An Explainable Real-Time Driver Drowsiness Detection System Using Vision Transformers and Behavioral Fatigue Indicators

Pisolla Rahul, Mohammed Afridi; Guguloth Venkatesh, Dr S Dheeraj · 2026 · Crossref

DOI: 10.48175/ijarsct-33444

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, a leading cause of road accidents during prolonged or monotonous driving. While existing solutions rely on intrusive physiological sensors or vehicle-based metrics, vision-based methods offer a non-intrusive alternative. However, traditional deep learning models, particularly Convolutional Neural Networks (CNNs), often function as "black boxes," lacking the explainability required for trust and regulatory acceptance in safety-critical applications like Advanced Driver Assistance Systems (ADAS). To bridge this gap, the authors propose an explainable, real-time driver drowsiness detection system that integrates Vision Transformers (ViT) with interpretable behavioral fatigue indicators. The proposed system utilizes a standard RGB camera to capture real-time video streams. Facial landmarks are extracted using MediaPipe FaceMesh, identifying 468 key points to compute heuristic fatigue metrics: Eye Aspect Ratio (EAR) for eye closure, Mouth Aspect Ratio (MAR) for yawning, blink frequency, and PERCLOS (percentage of eye closure). Simultaneously, facial image regions are processed by a Vision Transformer, which leverages self-attention mechanisms to capture global spatial dependencies and classify alertness states. A hybrid decision fusion mechanism combines the outputs of these rule-based indicators and the ViT model. To ensure robustness, temporal smoothing is applied to stabilize predictions across consecutive frames, triggering audio and visual alerts only when sustained drowsiness is detected, thereby minimizing false alarms. Experimental evaluations on multiple public datasets and real-time camera inputs demonstrate that the system achieves high detection accuracy with low latency. The hybrid approach outperforms standalone CNN and ViT models in both accuracy and explainability. The system effectively reduces false positives through its temporal smoothing strategy and provides transparent decision-making by relying on human-understandable metrics alongside deep learning outputs. The authors note limitations regarding performance in low-light conditions, face occlusion, and the computational cost of ViT models. The significance of this work lies in its contribution to explainable AI in intelligent transportation systems. By combining the high performance of Vision Transformers with the interpretability of behavioral indicators, the system offers a practical, non-intrusive solution for real-world deployment. It enhances driver safety by providing reliable, early-stage drowsiness detection while maintaining the transparency necessary for user trust and regulatory compliance. The framework is designed for scalability and integration into next-generation smart vehicle technologies, with future work directed toward edge AI deployment and multi-modal detection.

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-nvidia on 2026-08-10; verification: verified.

Topics

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

• drowsiness detection algorithms
• distraction detection algorithms
• gaze based attention detection
• drowsiness
• dms validation

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data
• Methodological Resource: tool software