AI Conflict Observer: Conflict severity and scenario identification for intersection based on Ensemble Transformer.

The application of perception and edge computing technologies provides extensive data support for intersection conflict analysis. However, traditional threshold-based and semantic rule-based conflict analysis methods struggle to address the challenges posed by intersection heterogeneity and data diversity. This study integrates the kinematic features of trajectory data from eight different Holographic Intersections. Based on the proposed definitions for conflict severity and scenarios, 5,339 typical conflict events were labeled as the training dataset. Subsequently, four Transformer encoders with different combinations of heads and layers were trained, then integrated using the Weighted Voting method. The Ensemble Transformer was selected as the benchmark to construct the AI Conflict Observer (AICO) model. For the tasks of classifying conflict severity and scenarios, the Weighted F1 Scores of AICO reached 0.846 and 0.902, respectively. To validate the model's generalization performance, this study conducted case studies using conflict events from a 4-leg intersection and a 3-leg intersection sourced from different datasets as ground truth. A total of 560 and 136 conflict events were identified through threshold-based preliminary screening and manual verification, respectively. The conflict recognition results of the AICO model were then compared with the ground truth. The results indicate: 1) At the 4-leg intersection and the 3-leg intersection, AICO achieved classification accuracies of 89.97% and 86.13% for conflict severity, and 88.71% and 90.97% for conflict scenario, respectively. 2) Regarding spatial distribution, the kernel density values of conflicts identified by AICO were highly consistent with the ground truth. 3) For temporal distribution, AICO's results demonstrated a high degree of goodness-of-fit. The R2 values for common and serious conflicts were 0.927 and 0.987 at the 4-leg intersection, and 0.934 and 0.945 at the 3-leg intersection, respectively. 4) In terms of conflict severity distribution, there was no significant difference between AICO and the ground truth in identification of TTC. However, significant differences were observed in conflict duration identification (p < 0.05). AICO employed a more conservative strategy, tending to assign longer durations. The AICO model overcomes the limitations of traditional threshold and rule-based methods. It can generalize to different types of signalized and unsignalized intersections and achieve batch conflict identification. The model can be applied to near-real-time analysis of intersection operational risks, extraction of critical risk scenarios, providing decision support for intersection improvement, governance, and management.