Drowsy Driver Detection Using Two Stage Convolutional Neural Networks

Abstract

Driver drowsiness and fatigue is one of the most significant causes of road accidents. Accidents involving drowsy drivers have claimed millions of lives in the past years making automated driver drowsiness detection an important computer vision problem. In order to tackle this problem, a real time driver drowsiness detection system is implemented based on the deep learning using Convolutional Neural Network (CNN). In the proposed method, drowsiness detection is treated as an object detection and classification task. In this method, detection and localization of face region is done using the YOLOv3 real-time object detection algorithm, while the Inception-v3 pre-trained neural network is used to classify the detected face as either drowsy or non-drowsy. The deep learning model was trained and tested on the standard datasets: Closed Eyes in the Wild (CEW) database, National Tsuing Hua University (NTHU) Driver Drowsiness Detection database and a custom database. The proposed methodology gives an accuracy of 80.32%, 79.34%and 89.90% respectively, on the three databases. Proposed system can process the input video stream in real-time without any expensive hardware like GPUs and hence is computationally efficient and cost effective, since it is able to process an incoming video stream in real-time, on a standalone device, without the need for any expensive hardware support. Also the methodology adopted does not involve face feature extraction which is a common procedure in most of the vision based drowsiness detection systems, making the model efficient without compromising on its prediction speed or accuracy.