Applying Deep Learning and High-Resolution Data to Predict Crashes for Transportation Safety Planning

In the context of transportation safety planning (TSP), it is essential to accurately predict crashes. In the recent decade, considerable efforts have been devoted to providing better prediction results with the consideration of zonal systems, mathematical methods, input variables, etc. The previous prediction methods required all input variables to be aggregated at the zonal level. With the aggregation process, the collected data fell into low resolution and lost details, which may introduce low accuracy and even biases. To improve the prediction performance, this study suggests applying a deep learning approach to predict crashes for TSP using high-resolution data. A framework of collecting high-resolution data is first introduced. Then, a deep learning architecture of a convolutional neural network (CNN) is adopted to predict traffic crashes by using the collected data. To validate the proposed method, an empirical study is conducted and the proposed method is compared with two counterparts: a statistical model (i.e., negative binomial model) and a traditional machine learning model (i.e., artificial neural network) using low-resolution data (i.e., data that are aggregated based on zones). The results indicate that the proposed deep learning method with high-resolution data could provide significantly higher prediction accuracy than the two conventional models using low-resolution data, which validates the concept of using detailed data for traffic crash prediction. It is expected that the deep learning approach for traffic crash prediction in this study could provide new and valuable insights into future directions of transportation safety planning.

  • Supplemental Notes:
    • This paper was sponsored by TRB committee ABJ70 Standing Committee on Artificial Intelligence and Advanced Computing Applications.
  • Corporate Authors:

    Transportation Research Board

  • Authors:
    • Cai, Qing
    • Abdel-Aty, Mohamed
    • Sun, Yangyang
    • Lee, Jaeyoung
    • Yuan, Jinghui
  • Conference:
  • Date: 2019


  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; Tables;
  • Pagination: 6p

Subject/Index Terms

Filing Info

  • Accession Number: 01697452
  • Record Type: Publication
  • Report/Paper Numbers: 19-02685
  • Files: TRIS, TRB, ATRI
  • Created Date: Dec 7 2018 9:27AM