Research on conflict analysis and prediction of unsignalized intersections based on multiple alternative safety measures

Unsignalized intersections, due to the frequent interactions between traffic participants, have become high-risk areas for road safety management. In recent years, the continued growth in the number of non-motorized vehicles (NMV), particularly those on electric vehicles, has exacerbated safety issues on these roads. This study, focusing on a typical unsignalized intersection, used drone aerial photography to capture high-angle video data. Combined with computer vision-based trajectory extraction techniques, the study employed multiple alternative safety metrics to quantitatively assess traffic conflicts at the intersection. Subsequently, the machine learning-based traffic conflict prediction model was constructed using the identified conflict events as a sample. The results showed that potential conflicts were primarily concentrated at intersection entrances and exits, with right-turn exits being the most frequent. The conflict risk and severity of non-motor vehicles interacting with motor vehicles were significantly higher than those of other traffic participants, and the rate of both parties yielding in high-risk situations was lower. Among the five prediction models, XGBoost performed best in terms of accuracy (83.6%), precision, and generalization performance. These results demonstrate that the conflict identification framework based on drone video and machine learning can effectively support proactive safety assessments at unsignalized intersections, providing data support for traffic management departments to formulate safety interventions and optimize road infrastructure design.