Advanced traffic conflict analysis for safety evaluation at roundabouts under mixed traffic using extreme value...

Roundabout safety evaluation in non-lane-based, heterogeneous traffic conditions in low-middle-income countries brings challenges due to unavailable/unreliable crash data, thereby switching to the utilization of safety surrogates. This study employed high-resolution drone videos and advanced image-processing techniques to extract vehicular trajectories. Traffic conflicts were identified using Surrogate Safety Measures (SSMs), namely, Time-to-Collision (TTC) and maximum post-collision (hypothetical) velocity difference (MaxDeltaV). Extreme Value Theory (EVT) was used to determine threshold values for TTC (1.25 s) and MaxDeltaV (4.5 m/s). By employing the determined thresholds in the Surrogate Safety Assessment Model (SSAM) tool developed by the Federal Highway Administration (FHWA), conflicts were classified into lane-change (42 %), rear-end (33 %), and crossing (25 %) types. An AdaBoost regressor model was developed using a negative binomial objective function for conflict frequency prediction. SHAP analysis revealed that increased circular road widths reduced conflict frequency, while higher conflicting and approaching traffic volumes increased the likelihood of conflicts. Based on TTC and MaxDeltaV values, hierarchical and two-step clustering techniques classified the identified conflicts into four severity levels: low (33.8 %), moderate (34.1 %), high (19.3 %), and very high severity (12.8 %). An Ordered Probit model predicted conflict severity based on geometric, traffic, built-environmental factors, and conflict types, with a prediction accuracy of 88.4 %. Larger central island diameters and circulatory road widths reduced severity, while higher average speeds increased conflict severity. This research presents a novel framework that applies EVT to establish SSM thresholds for unsignalized roundabouts under mixed traffic, integrating advanced statistical and machine learning techniques to assess conflict frequency and severity proactively.