Investigating spatiotemporal traffic dynamics toward conflict risk levels using trajectory data in heterogeneous...

This study aimed to identify dynamic spatiotemporal traffic factors influencing conflict risk levels on National Highways under heterogeneous traffic conditions in India. The research addresses gaps by capturing vehicle interactions using high-resolution UAV-based trajectory data and proposes a novel two-stage methodology for real-time conflict risk evaluation, moving beyond traditional binary risk classifications to a four-level framework (High, Moderate, Low, No-Risk).

Over 40,000 conflict risk sequences were classified into four severity levels using the Modified Time-to-Collision (MTTC) surrogate safety measure. Lane-specific spatiotemporal variables including vehicle dynamics (speed and acceleration variability, traffic flow, traffic density, proportion of heavy vehicles, and lane-change frequency) were systematically extracted up to five seconds preceding each conflict event, resulting in a comprehensive dataset of 65 variables. In the first stage, influential variables were identified using an aggregated normalized feature importance analysis from Random Forest, XGBoost, and LightGBM models. The second stage employed an ordered probit model to statistically evaluate the causal impact of selected critical variables on the categorized risk levels.

Incremental feature importance analyses revealed that immediate vehicle dynamics, particularly speed and acceleration variability within one second before the event, were the strongest predictors of conflict risk. Variables such as lane-change frequency and adjacent-lane traffic flow became influential over the broader five-second timeframe. The ordered probit model further quantified these relationships: each unit increase in speed variability and lane-change frequency elevated the high-risk probability by approximately 10%, whereas higher traffic density significantly reduced high-risk probability. A one-unit increase in traffic flow within adjacent lanes was associated with a 3.6% reduction in the likelihood of high-risk events, suggesting spatial interaction effects on safety.

The study introduces a robust methodological approach that combined advanced machine learning and statistical modeling to understand complex spatiotemporal dynamics influencing traffic conflict risks under mixed traffic conditions. By capturing these interactions at fine resolution and demonstrating their impact on conflict severity, this research offers significant implications for developing real-time risk-alert systems. Such systems can proactively warn drivers, enabling safer and more informed driving decisions, and thus enhancing overall highway safety in heterogeneous traffic environments.