[Paper Review] Eye movement simulation and detector creation to reduce laborious parameter adjustments

Wolfgang Fuhl, Thiago Santini+4 more|arXiv (Cornell University)|Mar 28, 2018

Gaze Tracking and Assistive Technology49 references11 citations

TL;DR

This paper proposes a machine learning-based eye movement detection framework that uses a physics-informed simulator to generate synthetic velocity profiles for fixations, saccades, and smooth pursuits, enabling robust detector training independent of sampling rate variability. The method outperforms state-of-the-art algorithms in detection accuracy across multiple datasets, achieving balanced performance on all movement types without manual parameter tuning.

ABSTRACT

Eye movements hold information about human perception, intention and cognitive state. Various algorithms have been proposed to identify and distinguish eye movements, particularly fixations, saccades, and smooth pursuits. A major drawback of existing algorithms is that they rely on accurate and constant sampling rates, impeding straightforward adaptation to new movements such as micro saccades. We propose a novel eye movement simulator that i) probabilistically simulates saccade movements as gamma distributions considering different peak velocities and ii) models smooth pursuit onsets with the sigmoid function. This simulator is combined with a machine learning approach to create detectors for general and specific velocity profiles. Additionally, our approach is capable of using any sampling rate, even with fluctuations. The machine learning approach consists of different binary patterns combined using conditional distributions. The simulation is evaluated against publicly available real data using a squared error, and the detectors are evaluated against state-of-the-art algorithms.

Motivation & Objective

• To address the challenge of parameter sensitivity and lack of generalization in existing eye movement detection algorithms due to variable sampling rates and noise.
• To reduce reliance on manual annotation and heuristic thresholding by creating data-driven detectors trained on simulated eye movement profiles.
• To develop a flexible simulation framework capable of modeling saccades with gamma-distributed peak velocities and smooth pursuits with sigmoid onsets.
• To enable detector training for novel or rare eye movement types (e.g., skewed saccades) without requiring gold-standard annotated data.
• To create a system that supports both offline and online detection with potential for validation and refinement using known velocity profiles.

Proposed method

• The simulator generates synthetic eye movement velocity profiles using gamma distributions for saccades and sigmoid functions for smooth pursuit onsets, based on empirical oculomotor data.
• The system models variable and fluctuating sampling rates, allowing training on realistic, non-uniformly sampled data without requiring constant frame rates.
• Noise, sampling jitter, and tracker inaccuracies are simulated using configurable probability distributions to reflect real-world eye tracker limitations.
• A random ferns-based machine learning pipeline combines ten binary decision patterns with conditional probability distributions to learn robust detectors.
• Detectors are trained exclusively on simulated data and evaluated on real annotated datasets without post-processing or segmentation.
• The approach enables end-to-end training and evaluation without reliance on manual threshold tuning or data-specific initialization.

Experimental results

Research questions

• RQ1 Can a simulation-based approach generate realistic eye movement velocity profiles that accurately reflect biological oculomotor behavior?
• RQ2 Can machine learning detectors trained solely on simulated data achieve competitive performance on real-world eye tracking data across diverse sampling rates?
• RQ3 To what extent can the proposed method generalize to novel eye movement types (e.g., skewed saccades) without requiring manual annotation?
• RQ4 How does the performance of the proposed detector compare to state-of-the-art algorithms in terms of detection accuracy and balance across fixation, saccade, and pursuit types?
• RQ5 Can the system maintain robustness under variable and fluctuating sampling rates, a common challenge in real eye-tracking applications?

Key findings

• The proposed detector achieved a detection rate of 61.86% for fixations, 76.56% for saccades, and 61.88% for smooth pursuits on the GazeCom dataset, outperforming I-BDT and LS in balance and overall accuracy.
• On the I-BDT dataset, the proposed method achieved 72.50% for fixations, 69.44% for saccades, and 64.83% for smooth pursuits, demonstrating consistent performance across movement types.
• The detector maintained balanced performance across all movement types without requiring manual threshold tuning or data-specific initialization.
• Compared to EV (Nyström and Holmqvist, 2010), which fails to detect smooth pursuits entirely, the proposed method achieved significantly higher detection for all movement types.
• The method outperformed LS (Larsson et al., 2013) in detection accuracy for fixations and saccades on all evaluated datasets, despite LS being a self-adapting segmentation-based algorithm.
• The results suggest that the simulator-generated training data effectively captures the statistical properties of real eye movement velocity profiles, enabling high generalization to unseen real data.