Network-level crash risk analysis using large-scale geometry features.

Area of Science:

• Transportation Engineering
• Road Safety
• Machine Learning Applications

Background:

• Road traffic crashes pose significant societal risks and economic losses.
• Identifying crash causes is complex due to interacting factors like drivers, vehicles, roads, and environment.
• Large-scale crash prediction faces challenges in data collection and cost.

Purpose of the Study:

• To develop a cost-effective method for large-scale road network crash risk identification using open-source data.
• To leverage road geometry, specifically horizontal curves (H-curves) and vertical curves (V-curves), for crash risk assessment.
• To introduce an optimized machine learning model for enhanced crash prediction.

Main Methods:

• Feature extraction from horizontal and vertical road curves.
• Development of the Harris Hawks Optimization (HHO) algorithm combined with the XGBoost model (HHO-XGBoost).
• Creation of a specialized road geometry-crash risk dataset for model training and validation.

Main Results:

• The HHO-XGBoost model adaptively identified optimal XGBoost hyperparameters.
• Favorable outcomes were achieved in crash risk prediction using the developed model.
• A tiered risk analysis of "region-road-segment" was successfully completed for large-scale road networks.

Conclusions:

• The proposed method offers a viable approach for large-scale road network crash risk identification.
• The HHO-XGBoost model demonstrates effectiveness in analyzing road geometry for safety.
• The study provides a 3D road geometry database and insights into swarm intelligence for integrated learning models.