Crash frequency prediction based on extreme value theory using roadside lidar-based vehicle trajectory data.

Area of Science:

• Traffic safety engineering
• Transportation data analytics
• LiDAR technology applications

Background:

• Traditional crash prediction models (CPMs) often rely on unreliable historical crash data.
• Near-crash prediction offers a proactive approach to enhance traffic safety.
• Roadside LiDAR provides high-resolution vehicle trajectory data for detailed movement analysis.

Purpose of the Study:

• To develop a methodology for identifying near-crashes using Roadside LiDAR data.
• To apply surrogate safety measures and extreme value theory (EVT) for crash probability estimation.
• To evaluate the effectiveness of different surrogate measure pairs in predicting crash frequencies.

Main Methods:

• Utilized Roadside LiDAR to collect microscopic vehicle trajectory data.
• Identified near-crashes using surrogate indicators: Time to Collision (TTC), Post Encroachment Time (PET), Anticipated Collision Time (ACT), and Maximum Deceleration (MaxD).
• Applied bivariate extreme value theory (EVT) to combine time-based and evasive-action-based surrogate measures for crash probability prediction.

Main Results:

• The bivariate EVT model showed a better fit for conflict extremes and improved crash frequency prediction compared to univariate models.
• The ACT and MaxD pair demonstrated the highest accuracy in the bivariate model.
• The TTC and MaxD pair effectively reflected relative threat levels at intersections.

Conclusions:

• The proposed methodology enables proactive safety analysis of signalized intersections using Roadside LiDAR data.
• Combining surrogate safety measures with bivariate EVT enhances the accuracy of near-crash prediction.
• This approach provides a foundation for data-driven traffic safety improvements.