A Novel Bayesian Convolutional Neural Network with Variational Dropout in Crash Severity Estimation

Due to high socio-economical costs of traffic crashes, accident injury prediction is an important aspect in safety related studies. Although statistical methods have been traditionally applied in crash severity prediction, there are some limitations regarding their application, hence limiting their usage in various tasks? With the aim of overcoming such limitations, Machine Learning methods have been developed and successfully applied in numerous fields including traffic safety. Deep Neural Networks are a field of research in Machine Learning which have been able to achieve state of the art performance in variety of fields. However, their usage has been limited in safety critical fields due to number of drawbacks. The current study addresses two of such drawbacks; namely, their inability to provide uncertainty measures in their predictions and time consuming hyperparameter optimization, due to high computational complexity. Hence, the current study proposed a novel approach to Deep Neural Networks. The proposed LRT-CNN-VD model aims to provide uncertainty measures using Bayesian Neural Networks, and simultaneously provides a Bayesian justification for Dropout regularization which renders grid search methods for regularization’s hyperparameter tuning useless. In order to evaluate our model’s performance, comparisons are made between the proposed model and two Convolutional Neural Network models and a binary Logistic Regression model. Results show improvement across different evaluation metrics, and regularization performance on par with dropout regularization.