[Paper Review] Driver Drowsiness Classification Based on Eye Blink and Head Movement Features Using the k-NN Algorithm

Mariella Dreißig, Mohamed Hedi Baccour+2 more|arXiv (Cornell University)|Sep 28, 2020

Sleep and Work-Related Fatigue37 references4 citations

TL;DR

This study proposes a k-NN-based driver drowsiness classification system using eye blink and head movement features extracted from driver monitoring camera data. It achieves 84.2% balanced accuracy in binary classification and identifies key behavioral features—such as PERCLOS2, eyelid cleft, and head bobbing—that strongly correlate with drowsiness, enabling more robust real-time monitoring systems.

ABSTRACT

Modern advanced driver-assistance systems analyze the driving performance to gather information about the driver's state. Such systems are able, for example, to detect signs of drowsiness by evaluating the steering or lane keeping behavior and to alert the driver when the drowsiness state reaches a critical level. However, these kinds of systems have no access to direct cues about the driver's state. Hence, the aim of this work is to extend the driver drowsiness detection in vehicles using signals of a driver monitoring camera. For this purpose, 35 features related to the driver's eye blinking behavior and head movements are extracted in driving simulator experiments. Based on that large dataset, we developed and evaluated a feature selection method based on the k-Nearest Neighbor algorithm for the driver's state classification. A concluding analysis of the best performing feature sets yields valuable insights about the influence of drowsiness on the driver's blink behavior and head movements. These findings will help in the future development of robust and reliable driver drowsiness monitoring systems to prevent fatigue-induced accidents.

Motivation & Objective

• To enhance driver drowsiness detection in vehicles by leveraging non-invasive signals from a driver monitoring camera.
• To identify and evaluate a comprehensive set of 35 blink and head movement features indicative of drowsiness.
• To develop and optimize a k-Nearest Neighbor (k-NN) classifier for accurate driver state classification using feature selection.
• To provide insights into how drowsiness affects blink behavior and head movements to inform future robust monitoring systems.

Proposed method

• Collected driving simulator data from participants under awake, questionable, and drowsy states to generate a dataset of 35 behavioral features.
• Extracted features including blink duration, opening/reopening duration, PERCLOS1/PERCLOS2, head bobbing, eyelid cleft, and amplitude-velocity ratios.
• Applied wrapper-based feature selection with k-NN to identify optimal feature subsets for binary and multiclass classification.
• Used one-vs-one (OvO) classification strategy to improve multiclass performance and gain insights into class separability.
• Optimized k-NN model by tuning the value of k and selecting appropriate distance metrics through cross-validation.
• Evaluated model performance using balanced accuracy and analyzed feature importance across different classification tasks.

Experimental results

Research questions

• RQ1 Which combinations of blink and head movement features most effectively distinguish between awake, questionable, and drowsy driver states?
• RQ2 How does the k-NN algorithm perform in classifying driver drowsiness when using a large set of behavioral features?
• RQ3 Can feature selection improve the accuracy and robustness of k-NN-based drowsiness classification?
• RQ4 What are the most discriminative features for separating drowsy states from alert states, and how do they vary across classification strategies?
• RQ5 Does the one-vs-one (OvO) classification approach yield better performance and deeper insights than standard multiclass k-NN for this problem?

Key findings

• The k-NN classifier achieved a balanced validation accuracy of 84.2% in binary classification (awake vs. drowsy), demonstrating strong performance with the selected feature set.
• The multiclass classification achieved a balanced accuracy of 70.0%, indicating challenges in distinguishing between three distinct states with the standard k-NN setup.
• The OvO classification approach yielded slightly better results than standard multiclass classification, with improved sensitivity and class-specific insights.
• Features such as PERCLOS2, eyelid cleft, head bobbing, and reopening duration were consistently selected across optimal feature subsets, indicating their high discriminative power.
• The model revealed that eyelid cleft and reopening duration are particularly effective in distinguishing the awake state from drowsy or questionable states.
• Smaller, compact k-NN models with fewer features and lower k values outperformed more complex models, suggesting that simplicity enhances generalization in this context.