Benchmarking Deep Learning Models for Driver Distraction Detection

The World Health Organisation reports distracted driving as one of the main causes of road traffic accidents. Current studies to detect distraction postures focus on analysing image features. However, there is lack of a comprehensive evaluation of state-of-the-art deep learning techniques employed for driver distraction detection, which limits or misguides future research in this area. In this paper, we conduct an in depth review of deep learning methods used in driver distraction detection and benchmark these methods including other popular state-of-the-art CNN and RNN techniques. This will assist researchers to compare their novel deep learning methods with state-of-the-art models for driver distraction posture identification. We evaluate 10 state-of-the-art CNN and RNN methods using the average cross-entropy loss, accuracy, F1-score and training time on the American University in Cairo (AUC) Distracted Driver Dataset, which is the most comprehensive and detailed dataset on driver distraction to date. Results show that pre-trained InceptionV3 CNNs coupled with stacked Bidirectional Long Short Term Memory outperforms state-of-the-art CNN and RNN models with an average loss and F1-score of 0.292 and 93.1% respectively.