LRT-CNN-VD Model in Crash Duration Prediction

Abstract

Traffic crashes significantly contribute to non-recurrent congestion, resulting in lower levels of service in transportation networks. Accurate predictions of crash clearance times could significantly enhance safety and increase efficiency, but prediction is challenging because of the stochastic nature of crashes. This study applies an LRT-CNN-VD model—a deep Bayesian convolutional neural network using the local reparameterization trick, variational inference, and variational dropout—to train efficiently while reducing overfitting. The model was tested on the Pennsylvania crash dataset and compared with two other deep-learning models and two machine-learning algorithms. The LRT-CNN-VD model outperformed the alternatives across the reported metrics, showed comparable overfitting control to regular dropout, and substantially reduced regularization hyperparameter tuning and training time.