A Deep Neural Network for Real-Time Driver Drowsiness Detection
DOI: 10.1587/transinf.2019edl8079
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 challenge of developing a real-time, video-based driver drowsiness detection (DDD) system that balances high accuracy with low computational cost. Drowsy driving contributes significantly to traffic accidents, yet existing solutions often rely on inconvenient wearable sensors or computationally expensive deep learning models that require post-processing or window-based predictions. The authors aim to create a visual DDD system that processes global faces directly, avoiding the latency and complexity associated with frame-by-frame face alignment, thereby enabling practical, real-time application. The proposed method utilizes a deep neural network (DNN) composed of three components: a convolutional neural network (CNN), a convolutional control gate-based recurrent neural network (ConvCGRNN), and a voting layer. The preprocessing stage extracts driver faces from video streams using a combination of face detection (MTCNN) and tracking (dlib correlation filter), which is reinitialized periodically to prevent drift. This approach is significantly faster than performing detection on every frame. The CNN, structured in a VGG-style architecture, extracts spatial facial representations from grayscale, resized faces. These features are passed to the ConvCGRNN, which learns temporal dependencies while preserving spatial properties through a control gate mechanism. Finally, a voting layer, acting as an ensemble of sub-classifiers, predicts the drowsiness state. The model processes frames sequentially without resetting hidden states, minimizing inference time. Experiments were conducted on the public NTHU-DDD dataset, which includes videos of various subjects under different conditions (e.g., bareface, glasses, night scenarios). The model was pretrained on the FER+ dataset and trained using binary cross-entropy loss with Adam optimization. The proposed DNN achieved an overall accuracy of 84.81% on the test set, outperforming previous state-of-the-art methods, including human performance (80.83%). Performance varied by scenario, with higher accuracy in night-time conditions (88.00–88.82%) compared to occluded scenarios like sunglasses (74.08%). Crucially, the system demonstrated high efficiency, with the DNN model achieving an inference speed of approximately 100 frames per second (fps). The combined preprocessing and detection pipeline operated at 60 fps, which is twelve times faster than using face detection alone. The significance of this work lies in its demonstration that high-accuracy drowsiness detection can be achieved without computationally intensive post-processing or window-based prediction methods. By leveraging ConvCGRNN for efficient spatiotemporal modeling and optimizing the preprocessing pipeline, the authors provide a solution that meets the strict real-time requirements for in-vehicle monitoring systems. This approach offers a practical alternative to slower, more complex deep learning architectures, facilitating the deployment of active driver monitoring systems that can provide timely warnings to prevent accidents.
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | OpenAlex-citations | — | — | 1 | 2026-06-24 |
| archive | success | unpaywall | — | — | 2 | 2026-06-26 |
| extract | success | cached | — | — | 2 | 2026-06-26 |
| clean | success | clean | — | — | 1 | 2026-06-25 |
| chunk | success | chunk | — | — | 1 | 2026-06-25 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-25 |
| promote | success | — | — | — | 1 | 2026-06-24 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 1 | 2026-06-26 |
| tag | success | vector_similarity | — | — | 6 | 2026-06-25 |
| verify | success | — | — | — | 1 | 2026-06-26 |
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.
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