Correcting inter-sectional accuracy differences in drowsiness detection systems using generative adversarial networks (GANs)

Abstract

oad accidents contribute to many injuries and deaths among the human population. There is substantial evidence that proves drowsiness is one of the most prominent causes of road accidents all over the world. This results in fatalities and severe injuries for drivers, passengers, and pedestrians. These alarming facts are raising the interest in equipping vehicles with robust driver drowsiness detection systems to minimise accident rates. One of the primary concerns of motor industries is the safety of passengers and as a consequence they have invested significantly in research and development to equip vehicles with systems that can help minimise to road accidents. A number research endeavours have attempted to use Artificial intelligence, and particularly Deep Neural Networks (DNN), to build intelligent systems that can detect drowsiness automatically. However, datasets are crucial when training a DNN. When datasets are unrepresentative, trained models are prone to bias because they are unable to generalise. This is particularly problematic for models trained in specific cultural contexts, which may not represent a wide range of races, and thus fail to generalise. This is a specific challenge for driver drowsiness detection task, where most publicly available datasets are unrepresentative as they cover only certain ethnicity groups. This thesis investigates the problem of an unrepresentative dataset in the training phase of Convolutional Neural Networks (CNNs) models. Firstly, CNNs are compared with several machine learning techniques to establish their superior suitability for the driver drowsiness detection task. An investigation into the implementation of CNNs was performed and highlighted that publicly available datasets such as NTHU, DROZY and CEW do not represent a wide spectrum of ethnicity groups and lead to biased systems. A population bias visualisation technique was proposed to help identify the regions, or individuals where a model is failing to generalise on a picture grid. Furthermore, the use of Generative Adversarial Networks (GANs) with lightweight convolutions called Depthwise Separable Convolutions (DSC) for image translation to multi-domain outputs was investigated in an attempt to generate synthetic datasets. This thesis further showed that GANs can be used to generate more realistic images with varied facial attributes for predicting drowsiness across multiple ethnicity groups. Lastly, a novel framework was developed to detect bias and correct it using synthetic generated images which are produced by GANs. Training models using this framework results in a substantial performance boost.