Convolutional Neural Networks for Driver Behaviour Prediction: A Comparative Analysis of AlexNet, VGGNet and ResNet (#441)

Authors

Melgarejo-Graciano, Melquiades

Iparraguirre-Villanueva, Orlando

Abstract

Understanding driver activity in real time is complex, yet it is important for in-vehicle systems that aim to reduce car crashes. This work addressed the problem by relying on state-of-the-art methods, specifically, evaluating three convolutional neural network (CNN) models, such as: AlexNet, VGGNet and ResNet with the aim of predicting real-time driver activity and behaviour during driving. For the development, a dataset consisting of 10,751 images was used (The dataset was obtained from the Kaggle platform). To achieve good results, there are multiple factors, such as the volume of the dataset, the quality of the data, the application of optimisation techniques, among others. The findings of the proposal showed the VGGNet model to be the most efficient model for efficiently classifying and predicting driver’s behaviour, achieving an accuracy rate of over 98%. It is closely followed by the AlexNet model, with an accuracy rate of 98%, a very significant result for this type of task, this model also obtained an F1-Score of 86%, and a count rate of 75%. However, the metrics obtained by the ResNet model are much lower than the compassion of the other models, it only achieved an accuracy rate of 35.28%, which indicates that it has limitations in identifying features and predicting driver behaviour. Finally, it is concluded that the VGGNet model slightly outperforms the AlexNet model, which shows that both models are efficient for this type of task