ABSTRACT：Real-time crash risk prediction is an important area of research that focuses on identifying hazardous traffic conditions as part of proactive traffic safety management. Although there is a plethora of classification algorithms applied to predict an unsafe traffic condition, they cannot capture spatio-temporal variability in traffic dynamics and are not transferable. In this paper, a state-of-the-art approach based on supervised machine learning - recurrent neural network (RNN) is developed and implemented to address the challenges of predictability of crash risk models. In relation to existing techniques, one of the unique features of RNN is to employ feedback loops where the output from each of the steps is feedback to the RNN to affect the outcome of the current step. It also has a self-updating ability of model parameter via a time sequence, which is helpful for the model adaptability by overcoming the spatial-temporal variability of traffic dynamics. Historical crash data and real-time traffic data from Shanghai Urban Expressway System were matched and split into a training dataset and a test dataset: the training dataset was designed in the matched case-control study and used to develop the crash risk prediction models; the test dataset was a full set including all cases and employed to evaluate the performance of the models via the area under ROC curve (AUC) and sensitivity. In addition, the prediction results were compared with those given by other frequently used classification algorithms, including logistic regression and support vector machine (SVM). The results proved that RNN had a better prediction performance. It could increase the crash prediction accuracy by an average of 13.3% and 7.9% compared to the SVM and logistic regression model, respectively. Furthermore, the optimal ratio of crashes to non-crashes has found to be 1:4 for the model development.
Kui Yang, Xuesong Wang*, Mohammed Quddus, Rongjie Yu. Predicting Real-Time Crash Risk on Urban Expressways Using Recurrent Neural Network. Transportation Research Board 98th Annual Meeting, Washington D.C., USA, 2019. 1.13-17.