Proactive traffic conflict prediction: uncovering multi-dimensional risk mechanisms through spatiotemporal...

Proactive traffic conflict prediction often relies on average-state macroscopic variables (e.g., mean speed, flow, density), which can obscure the underlying traffic flow instability and microscopic heterogeneity that precipitate critical interactions. The objective of this study is to propose and validate a physics-informed feature set that quantifies spatiotemporal heterogeneity to improve the current-window identification and assessment of multidimensional ordinal traffic-conflict risks.

Using 29,848 traffic-state samples (10-second intervals) extracted from the high-precision highD vehicle trajectory dataset, a novel set of instability features was formulated, including spatial gradients of flow and density, speed variation coefficients ($C{V}_v$) and braking intensity exposure ($E_{\text{BRK}}$). To capture the multidimensional nature of risk, conflicts were simultaneously evaluated in three complementary surrogate safety measures: Time-to-Collision (TTC), Deceleration Rate to Avoid Crash (DRAC), and Proportion of Stopping Distance (PSD). These measures were categorized into a four-level ordinal risk framework. A dual-track modeling approach was employed, benchmarking an Ordered Probit Model (OPM) for parametric statistical inference against an XGBoost model for non-linear prediction, augmented by SHapley Additive exPlanations (SHAP) for interpretability.

The incorporation of the proposed spatiotemporal heterogeneity features systematically and significantly improved the predictive accuracy of all models and risk dimensions compared to conventional macroscopic variables. In particular, in the challenging PSD risk task, the XGBoost model achieved a Quadratic Weighted Kappa (QWK) increase of 0.1932 to 0.3331 (a relative improvement of 72.4%). Both OPM statistical tests and SHAP non-linear analyzes consistently identified the speed variation coefficient ($C{V}_v$) and braking intensity exposure ($E_{\text{BRK}}$) as the most dominant predictors of high-risk states, frequently surpassing traditional volume and speed metrics.

The study empirically shows that the genesis of traffic conflicts lies in microscopic instability rather than merely average flow states. The findings elucidate a distinct process-event mechanism in risk evolution: $C{V}_v$ acts as a continuous indicator of the gradual accumulation of traffic flow disorder, while $E_{\text{BRK}}$ serves as an acute trigger signaling an imminent critical state. The proposed multidimensional risk vectors and empirically derived thresholds provide an interpretable basis for future active traffic management, but their operational use in Variable Speed Limit (VSL) control or connected-vehicle applications requires real-time monitoring of heterogeneity features and further validation under lagged prediction settings.