A novel deep ensemble based approach to detect crashes using sequential traffic data

Abstract

Crash is one of the leading causes of death in the United States. Real time detection of crashes plays a pivotal role in increasing safety of highways. In this study, a deep ensemble modelling approach is proposed in which we first employed three powerful deep learning techniques, Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs), and Deep Neural Network (DNN), then these three models are combined using eight ensemble techniques to detect crashes in real time. Loop detectors' traffic condition, crash information, and weather condition are the main sources of data used in this study. In addition, since the dataset of this study includes 241 crash and 6038 non-crash cases, Synthetic Minority Over-sampling Technique (SMOTE) is used to overcome problem of imbalanced data before training the models. The results show that while deep learning models are performing well in detecting crashes, ensemble of these three deep learning models using Multilayer Perceptron (MLP) and Random Forest Classifier (RFC) can improve detection performance. Interestingly, MLP and RFC ensemble models achieved the highest detection rate and lowest false alarm rate values, respectively, among all the studied models. Comparing the models regarding area under curve (AUC) of ROC curve also shows that the best five models are MLP ensemble, RFC ensemble, DNN, GRU, and LSTM, respectively, with AUC of 97.2%, 95.2%, 93.7%, 91.8%, and 91.3%.

Keywords

• Crash detection
• Long short-term memory (LSTM)
• Gated recurrent unit (GRU)
• Deep neural network (DNN)
• Ensemble model