Real-time Secondary Crash Prediction

A new study presents a system for real-time prediction of secondary crash likelihood on freeways. The key feature of this system is that it does not rely on data related to the primary crash (crash type, severity, etc.), which is often not immediately available.

Methodology

The proposed framework uses a hybrid approach. A dynamic spatiotemporal window system extracts real-time data on traffic flow and environmental conditions from the location of the primary crash and upstream road segments. The system includes three models:

  • A model to estimate the likelihood of the primary crash.
  • Two models to evaluate traffic conditions at the crash location and upstream segments, in different comparative scenarios.

To improve predictive performance, an ensemble learning strategy integrating six machine learning algorithms was implemented. A voting-based mechanism combines the outputs of the three models.

Results

Tests on Florida freeways demonstrated that the framework correctly identifies 91% of secondary crashes, with a low false alarm rate of 0.20. The Area Under the ROC Curve (AUC) improved from 0.654, 0.744, and 0.902 for the individual models to 0.952 for the hybrid model, outperforming previous studies.