The Use of Convolutional Neural Networks for Traffic Incident Detection at a Network Level

Urban road networks are often affected by disruptive traffic anomalies, such as accidents and street works, giving rise to congestion and delays, which can in turn create a wide range of negative impacts to the economy, environment, safety and security. Substantial research has been dedicated into establishing automatic incident detection approches, typically based on traffic flow theory and machine learning techniques. However, most existing studies focus on individual links or linear corridors, ignoring the wider spatio-temporal aspects of incident causation and propagation. Taking these spatio-temporal considerations into account has the potential to improve the coverage and accuracy of incident detection. This paper uses a deep learning technique, Convolutional Neural Network (CNN), to develop an incident detection model that takes account of spatio-temporal network and traffic inherent structure. The CNN model is calibrated based on the simulated Inductive Loop Detectors (ILDs) data and pre-defined labelled incident data, using the Sioux-Falls network as a base. The performance of the CNN is compared with the performance of conventional link-based approaches, implemented using Multilayer Perceptron, Random Forest and Gradient Boosted Tree under a range of incidents. The comparison is made using evaluation criteria based on confusion matrix (including Detection Rate and False Positive Rate), Receiver Operating Characteristic (ROC) curve, Area under the Curve (AUC) and Mean Time to Detection. The result shows that the detection accuracy of the CNN is generally superior to that of conventional alternatives.

  • Supplemental Notes:
    • This paper was sponsored by TRB committee ABJ70 Standing Committee on Artificial Intelligence and Advanced Computing Applications. The Use of Convolutional Neural Networks for Incident Detection at a Network Level: This is an alternate title.
  • Authors:
    • Zhu, Lin
    • Guo, Fangce
    • Krishnan, Rajesh
    • Polak, John W
  • Conference:
  • Date: 2018


  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; References; Tables;
  • Pagination: 14p

Subject/Index Terms

Filing Info

  • Accession Number: 01657942
  • Record Type: Publication
  • Report/Paper Numbers: 18-00321
  • Files: TRIS, TRB, ATRI
  • Created Date: Jan 8 2018 10:06AM