Machine learning in traffic safety: Techniques for injury severity prediction

Abstract

In this research, we look at how deep learning and TL models may be used to forecast how bad traffic crash will be. It is critical to create trustworthy crash severity prediction models since road crash are on the rise and have major social and economic consequences. Research in this area makes use of CAS crash data and employs a number of deep learning and transfer learning models, including ResNet, EfficientNetB4, InceptionV3, Xception, and MobileNet, as well as a number of convolutional neural networks (CNNs), multilayer perceptrons (MLPs), and long short-term memories (LSTMs). MobileNet achieved the highest performance metrics, including an F1-score of 98.9%, a precision of 98.5%, and an accuracy of 98.2%. The model's predictions were further analyzed using SHapley Additive exPlanations (SHAP) to determine the most important components that contributed to the severity of the disaster.   Findings show that MobileNet is very effective when it comes to crash severity prediction using transfer learning's strong and generalizable architecture