Abstract:
Distracted driving is a significant contributor to traffic accidents worldwide,
necessitating effective driver behavior monitoring systems. This study
investigates the efficacy of state-of-the-art Convolutional Neural Network (CNN)
architectures—VGG16, ResNet50, InceptionV3, Xception, and
Xception_Customize_Model—for real-time driver behavior detection. The
research employs a systematic methodology to evaluate these architectures on
parameters such as accuracy, computational efficiency, and generalization
ability. Through data preprocessing techniques like image augmentation and
normalization, models were trained to classify driver behaviors such as safe
driving, texting, talking on the phone, turning, and others. The experimental
results highlight that Xception_Customize_Model achieved the highest
validation accuracy of 97.07% with a consistently low validation loss,
demonstrating superior performance and stability. The Xception model followed
closely with competitive accuracy but exhibited occasional fluctuations in
validation performance. ResNet50 also performed well, with a validation
accuracy of 88.25%, reflecting robust classification capabilities despite requiring
more epochs for stabilization. In contrast, InceptionV3, VGG16, and CNN
demonstrated lower performance, with validation accuracies not exceeding 70%,
largely due to higher training losses and limited generalization ability. The
findings of this study contribute to the development of intelligent transportation
systems by enhancing real-time detection of driver distractions, thereby
promoting road safety. This research provides valuable insights into optimizing
CNN architectures for real-world applications, offering a pathway for
practitioners to implement scalable, efficient, and accurate driver monitoring
solutions.
