A Bayesian dynamic updating approach for urban expressway real-time crash risk evaluation

Real-time crash risk evaluation is an emerging analysis topic which can identify crash-prone traffic conditions and further benefit the implementation of the active traffic safety management systems on urban expressways. Studies have been conducted to develop crash risk evaluation models with different modeling techniques. However, they were developed and calibrated with fixed model parameters, which cannot account for different traffic patterns with the spatial-temporal variability, and the on-line model updating issue with in-field streaming traffic data has not been well analyzed. In this study, the Bayesian dynamic logistic regression (Bayesian Dynamic LR) was introduced to develop the real-time crash risk evaluation model, where the model parameters could dynamically change by efficiently integrating a new instance with prior knowledge when it arrives. Firstly, a Bayesian Dynamic LR model was developed and compared with a standard logistic regression (Standard LR) model and four support vector machine (SVM) models with different kernel functions for the prediction accuracy evaluated by the Area under the Receiver Operating Characteristics curve (AUC). Then, the sensitivity analysis of the Bayesian Dynamic LR model was conducted on its two key parameters, which are forgetting parameter λ and the ratio of crash to non-crash observations. The analyses results concluded as (i) Bayesian Dynamic LR model has by far the best predictive performance, where AUC could reach around 0.987 which is also the very highest in existing related literatures to the best of the authors' knowledge; (ii) using a less informative prior in the updating equation stage, makes the model performance better; (iii) when the crash to non-crash ratio increases from 1:1 to 1:10, the model performance is gradually enhanced and it tends to be stable at the AUC level of 0.971. Finally, an on-line updating modeling framework was proposed from the model implementation perspective and its prediction performance was proved by a full set in real-world, and the implementation challenges in real-world was discussed.


  • English

Media Info

Subject/Index Terms

Filing Info

  • Accession Number: 01682870
  • Record Type: Publication
  • Files: TRIS
  • Created Date: Oct 2 2018 3:04PM