Highlights Introduce a GAN model to generate traffic data related to crashes for data balance. Use the CNN model to incorporate the correlations between variables. Develop various models to compare oversampling and undersampling data.

    Abstract Real-time crash prediction is essential for proactive traffic safety management. However, developing an accurate prediction model is challenging as the traffic data of crash and non-crash cases are extremely imbalanced. Most of the previous studies undersampled non-crash cases to balance the data, which may not capture the heterogeneity of the full non-crash data. This study aims to use the emerging deep learning method called deep convolutional generative adversarial network (DCGAN) model to fully understand the traffic data leading to crashes. With the full understanding of the traffic data of crashes, the DCGAN model could generate more synthetic data related to crashes to balance the dataset. All non-crash data could be used for developing the prediction models. To capture the correlations between different variables, the data are augmented to 2-D matrix as the input for the DCGAN model. The suggested model is evaluated based on data from expressways and compared to two counterparts: (1) synthetic minority over-sampling technique (SMOTE); (2) random undersampling technique. The results suggest that the DCGAN could better understand the crash data characteristics by generating data with better fit of the real data distribution. Four different crash prediction algorithms (i.e., logistic regression model, support vector machine, artificial neural network, and convolutional neural network) are developed based on each balanced data and totally twelve models were estimated. The results indicate that the convolutional neural network model based on the DCGAN balanced data could provide the best prediction accuracy, validating that the proposed oversampling method could be used for the data balance. Besides, compared to other two models, only the DCGAN-based model could identify the significant effects of speed difference between the upstream and downstream locations which could help guide traffic management strategies. With the prediction model developed based on the balanced data by DCGAN, it is expected that more crashes could be predicted and prevented with more appropriate proactive traffic safety management strategies such as Variable Speed Limits (VSL) and Dynamic Message Signs (DMS).


    Zugriff

    Zugriff prüfen

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    Real-time crash prediction on expressways using deep generative models


    Beteiligte:
    Cai, Qing (Autor:in) / Abdel-Aty, Mohamed (Autor:in) / Yuan, Jinghui (Autor:in) / Lee, Jaeyoung (Autor:in) / Wu, Yina (Autor:in)


    Erscheinungsdatum :

    2020-06-01




    Medientyp :

    Aufsatz (Zeitschrift)


    Format :

    Elektronische Ressource


    Sprache :

    Englisch