Low-light driver drowsiness detection for real-time safety assistance using dual attention mechanisms in deep learning model.

Saxena S; Angel; Khurana M; Tiwari S; Shakya, HK · 2026 · PubMed Central

DOI: 10.1038/s41598-026-44442-3

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical safety issue of driver drowsiness, particularly under challenging real-world conditions such as low-light environments and facial occlusions (e.g., glasses, masks). Motivated by the high incidence of fatigue-related accidents, especially in regions with erratic traffic and inconsistent infrastructure like India, the authors aim to develop a robust, real-time detection system that overcomes the limitations of existing Advanced Driver Assistance Systems (ADAS). Current solutions often fail in heterogeneous conditions due to poor illumination handling, lack of temporal analysis, and insufficient interpretability, which hinders user trust. To address these gaps, the study proposes a deep learning framework integrating a fine-tuned InceptionV3 baseline enhanced with dual attention mechanisms (spatial and channel) to focus on fatigue-relevant facial cues. The system incorporates a Low-Light Fine-Tuned LLFormer for image enhancement and utilizes ResNet-50 for feature extraction. The methodology involves aggregating data from multiple public datasets, including YAWDD, NTHU-DDD, LOL, and ExDark, to ensure diversity in lighting and occlusion scenarios. Preprocessing steps include pixel normalization, histogram equalization, and noise reduction. The model analyzes multiple drowsiness indicators—head tilting, blinking, and yawning—using temporal factors and facial landmark detection. Additionally, Explainable AI (XAI) techniques, such as Grad-CAM and LRP, are employed to provide transparency and interpretability to the model’s predictions. Experimental results demonstrate that the proposed system achieves state-of-the-art performance, reaching up to 98.4% accuracy even under challenging conditions involving low light, eyewear, and varied facial occlusions. The model was validated across diverse benchmark datasets and optimized for computational efficiency, enabling real-time deployment on mobile and embedded platforms with minimal overhead. Ablation studies confirmed the contribution of each component, including the attention mechanisms and low-light enhancement module, to the overall performance. The significance of this work lies in its potential to significantly reduce drowsy driving-related risks by providing a scalable, interpretable, and robust tool for road safety. The system supports a tiered alert mechanism, ranging from mild alarms for low drowsiness to automatic emergency notifications for extreme cases, thereby offering proactive safety assistance. By addressing the specific challenges of low-light detection and interpretability, the study offers a practical solution that can be integrated into diverse vehicular settings, enhancing the reliability and accessibility of driver monitoring systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms