Efficient mapping of crash risk at intersections with connected vehicle data and deep learning models

Abstract

Traditional methods for identifying crash-prone roadways are mainly based on historical crash data. It usually requires more than three years to collect a sufficient amount of dataset for road safety assessment. However, the emerging connected vehicles (CVs) technology generates rich instantaneous information, which can be used to identify dangerous road sections proactively. Information about the identified crash-prone intersections can be shared with the surrounding vehicles via CVs communication technology to promote cautious driving behaviors; in the longer term, such information will guide the implementation of countermeasures to prevent potential crashes. This study proposed a deep-learning based method to predict the risk level at intersections based on CVs data from the Michigan Safety Pilot program and historical traffic and intersection crash data in areas around Ann Arbor, Michigan, USA. One month of data by CVs at intersections were used for analyses, which accounts for about 3%–12% of overall trips. The risk levels of 774 intersections (i.e., low, medium and high risk) are determined by the annual crash rates. Feature extraction process is applied to both CV’s data and traffic data at each intersection and 24 features are extracted. Two black-box deep-learning models, multi-layer perceptron (MLP) and convolutional neural network (CNN) are trained with the extracted features. A number of hyperparameters that affect prediction performance are fine-tuned using Bayesian optimization algorithm for each model. The performance of the two deep learning models, which are black-box models, were also compared with a decision tree model, a white-box type of simple machine learning model. The results showed that the accuracies of deep learning (DL) models were slightly better (both over 90 %) than the decision tree model (about 87 %). This indicated that the DL models were capable of uncover the inherent complexity from the dataset and therefore provided higher accuracy than the traditional machine learning model. CNN model achieves slightly higher accuracy (93.8 %) and is recommended as the classifier to predict the risk level at intersections in practice. The interpretability analysis of the CNN model is conducted to confirm the validity of the model. This study shows that combination of CVs data (V2V and V2I) and deep learning networks (i.e. MLP and CNN used in this paper) is promising to determine crash risks at intersections with high time efficiency and at low CV penetration rates, which help to deploy countermeasures to reduce the crash rates and resolve traffic safety problems.