Convolutional neural network-based real-time drowsy driver detection for accident prevention
Abstract
Drowsy driving significantly threatens road safety, contributing to many accidents globally. This paper presents a convolutional neural network (CNN)-based real-time drowsy driver detection system aimed at preventing such accidents, particularly for deployment in Android applications. We propose a lightweight CNN architecture that effectively identifies drowsiness and microsleep episodes by categorizing driver facial expressions into four distinct categories: close-eye expressions, open-eye expressions, yawns, and no yawns. Our model, which employs facial landmark detection and various pre-processing techniques to enhance accuracy, achieves an impressive 96.6% accuracy. This performance surpasses several popular CNN architectures, including VGG16, VGG19, MobileNetV2, ResNet50, and DenseNet121. Notably, our proposed model is highly efficient, with only 0.4 million parameters and a memory requirement of 1.51 MB, making it ideal for real-time applications. The comparative analysis highlights the superior balance between accuracy and resource efficiency of our model, demonstrating its potential for practical deployment in reducing accidents caused by driver fatigue.