[Paper Review] Driver Gaze Estimation Without Using Eye Movement

Lex Fridman, Philipp Langhans+2 more|arXiv (Cornell University)|Jul 16, 2015

Gaze Tracking and Assistive Technology14 references12 citations

TL;DR

This paper proposes a real-time driver gaze estimation system that uses facial feature detection and head pose classification instead of direct eye tracking. By classifying spatial configurations of facial features into six gaze regions, it achieves 91.4% accuracy at 11 Hz on a 50-driver on-road dataset, offering a robust alternative to traditional eye tracking under challenging visual conditions.

ABSTRACT

Abstract—Automated estimation of the allocation of adriver’s visual attention may be a critical component offuture Advanced Driver Assistance Systems. In theory,vision-based tracking of the eye can provide a good estimateof gaze location. In practice, eye tracking from videois challenging because of sunglasses, eyeglass reflections,lighting conditions, occlusions, motion blur, and otherfactors. Estimation of head pose, on the other hand, isrobust to many of these effects, but cannot provide as fine-grained of a resolution in localizing the gaze. However,for the purpose of keeping the driver safe, it is sufficientto partition gaze into regions. In this effort, we proposea system that extracts facial features and classifies theirspatial configuration into six regions in real-time. Ourproposed method achieves an average accuracy of 91.4%at an average decision rate of 11 Hz on a dataset of 50drivers from an on-road study.Index Terms—Head pose estimation, gaze tracking, driverdistraction, driver assistance systems, on-road study.

Motivation & Objective

• To develop a robust driver gaze estimation system that does not rely on eye movement tracking.
• To address the limitations of traditional eye tracking, such as reflections from glasses and poor lighting.
• To improve safety in Advanced Driver Assistance Systems by estimating visual attention distribution.
• To classify gaze into discrete regions with high accuracy using only facial feature configurations.
• To achieve real-time performance suitable for on-road deployment.

Proposed method

• The system detects key facial features using a real-time facial landmark detection algorithm.
• It classifies the spatial configuration of these facial features into six predefined gaze regions.
• Head pose estimation is used as a proxy for gaze localization, avoiding direct eye tracking.
• The method leverages geometric relationships between facial landmarks to infer gaze direction.
• A machine learning classifier is trained to map facial feature configurations to gaze regions.
• The system operates at an average decision rate of 11 Hz, enabling real-time performance.

Experimental results

Research questions

• RQ1 Can facial feature configuration be reliably used to estimate driver gaze without eye tracking?
• RQ2 How accurately can gaze be classified into discrete regions using only head pose and facial landmarks?
• RQ3 What is the performance of the system under real-world on-road conditions with visual obstructions?
• RQ4 Can the system maintain real-time operation while achieving high accuracy in gaze region classification?
• RQ5 How does the method compare to traditional eye tracking in terms of robustness to environmental factors?

Key findings

• The proposed system achieves an average accuracy of 91.4% in classifying driver gaze into six regions.
• The system operates at an average decision rate of 11 Hz, suitable for real-time applications.
• The method demonstrates robustness to common challenges in eye tracking, such as eyeglass reflections and lighting variations.
• Gaze estimation accuracy is maintained across a diverse set of 50 drivers in an on-road study.
• The approach effectively replaces direct eye tracking by using facial feature geometry and head pose for gaze region classification.