Distracted Driver Detection Using Deep Learning

• Main Author: Koh, Qi Zhe
• Format: Final Year Project Report
• Language: English; English
• Published: Universiti Malaysia Sarawak, (UNIMAS) 2023
• Subjects: T Technology (General)
• Online Access: http://ir.unimas.my/id/eprint/43104/1/Koh%20Qi%20Zhe%2024%20pgs.pdf; http://ir.unimas.my/id/eprint/43104/6/Koh%20Qi%20Zhe%20ft.pdf; http://ir.unimas.my/id/eprint/43104/

Driving involves a wide range of complex operations and the coordination of multiple senses, making it a task that requires utmost attention and focus. Various factors, such as cell phone use, adjusted audio equipment, smoking, consumed food and drinks, conversed with passengers, or experienced fatigue, distracted drivers and jeopardized their safety, resulting in car collisions and injuries. The rising prevalence of distracted driving poses significant risks to road safety, leading to increased accidents and fatalities. Various studies explore different approaches to detect driver distraction, from traditional machine learning to advanced deep learning. Deep learning, especially CNN-based methods, show better accuracy and real-time performance. In this project, a lightweight deep learning detection model based on MobileNetV2 was proposed to address the detection of distracted driver actions without causing discomfort to the driver, as some current technologies involve wearing sensors that can be uncomfortable. The proposed model was enhanced by incorporating attention mechanisms like the SE module. The model was trained and tested using the American University in Cairo (AUC) distracted driver dataset, which encompassed 10 distraction categories. Techniques like hyperparameter tuning, data augmentation, and class weighting were utilized to optimize the model, achieving an impressive accuracy of 93% with the configuration of batch size 32, learning rate 0.0001, 21 epochs. Evaluation metrics, including confusion matrix, FPS, accuracy, precision, recall, and F1 score, were employed to assess the effectiveness of the model. Additionally, the proposed method was compared to MobileNetV2 model and other existing architectures in terms of accuracy and parameters. It outperformed unmodified deep learning models and maintained a balance between accuracy and parameter utilization, while some other modified models perform slightly better. The proposed method exhibiting promising potential in accurately detecting distracted drivers with efficiency.