Deep learning for real-time crash prediction on urban expressways

ABSTRACT:Real-time crash prediction is an important area of research focusing on predicting crash-prone traffic conditions, and has strong implication to active traffic safety management systems. One of the challenging aspects of real-time crash prediction relates to the data imbalance problem arising from the inappropriate mixture of traffic conditions relating to crash and non-crash cases. Existing studies addressed this problem by employing approximately balanced samples in developing crash prediction models without considering the entire population. This often resulted in biased models and even erroneous conclusions. Additionally, previous work mainly uses shallow crash prediction models, which are still unsatisfying for the real-world application with the imbalanced big data. This study therefore structured the full set and adopted a novel Deep Learning methodology—Deep Neural Network (DNN) for model development. Historical crash data and real-time traffic data from two expressways in Shanghai were combined and divided into training data and test data, with both datasets including all crash and non-crash traffic conditions. Training data was utilized to develop DNN models and validate their performance with the help of K-fold cross-validation. Test data was used to evaluate the performance of the models with several evaluation criteria. The results indicated that the DNN technique has powerful identification ability in real-time crash prediction: 63%-65% of crashes can be identified at the cost of just a 5% false alarm rate. In addition, class balancing for imbalanced training data may reduce the adaptability of the DNN model, and further weaken its prediction performance before retraining by new data. With the size increase of class balanced training data, the difference between training data and imbalanced data becomes larger and further degrades the prediction performance of the DNN model.

Kui Yang, Xuesong Wang*, Mohammed Quddus, Rongjie Yu. Deep learning for real-time crash prediction on urban expressways. Transportation Research Board 97th Annual Meeting, Washington D.C., USA, 2018. 1.7-11.