Driver Distraction Detection Using Advanced Deep Learning Technologies Based on Images

Abstract

Driver distraction is an important factor leading to traffic accidents. The identification of drivers’ distracted states is of great significance to improve traffic safety. This paper presents a new driver distraction discrimination method by comprehensively deploying transfer learning technologies in deep learning algorithms. We use three publicly available datasets (i.e., state farm, origin, and 3MDAD) to verify the effectiveness of the developed methods based on Alexnet, VGG16, and Resnet18. The images with high recognition errors are further investigated by class activation map. The experiment results show that Resnet18 on origin has the best performance. The accuracies of the developed methods based on Alexnet, VGG16, and Resnet18 can achieve 99.92%, 100%, and 99.99% on the origin dataset, respectively. The accuracies of the Resnet18 -based method on state farm and 3MDAD are 84.82% and 97.7%, respectively. Our proposed methods can well support the development of distraction detection applications in transportation systems.