Real-time crash risk prediction on arterials based on LSTM-CNN

• Published in: Accident analysis and prevention Vol. 135; p. 105371
• Main Authors: Li, Pei, Abdel-Aty, Mohamed, Yuan, Jinghui
• Format: Journal Article
• Language: English
• Published: England Elsevier Ltd 01-02-2020
• Subjects: Accidents, Traffic - prevention & control; Bayes Theorem; Deep learning; Humans; Logistic Models; Neural Networks, Computer; Real-time crash risk; Recurrent neural network; Urban arterials; Real-time crash risk; Urban arterials; Deep learning; Recurrent neural network
• ISSN: 0001-4575; 1879-2057
• DOI: 10.1016/j.aap.2019.105371

•A LSTM Convolutional Neural Network (LSTM-CNN) network is proposed to predict crash risk in real-time on arterials.•The possibilities of using various data sources for real-time crash prediction are explored. Such as Bluetooth data, detector data and weather data. One year’s data are analyzed extensively. Different data preparation techniques are used.•Synthetic minority over-sampling technique (SMOTE) is utilized for re-sampling the extremely imbalanced crash dataset.•The performance of LSTM-CNN is compared with other common approaches on the same dataset. Results suggest that the LSTM-CNN outperforms the others with various evaluation metrics, i.e., Area Under the Curve (AUC), sensitivity and false alarm rate.
Real-time crash risk prediction is expected to play a crucial role in preventing traffic accidents. However, most existing studies only focus on freeways rather than urban arterials. This paper proposes a real-time crash risk prediction model on arterials using a long short-term memory convolutional neural network (LSTM-CNN). This model can explicitly learn from the various features, such as traffic flow characteristics, signal timing, and weather conditions. Specifically, LSTM captures the long-term dependency while CNN extracts the time-invariant features. The synthetic minority over-sampling technique (SMOTE) is used for resampling the training dataset. Five common models are developed to compare the results with the proposed model, such as the XGBoost, Bayesian Logistics Regression, LSTM, etc. Experiments suggest that the proposed model outperforms others in terms of Area Under the Curve (AUC) value, sensitivity, and false alarm rate. The findings of this paper indicate the promising performance of using LSTM-CNN to predict real-time crash risk on arterials.