| Summary: | A dissertation submitted in fulfilment of the requirements for the degree Master of Science in the Faculty of Science, the University of the Witwatersrand, 2020 === According to World Health Organization, driver drowsiness and distraction are leading factors of road crashes. In 2013, approximately 1.25 million deaths were caused by road traffic injuries [1] and by 2018, this number had swelled up to 1.35 million [2]. This report can be interpreted to mean that approximately one person experiences (or will experience) a fatal crash every twenty four seconds [3] [2]. Thus, because of the rising drowsiness related accidents, it is pertinent to know the state of a driver in real time to allow for a prior warning to either take a break or hand over the steering wheel to another person (or an artificial intelligent smart vehicle where applicable). To achieve this goal, there is a need to develop a system which continuously monitors the driver’s alertness in real time. In this research, we introduce a non-intrusive novel approach to monitor driver alertness in real time using deep neural network architectures. Specifically, we used deep convolutional neural network architectures to design a system which monitors driver alertness in real time. “Convolutional neural networks (CNNs)” have been notably known as the best machine learning architectures when it comes to image classification. Thus, in this research, we explored different convolutional neural network architectures in developing an efficient face-state detection model. The decisions of these CNNs are fed to our drowsiness evaluator which is formed with conditions based on our findings from the literature. Specifically, we focused on LeNet-5 and four other notable convolutional neural network architectures which have won (or performed excellently) at the ImageNet challenge and these include “AlexNet,” “Inception-ResNet V2,” “VGGNet”, and “ResNext-50.” The neural networks were trained on publicly available dataset as well as publicly available images we compiled using Google and Bing APIs. We compared the performance of our proposed driver alertness monitoring framework using the different CNNs against one another in terms of their accuracy, computation speed, Video Error and Average Video Error. Experimental results show ”ResNext-50” to be our best performing model, even though it only had the second best Average Video Error, after “Inception ResNet V2,” by a gap of 0.006. In previous literature, researchers have focused on predicting drowsiness with single frames classification, however, this is not an efficient approach as it raises a lot of false positives and is not robust to other drowsiness conditions. Furthermore, previous literatures focus on drowsiness classification and not detection, however, in this study, we have presented a framework for detecting and monitoring drowsiness, and categorizing alertness into different levels which are: Alert, Slightly Drowsy, Moderately Drowsy, Very Drowsy, and Dangerously Drowsy === CK2021
|
|---|
