Vehicle crash prediction using transformer networks

Abstract

Road accidents pose a significant threat to human life, causing numerous injuries, fatalities, and economic damage worldwide. Recently, there has been growing interest in leveraging Artificial Intelligence (AI) to create systems that can predict vehicle crashes. This thesis focuses on vehicle crash prediction and aims to develop a solution using pre-trained Convolutional Neural Networks (CNN) and transformer networks to mitigate the occurrence of such accidents. By leveraging advanced deep learning techniques, this research addresses the limitations of traditional crash analysis methods. The objectives of this study involve critically reviewing the evolution of vehicle crash prediction systems, studying modern technologies employed in vehicle crash prediction, designing and implementing a crash prediction system using pre-trained CNN and transformer networks, and evaluating the solution's effectiveness. The motivation behind this research stems from the urgency to reduce the growing frequency of vehicle crashes and their associated consequences. Current approaches to road safety primarily rely on reactive measures, prompting the need for proactive strategies to prevent accidents. The Car Learning to Act (CARLA) simulator was used for data gathering, in that an ego-vehicle attached to RGB and RGB-Depth cameras were used. Ego-vehicle traveled through a simulated environment with different weather conditions and maps and captured a sequence of RGB and RGB-Depth (RGB-D) images in safe and unsafe scenarios. To extract features from RGB and RGB-D images pre-trained CNN were utilized. Four pre-trained CNNs were used for feature extraction. With those extracted features, a transformer network was employed to train a model. After model training and testing, it was observed that the transformer model trained with VGG16-based feature extraction performs better than other methods. This research can be implemented in the real world with real-time predictions and addresses the shortcomings of visual-based systems. Ultimately, it aims to improve road safety by enabling proactive crash prevention and minimizing the human and economic costs associated with accidents.