Design of an Efficient Distracted Driver Detection System: Deep Learning Approaches

Abstract

Distracted driving is any activity that deviates an individual’s attention from driving. Some of these activities include talking to people in the vehicle, using hand-held devices such as mobile phones or tablets, eating or drinking, and adjusting the stereo or navigation systems while driving. To counter the effects caused by distracted driving, many countries around the world have imposed rules and charged fines on distracted drivers in order to ensure safe driving. Owing to the technological advancement in recent times, modern-day technologies such as computer vision, image processing, and machine learning techniques can further support the efforts of governments to prevent accidents caused by distracted driving. In this paper, an efficient distracted driver detection scheme (DDDS) has been proposed using two robust deep learning architectures, mainly visual geometric groups (VGG-16) and residual networks (ResNet-50). The proposed DDDS scheme contains the pre-processing module, image augmentation techniques, and two classification modules based on deep learning architectures. Both the architectures are implemented, and the results have been compared in terms of performance indices, namely accuracy, and logarithmic loss. The two-dimensional (2D) dashboard images derived from the State-Farm dataset are pre-processed and are used for training, testing, and validation of the proposed architectures. Accuracy of 86.1% and 87.92% are achieved with VGG-16 and ResNet-50 models, respectively, and it is observed that the DDDS scheme is found highly efficient for c4, c5, and c7 categories of the State-Farm dataset. The results obtained with the proposed DDDS methodology are compared with existing literature and found to be satisfactory. The algorithms developed and discussed for the proposed DDDS can be instrumental in reducing the fatalities and injuries caused due to distracted driving.