Recognition of driver distractions using deep learning

Driver distraction has a great impact on the safety of people and it is a relevant topic for a number of applications, from autonomous driving assistance to insurance companies and investigations. In this paper we address the problem of automatic recognition of driver distractions by exploiting deep learning and convolutional neural networks. We propose and present a comparison of different deep learning-based methods to classify driver's behaviour using data from 2D cameras. Evaluation has been carried out on the State Farm dataset, which consists of 10 different actions performed by 26 subjects such as, normal driving, texting, talking on the phone, operating the radio, drinking, reaching behind, etc. Results, achieved using 3 rounds of 5- fold cross validation, show that all the evaluated methods exceed the 90% of accuracy with the best achieving about 97%.