A crash prediction method based on bivariate extreme value theory and video-based vehicle trajectory data

Area of Science:

• Traffic Safety Engineering
• Extreme Value Theory
• Computer Vision

Background:

• Traditional crash prediction models often require extensive historical data and suffer from poor data quality.
• Existing methods may not adequately capture the complex interactions leading to traffic accidents.
• There is a need for robust crash prediction tools applicable even with limited historical crash data.

Purpose of the Study:

• To propose and validate a novel crash prediction method using a bivariate extreme value theory (EVT) framework.
• To assess the efficacy of different conflict metrics derived from vehicle trajectory data for crash prediction.
• To compare the performance of bivariate EVT models against univariate models for traffic safety evaluation.

Main Methods:

• Utilized unmanned aerial vehicles (UAVs) to collect high-resolution video data from ten intersections.
• Extracted vehicle trajectory data using the Kanade-Lucas-Tomasi (KLT) technique.
• Derived four key conflict metrics: Time-to-accident (TA), Post-encroachment Time (PET), minimum Time-to-collision (mTTC), and Maximum Deceleration Rate (MaxD).
• Developed and applied univariate and bivariate EVT models to estimate crash probabilities and predict annual crash frequencies.

Main Results:

• Bivariate EVT models demonstrated superior performance compared to univariate models for both angle and rear-end crash predictions.
• The Time-to-accident (TA) metric was identified as a crucial factor for inclusion in bivariate EVT models.
• Optimal conflict metrics for predicting angle and rear-end crashes were found to differ, highlighting the need for context-specific analysis.

Conclusions:

• The proposed bivariate EVT framework offers a promising approach for traffic safety evaluation, particularly in scenarios with limited historical crash data.
• Integrating multiple conflict metrics within a bivariate EVT model enhances prediction accuracy.
• The study provides valuable insights into the selection of appropriate conflict metrics for different crash types.