An innovative approach for traffic crash estimation and prediction on accommodating unobserved heterogeneities

Since traffic crashes involve complex interactions among drivers, vehicles, roadway, traffic, and environmental elements and not all of the factors that could potentially determine the occurrences of traffic crashes can be observed and measured, new methods are needed to better perform traffic crash estimations and predictions and address the unobserved heterogeneity issues in crash data. Unlike the conventional methods, which generally are the statistical models with the observed crash counts as the dependent variables and the factors affecting the likelihood of a traffic crash as the independent variables, a dynamic state-space model with deep learning is proposed to analyze the traffic crashes. The proposed model includes three modules, an unsupervised feature learning module to identify functional network between the explanatory variables and the feature representations, a supervised fine tuning module to perform crash occurrence likelihood estimations, and a dynamic state-space module to perform crash count predictions. A multivariate Tobit model is incorporated in the supervised fine tuning module as the regression layer to account for the heterogeneity issues in correlated crash data. The results of deep learning are fed to the dynamic state-space model that contains a dynamic equation governing the state dynamics to improve the performances of estimation and prediction. The proposed model was applied to the dataset that was obtained from Knox County in Tennessee to validate the model effectiveness and efficiency. The results show that the proposed model has superior performances in terms of estimation and prediction power compared to the SVM and Random Forest (RF) models. The overall performances of the proposed model for all crashes show an 50.559% RMSD improvement over the SVM models and an 57.867% RMSD improvement over the RF models. The findings indicate that the feature learning module identifies relational information between the explanatory variables and feature representations, which reduces the dimensionality of the input and preserves the original information. The proposed model that includes a multivariate Tobit regression layer in the supervised fine tuning module can better account for differential distribution patterns in traffic crashes across injury severities and provides superior crash occurrence likelihood estimation results. The findings suggest that the proposed model can better address the heterogeneity issues in correlated crash data and is a superior alternative for traffic crash estimations and predictions.