A Transfer Learning Approach for Identification of Distracted Driving

Abstract

In recent years, the number of motor vehicle traffic crashes is rising around the world. According to the United States National Highway Traffic Safety Administration (NHTSA), the main cause of road fatalities and injuries is distracted drivers. Driver distraction is a specific sort of driver interference on the roadway. In this case, a deep learning-based system can detect and distinguish the source of distractions in real-time, to avoid traffic accidents and make better transport safety. This paper attempts to develop the system using transfer learning and fine-tuning methods with different model architectures. Various pre-trained weights were used with fine-tuning to improve accuracy and implemented using Mobile Net, VGG16, and ResNet50 models. Finally, the results illustrate that transfer learning on Mobile Net with frozen layers is the best model out of three models getting 99% (495 out of 500) of accuracy on the test dataset images.