Distracted Driver Detection with MobileVGG Network

Abstract

The escalation of road traffic fatalities in recent years has highlighted the issue of distracted driving as a significant problem that warrants attention. This paper presents a CNN-based approach for identifying and categorizing distracted driving behavior, catering to the requirements of advanced driver assistance systems. The proposed algorithm demonstrates an optimal balance between accuracy and efficiency, with respect to memory consumption and processing speed. The architecture employed, termed mobile VGG, is founded on the principles of deeply separable convolution. The outcome of de-duplicating the American University in Cairo's (AUC) dataset for distracted driving detection reveals that the proposed mobile VGG architecture has just 2.2M parameters and achieves 95.50% accuracy on the AUC dataset with 38% less computing time than alternative methods.