Abstract

Aim/Background: Drowsy driving is a leading cause of traffic accidents, with 139,258 incidents recorded in Indonesia in 2022, according to Badan Pusat Statistik (Central Bureau of Statistics). Fatigue reduces a driver’s attention, reaction time, and decision-making, increasing accident risks. Camera-based drowsiness detection has been developed, but their effectiveness is often hindered by varying lighting conditions. This study evaluates the impact of image enhancement techniques on drowsiness detection performance under different lighting intensities and directions to improve real-time detection accuracy and reliability.

Methods: In this study, the dataset comprises human facial images focusing on the eyes and mouth, categorized into "microsleep" and "normal" states. Four image enhancement methods—CLAHE, LIME, Single-Scale Retinex, and Multi-Scale Retinex—are applied. Haar Cascade Classifier detects facial features, Canny Edge Detection extracts edges, and Histogram of Oriented Gradients (HOG) represents features. A Support Vector Machine (SVM) performs classification, evaluated through accuracy, precision, recall, F1-score, and detection time.

Results: Results indicate that Single-Scale Retinex with an intensity variation of 45 provides optimal performance. Morning tests achieved 88% accuracy, 90% precision, 88% recall, 89% F1-score, and a detection time of 0.183 seconds, while afternoon tests recorded 90% accuracy, 92% precision, 90% recall, 91% F1-score, and 0.109-second detection time.

Conclusion: The study confirms that Single-Scale Retinex effectively enhances drowsiness detection under varying lighting conditions, improving accuracy and speed. This advancement enhances the reliability of real-time detection, contributing to reduced traffic accidents and improved road safety.

Key words: Drowsiness Detection, Single-Scale Retinex, Canny Edge Detection, Histogram of Oriented Gradients, Principal Component Analysis, Support Vector Machine