According to the World Health Organization (WHO) report, the number of road traffic deaths have been continuously increasing since last few years though the rate of deaths relative to world's population has stabilized in recent years. As per the survey of National Highway Traffic Safety Administration (NHTSA), distracted driving is a leading factor in road accidents. In this paper, we present a Convolutional Neural Network (CNN) based approach for detecting and classifying the driver distraction. In the development of safety features for Advanced Driver Assistance Systems, the algorithm not only has to be accurate but also efficient in terms of memory and speed. Hence, we focused on developing computationally efficient CNN while maintaining good accuracy. We propose a new architecture named as mobileVGG based on depthwise separable convolutions. We evaluate results of the proposed network on the American University in Cairo (AUC) distracted driver detection dataset as well as Statefarm's dataset on Kaggle and compare the performance with state-of-the-art CNN architectures from literature. Our proposed mobileVGG architecture with just 2.2M parameters outperforms earlier approaches while achieving 95.24% and 99.75% accuracy on AUC and Statefarm's dataset respectively with less computational complexity and memory requirement.
Read full abstract