Vision-Based Driver Drowsiness Detection: From Deep Learning Models to Real-Time Mobile Deployment

Jadhav, Hitesh Jitendra · 2026 · Crossref

DOI: 10.22214/ijraset.2026.81356

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, which contributes to 20–30% of highway fatalities globally. While deep learning models have achieved high accuracy in controlled laboratory settings, existing systems often fail in real-world deployment due to high computational demands, sensitivity to environmental variations, and reliance on intrusive physiological sensors. The authors identify a significant gap between benchmark performance and practical utility, specifically the lack of robust, non-intrusive systems capable of real-time inference on resource-constrained mobile devices. To bridge this divide, the study proposes a lightweight, vision-based detection system optimized for on-device execution using Android smartphones. The methodology employs a MobileNetV2 convolutional neural network (CNN) trained on benchmark datasets (NTHU-DDD, YawDD, CEW) and a custom dataset collected during real driving scenarios. The system extracts facial landmarks to compute geometric features, including Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR), which are concatenated with CNN activation vectors for hybrid classification. To ensure efficiency, the model undergoes post-training quantization, comparing Float32, Float16, and INT8 precision levels. The final architecture is deployed via TensorFlow Lite, ensuring all processing occurs locally on the device without cloud connectivity. The design incorporates a "Privacy-by-Design" protocol, utilizing volatile memory only and retaining no biometric data. Experimental results demonstrate that the proposed INT8-quantized MobileNetV2 model achieves 93.76% accuracy on the NTHU-DDD dataset, with a minimal accuracy loss of less than 0.5% compared to the baseline Float32 model. Crucially, the quantized model reduces the file size to 4.2 MB—a 125-fold reduction compared to VGG16—and achieves an average inference latency of 38 ms, enabling 26 frames per second on mid-range Android hardware. This represents a tenfold latency improvement over heavier architectures. The system maintains above 90% accuracy across challenging real-world conditions, including night driving, sunglasses usage, and partial face occlusion, though false alarm rates increase slightly under heavy occlusion. The significance of this work lies in its demonstration that high-accuracy drowsiness detection can be deployed scalably on commodity hardware without compromising privacy or battery life. By validating the feasibility of sub-40ms inference on mobile processors, the study provides a viable pathway for integrating automated safety monitoring into everyday vehicles. The authors conclude that such non-intrusive, cost-effective solutions are essential for transforming drowsiness detection from a specialized luxury into a standard safety infrastructure element, comparable to seatbelts or airbags. Future work will focus on extensive on-road validation across diverse driver populations to further refine robustness.

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
• 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).

• Methodological Resource: tool software