Driver drowsiness detection using deep learning approaches

Abstract

A drowsy driver appears to be far more dangerous on the road than one who is driving too fast as he might have microsleeps. Vehicle engineers are seeking to investigate this problem with a few mechanical solutions that would prevent such an emergency. In this work, we proposed a solution based on Deep Learning (Deep Learning (DL)) and Vehicular Adhoc Network (VANET) approaches. Our strategy is mainly divided into two parts: The frst one employed DL approaches including Transfer Learning (TL) and data augmentation techniques, in order to create an accurate model that classify driver images into drowsy or awake. In this case, drowsiness is detected using an auto-camera used in smart vehicles. Based on the obtained images, the driver state is predicted as awake or drowsy by a proposed convolutional neural network Convolutional Neural Network (CNN). To achieve that, four pre-trained CNN models were modifed and adapted to our problem. Each of Residual Network (ResNet)50V2, Visual Geometric Group (VGG)16, VGG19, and Inception V3 were employed in several trials and with various fne-tuning parameters to establish the most appropriate one for our case. We gathered more than 6000 images from two different datasets of drowsiness and yawn to train and evaluate our models. The obtained results were exceptionally satisfying according to the previous studies, putting InceptionV3 and ResNet50V2 ahead of the other models with 96.85% and 96.70% accuracy respectively, followed by VGG 19 with 92.23%. Meanwhile, the VGG 16 reached the lowest accuracy of 50.48%. The second part contains two scenario’s : First, broadcasting messages to alert the approximate vehicles of the car whose driver may be drowsy. As for the second scenario, an alarm was employed to awake the driver using site vibration, or smart devices (phone or watch).