An Ensemble Approach for Identification of Distracted Driver by Implementing Transfer Learned Deep CNN Architectures

Abstract

Road accidents are common issues caused by overpopulation in almost all countries around the world. In developed and developing nations, one of the main causes of road accidents is driver distraction. Due to the distraction of the drivers, the time required for decision-making can be reduced significantly and result in catastrophic damage. Texting or talking over the phone while driving, drinking, reaching behind, hair and makeup, operating the radio, and talking to the passenger are common causes of driver distraction happens. Previously, many works have been conducted in this domain as the lives of passengers depend on the successful driving of drivers. Diverse proposals have been introduced throughout the last decade for driver distraction detection including machine and deep learning approaches. Although some methods seem promising, most of them couldn’t achieve the desired performance due to the lack of important feature extraction. The recent advances in transfer learning, however, created an opportunity for contribution in this sector as these approaches are capable of extracting deep features. In this study, we have considered 10 states of driver distraction and utilized two transfer learning approaches i.e., DenseNet121 and MobileNet for successful recognition of driver distraction. However, further application of ensemble learning produced a better accuracy than DenseNet121 and MobileNet alone. The proposed ensemble model achieved an overall accuracy of 99.81% whereas DenseNet121 and MobileNet achieved an overall accuracy of 99.66% and 99.73% respectively.