Abstract: To examine the conflict features and the key impact factors of conflict severity in the weaving area of the cloverleaf interchange of the urban expressway, this paper uses the North Ring Interchange of Chongqing city as the research object, and collected vehicles natural driving videos by UAV (Unmanned Aerial Vehicle). A vehicle target detection and tracking framework based on the YOLOX and Deep-SORT was established, and a total of 10483 vehicle trajectory data were obtained from aerial video. The longitudinal and lateral conflicts were extracted based on the relative positions of the interacting vehicles. The conflict events were categorized as minor, general and severe based on the two- dimensional expansion of the collision time metrics. The assessment of indicators of the multivariate Logistic regression model, Random Forest model and CatBoost model were calculated in consideration of the factors of macro 17 explanatory variables including traffic flow parameters and vehicle micro- motion parameters. The models with better performance for longitudinal and lateral conflict identification were selected, and the key factors affecting the severity of conflicts within the weaving area were further analyzed. The results indicate that lateral conflict is the main type of conflict in the cloverleaf interchange weaving area, and the duration of conflict is longer and the risk of collision is higher than that of longitudinal conflict. The general conflict has the widest distribution in the intersection area, followed by serious conflict, and minor conflict has the most concentrated distribution, which is mainly distributed in the vicinity of merging triangles. There are some differences in the degree of influence of different explanatory variables on the severity of the conflict, longitudinal serious conflict is significantly related to six indicators such as relative speed difference and target vehicle length. For the lateral conflict, the top five variables in terms of importance are the relative speed difference, the average headway spacing, the target vehicle speed, and the target vehicle transverse longitudinal position. The risk of collision increases dramatically when the relative speed is more than 20 km·h^-1 , the speed of vehicle is more than 60 km·h^-1 , and the average headway spacing is below 12 meters.

Key words: traffic engineering, conflict severity, expanded time-to-collision, cloverleaf interchange, vehicle trajectory data

摘要： 为明确城市快速路苜蓿叶立交交织区的冲突特征及冲突严重程度的关键影响因素，以重庆市北环立交为研究对象，使用无人机拍摄交织区车辆自然驾驶视频。搭建基于 YOLOX 和Deep-SORT的车辆目标检测和追踪框架，从航拍视频中获取10483条车辆轨迹数据，根据交互车辆的相对位置提取纵向和侧向冲突样本，并基于二维拓展碰撞时间指标将冲突事件划分为轻微、一般和严重；考虑包括宏观交通流参数和车辆微观运动参数等17个解释变量，计算多元Logistic回归模型、随机森林模型和CatBoost模型的评估指标，并分别选择对纵向冲突和侧向冲突识别性能更优的模型，进一步分析影响交织区内冲突严重程度的关键影响因素。结果表明，侧向冲突是苜蓿叶立交交织区的主要冲突类型，冲突持续时间较纵向冲突更长，碰撞风险更高；一般冲突在交织区内分布范围最广，严重冲突其次，轻微冲突分布最集中，主要分布于合流三角区附近；不同解释变量对冲突严重程度的影响程度存在一定差异，纵向严重冲突与相对速度差和目标车长度等6种指标显著相关，对于侧向冲突，重要性前5的变量分别是相对速度差、平均车头间距、目标车速度、目标车横向位置及目标车纵向位置，且当相对速度超过20km·h^-1，车速超过60 km·h^-1及平均车头间距低于12 m时，侧向碰撞风险将大幅增加。研究成果可对苜蓿叶立交交织区危险驾 驶行为预警系统的设计和主动交通安全管理提供依据。

关键词: 交通工程, 冲突严重程度, 拓展碰撞时间, 苜蓿叶立交, 车辆轨迹数据