Detection of Driver Drowsiness and Distraction Using Computer Vision and Machine Learning Approaches

Abstract

Drowsiness and distracted driving are leading factor in most car crashes and near-crashes. This research study explores and investigates the applications of both conventional computer vision and deep learning approaches for the detection of drowsiness and distraction in drivers. In the first part of this MPhil research study conventional computer vision approaches was studied to develop a robust drowsiness and distraction system based on yawning detection, head pose detection and eye blinking detection. These algorithms were implemented by using existing human crafted features. Experiments were performed for the detection and classification with small image datasets to evaluate and measure the performance of system. It was observed that the use of human crafted features together with a robust classifier such as SVM gives better performance in comparison to previous approaches. Though, the results were satisfactorily, there are many drawbacks and challenges associated with conventional computer vision approaches, such as definition and extraction of human crafted features, thus making these conventional algorithms to be subjective in nature and less adaptive in practice. In contrast, deep learning approaches automates the feature selection process and can be trained to learn the most discriminative features without any input from human. In the second half of this research study, the use of deep learning approaches for the detection of distracted driving was investigated. It was observed that one of the advantages of the applied methodology and technique for distraction detection includes and illustrates the contribution of CNN enhancement to a better pattern recognition accuracy and its ability to learn features from various regions of a human body simultaneously. The comparison of the performance of four convolutional deep net architectures (AlexNet, ResNet, MobileNet and NASNet) was carried out, investigated triplet training and explored the impact of combining a support vector classifier (SVC) with a trained deep net. The images used in our experiments with the deep nets are from the State Farm Distracted Driver Detection dataset hosted on Kaggle, each of which captures the entire body of a driver. The best results were obtained with the NASNet trained using triplet loss and combined with an SVC. It was observed that one of the advantages of deep learning approaches are their ability to learn discriminative features from various regions of a human body simultaneously. The ability has enabled deep learning approaches to reach accuracy at human level.