Blinking Beyond EAR: A Stable Eyelid Angle Metric for Driver Drowsiness Detection and Data Augmentation

TL;DR

This work introduces Eyelid Angle ($ELA$), a 3D landmark–based metric that quantifies the relative orientation of the upper and lower eyelids to robustly track eye openness across camera viewpoints. By fitting eyelid planes from 3D landmarks and combining left/right eyes with a yaw-aware weighting, the authors extract temporally meaningful blink features and employ a $k$-NN classifier for drowsiness detection, while also generating synthetic data by animating a Blender avatar with $ELA$-driven blink dynamics. The approach is evaluated on public driver-monitoring datasets, showing improved viewpoint stability over EAR, reliable blink detection, and useful data augmentation capabilities via synthetic videos with controllable noise, viewpoints, and blink dynamics. The results underscore $ELA$ as a reliable biometric feature and a scalable tool for dataset expansion, with promising implications for real-time driver state monitoring and future multi-camera systems. The authors also propose to open-source the Blender-based data generation pipeline to promote reproducibility and further research.

Abstract

Detecting driver drowsiness reliably is crucial for enhancing road safety and supporting advanced driver assistance systems (ADAS). We introduce the Eyelid Angle (ELA), a novel, reproducible metric of eye openness derived from 3D facial landmarks. Unlike conventional binary eye state estimators or 2D measures, such as the Eye Aspect Ratio (EAR), the ELA provides a stable geometric description of eyelid motion that is robust to variations in camera angle. Using the ELA, we design a blink detection framework that extracts temporal characteristics, including the closing, closed, and reopening durations, which are shown to correlate with drowsiness levels. To address the scarcity and risk of collecting natural drowsiness data, we further leverage ELA signals to animate rigged avatars in Blender 3D, enabling the creation of realistic synthetic datasets with controllable noise, camera viewpoints, and blink dynamics. Experimental results in public driver monitoring datasets demonstrate that the ELA offers lower variance under viewpoint changes compared to EAR and achieves accurate blink detection. At the same time, synthetic augmentation expands the diversity of training data for drowsiness recognition. Our findings highlight the ELA as both a reliable biometric measure and a powerful tool for generating scalable datasets in driver state monitoring.

Figures (5)

• Figure 1: Overview of the proposed Eyelid Angle (ELA) framework: defining eyelid geometry, extracting blink dynamics, and generating synthetic datasets for drowsiness detection.
• Figure 2: 3D landmarks of the upper (green) and lower (red) eyelids with fitted planes. The relative orientation of these planes defines the ELA, providing a robust geometric representation of eyelid motion from noisy landmark data.
• Figure 3: A blink is defined as the interval between the last local maximum $\mathrm{ELA}_{\mathrm{start}}$ preceding the minimum derivative $m_1$ and the first local maximum $\mathrm{ELA}_{\mathrm{end}}$ following the maximum derivative $m_2$. The extrema of the derivative, $m_1$ and $m_2$, are classified using a $k$-means algorithm, as described in baccour_camera-based_2019. The intersections of the tangents (shown as red lines) define the temporal boundaries used to compute the closing, closed, and reopening durations of the blink. All features extracted from this process are summarized in Table \ref{['tab:blinkfeatures']}. The example shown in Fig. \ref{['fig:blinkfeatures']} was recorded using a standard webcam operating at $\qty{30}{\hertz}$.
• Figure 4: Comparison of the EAR and raw ELA values across varying head orientations relative to the camera, measured on two synthetic videos generated as described in Section \ref{['sec:synthbenchmarks']}. The vertical sweep ranges from below to above head level, and the horizontal sweep from right to left relative to the head. The eyelids were fixed at a constant angle of $\qty{60}{\degree}$. .
• Figure 5: Exemplary detection results and ground-truth labels from video 7 of the DMD dataset. $f_{\mathrm{det},i}$ denotes the frames classified as part of a detected blink, and $f_{\mathrm{gt},i}$ the frames labeled as blinks. Shown are one false negative, where $f_{\mathrm{det},2}$ does not overlap with any labeled blink, and two true positives corresponding to a slightly misleading label spanning two actual blinks. In this example, the subject squints, yawns, and blinks multiple times.