant differences between novice and experienced drivers, the D2DGE ellipse area did differ significantly (p = .02). Novice drivers exhibited a larger ellipse area, indicating a broader scanning region, whereas experienced drivers maintained a narrower observational field, likely due to quicker identification of potential hazards. The significance of this work lies in establishing D2DGE as a robust alternative to raw gaze points for data-driven visual attention analysis. By capturing richer information about gaze distribution, D2DGE enhances the performance of predictive algorithms and reveals behavioral distinctions between driver groups that discrete points obscure. The findings suggest that D2DGE can improve the design of adaptive driver assistance systems and autonomous vehicle interfaces. The authors note limitations regarding parameter calibration and scenario diversity, recommending future validation with on-road data and more advanced generative models to further refine the representation’s application in understanding driver behavior.

Key finding

The Differential 2D Gaussian Ellipse representation reduces noise in eye-tracking data and captures richer visual attention information than raw gaze points, revealing significant differences in scanning area between novice and experienced drivers.

Methodology

simulator

Sample size: 12

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• gaze based attention detection
• eye movements scanning
• attention allocation

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: behavioral performance data
• Methodological Resource: tool software, measurement protocol