Smart Vision-Based Driver Safety Monitoring System

Abstract

Many road accidents happen because a driver becomes tired or sleepy, and safety features such as seat belts and airbags only reduce the damage after a crash has already begun. This paper presents the design of a Smart Vision-Based Driver Safety Monitoring System that watches the driver through an ordinary dashboard or laptop camera and warns them when signs of drowsiness appear. The system looks at several clues together, namely how long the eyes stay closed, whether the driver is yawning, and how the head is positioned, instead of depending on a single clue. A deep learning model called a Convolutional Neural Network (CNN), improved with an attention mechanism known as CBAM, helps the system concentrate on the eye and mouth regions that matter most. To stay dependable when the view is poor, the system first checks the quality of each frame, adjusts how much weight each clue receives, smooths its decisions over a short period of time, and raises an alert only when two separate checks agree. A review of sixteen related studies from 2021 to 2026 shows the gaps that this design aims to close, and preliminary training results are reported to indicate the current stage of the implementation. The proposed solution is low-cost, needs no wearable sensors, and is intended to run on standard computing hardware.

Introduction

The text presents a Smart Vision-Based Driver Safety Monitoring System designed to detect driver drowsiness in real time and prevent road accidents caused by fatigue. Driver drowsiness reduces reaction time, weakens judgment, and increases the risk of collisions, particularly during long journeys and night-time driving. Traditional safety features mainly reduce injuries after an accident occurs, whereas the proposed system aims to identify early signs of fatigue and alert drivers before a dangerous situation develops.

The system uses a standard camera, computer vision, and deep learning techniques to monitor facial features without requiring wearable sensors. Its methodology consists of six major stages:

• Video Acquisition and Face Detection: OpenCV captures video frames, while HOG+SVM and dlib facial landmarks identify the driver's face, eyes, and mouth.

• Frame-Quality Assessment: Evaluates lighting, facial visibility, and occlusion to prevent unreliable frames from producing misleading predictions.

• Adaptive Multi-Cue Detection: Analyzes Eye Aspect Ratio (EAR), Mouth Aspect Ratio (MAR), and head pose to detect prolonged eye closure, yawning, and head nodding. The system dynamically adjusts the importance of each cue according to its reliability.

• CNN with CBAM Attention: Uses a Convolutional Neural Network enhanced with the Convolutional Block Attention Module to focus on relevant facial regions and improve feature extraction.

• Temporal Smoothing and Decision Verification: Analyzes consecutive frames to distinguish genuine drowsiness from normal blinking or brief movements. An alert is generated only when both the predicted drowsiness probability and model confidence exceed predefined thresholds.

• Alert and Monitoring Interface: Displays live driver status, the available detection cues, system warnings, and notifications when facial visibility or frame quality is insufficient.

The literature review examines existing approaches, including EAR-based methods, CNN classifiers, CNN-LSTM architectures, attention mechanisms, and lightweight edge-device models. It identifies several limitations in current systems, such as sensitivity to poor lighting and sunglasses, dependence on individual facial cues, excessive computational requirements, and limited adaptability when certain facial features are obscured. The proposed system addresses these challenges through quality-aware frame selection, adaptive multi-cue analysis, attention-enhanced feature extraction, and lightweight temporal processing.

The system is designed for deployment in private vehicles, commercial fleets, taxis, driver-training simulators, and low-cost hardware platforms such as Raspberry Pi and NVIDIA Jetson Nano. Its camera-only architecture aims to provide an affordable and non-intrusive driver monitoring solution.

Preliminary experimental results, however, reveal that the current classifier requires significant improvement. Training and validation accuracy remain at approximately 64.3% across 15 epochs. On a test dataset of 9,078 samples, the model predicts every sample as drowsy, achieving 100% recall for the drowsy class but failing to identify non-drowsy samples. The overall accuracy therefore reflects the proportion of drowsy samples in the dataset rather than effective discrimination between the two classes. Improving model training and class separation is the immediate priority.

Conclusion

This paper has presented the design of a Smart Vision-Based Driver Safety Monitoring System that combines computer vision and deep learning to detect driver drowsiness from an ordinary camera. By combining frame-quality gating, adaptive eye, mouth, and head-pose cues, a CNN+CBAM feature extractor, sliding-window smoothing, and a two-gate probability–confidence check, the proposed approach aims to remain reliable under sunglasses, poor lighting, and partial occlusion while staying light enough for real-time use on standard hardware. The CBAM-enhanced CNN of the base work [15] reports over 99% accuracy, but the preliminary results of the present implementation reflect only the class distribution and show that further refinement of the training is needed. Future work will complete this refinement and evaluate the full adaptive framework on real driving footage, validating its robustness, latency, and false-alarm rate.

References

[1] S. K. Sharma, A. Choudhary, and R. Kumar, “Driver drowsiness detection using deep learning and computer vision,” IEEE Access, vol. 9, pp. 92187–92199, 2021. [2] M. R. Islam, M. S. Hossain, and A. Alamri, “Vision-based driver drowsiness detection using CNN,” IEEE Access, vol. 9, pp. 34519–34530, 2021. [3] N. Gupta, R. Sharma, and P. Arora, “Real-time driver drowsiness detection using deep neural networks,” in Proc. 12th Int. Conf. on Computing, Communication and Networking Technologies (ICCCNT), 2021, pp. 1–6. [4] P. Singh, S. Jain, and A. K. Singh, “Real-time driver drowsiness detection using facial features and eye aspect ratio,” in Proc. Int. Conf. on Electronics and Sustainable Communication Systems (ICESC), 2022, pp. 1–6. [5] K. Verma, A. Mishra, and S. Srivastava, “Driver drowsiness detection using hybrid deep learning model,” in Proc. Int. Conf. on Intelligent Systems and Smart Infrastructure (ICISSI), 2023, pp. 1–5. [6] J. Li, X. Wang, et al., “Vision-based driver drowsiness detection using lightweight CNN and attention mechanism,” IEEE Trans. Intell. Transp. Syst., vol. 25, no. 2, pp. 1234–1245, 2024. [7] S. Kumar, R. Patel, et al., “Real-time driver drowsiness detection using YOLOv8 and facial landmark analysis,” Multimedia Tools Appl., Springer, vol. 83, no. 5, pp. 14321–14338, 2024. [8] A. Singh, P. Verma, et al., “Edge-based driver drowsiness detection using MobileNet and IoT,” IEEE Internet Things J., vol. 11, no. 4, pp. 6780–6791, 2024. [9] A. Bhanja, S. Paul, et al., “A real-time driver drowsiness detection system using deep learning,” Int. J. Mach. Learn. Cybern., Springer, vol. 16, no. 1, pp. 123–136, 2025. [10] M. A. Albadawi, H. Alshammari, et al., “Real-time driver drowsiness detection using facial analysis,” MDPI Sensors, vol. 25, no. 3, p. 812, 2025. [11] O. Nasir, M. A. Khan, et al., “SAFE-DRIVE-AI: A CNN–LSTM–attention framework for driver drowsiness detection,” Eng., Technol. Appl. Sci. Res. (ETASR), vol. 15, no. 1, pp. 1–8, 2025. [12] O. F. Hassan, A. Mahmood, et al., “Real-time driver drowsiness detection using transformer architectures: A novel deep learning approach,” Sci. Rep., Nature Publishing Group, vol. 15, no. 1, p. 9821, 2025. [13] H. Chen, Y. Zhao, et al., “Driver drowsiness detection using multimodal deep learning,” Neurocomputing, Elsevier, vol. 580, p. 127450, 2025. [14] M. Rahman, T. Islam, et al., “Real-time driver drowsiness detection using hybrid CNN-BiLSTM model,” Neural Comput. Appl., Springer, vol. 37, no. 3, pp. 2345–2360, 2025. [15] O. F. Hassan, A. F. Ibrahim, A. Gomaa, S. Alshathri, and W. El-Shafai, “Comprehensive benchmarking of attention-enhanced CNNs and transfer learning models for real-time driver drowsiness detection,” IEEE Access, vol. 14, pp. 12310–12330, 2026. [16] J. Menda, B. N. Marrapu, R. T. Bheri, and C. S. N. Chebodula, “Efficient and secure video streaming using multi-key cryptography,” International Research Journal of Modernization in Engineering Technology and Science, vol. 6, no. 11, Nov. 2024. [17] F. Chollet, Keras: The Python Deep Learning Library, 2017. [Online]. Available: https://keras.io. [18] NTHU Driver Drowsiness Detection Dataset, 2023. [Online]. Available: http://cv.cs.nthu.edu.tw/php/callforpaper/datasets/DDD/. [19] OpenCV, 2024. [Online]. Available: https://opencv.org. [20] dlib C++ Library, 2024. [Online]. Available: http://dlib.net.

Copyright

Copyright © 2026 Mrs. Sowmya D, Sanjana M L, Shreya S, Yashu A B, Khushi N. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.