区分冲突类型的路段实时碰撞风险预测模型

使用交通数据建立路段实时碰撞风险预测模型(RTCPM)是主动交通安全管理的基础,路侧精细感知的行车数据和替代安全指标(SSMs)在RTCPM领域有着潜在价值.基于此,采用路侧精细感知数据生成SSMs作为输入,提出一种区分冲突类型的路段实时碰撞风险预测模型.以路段精细交通数据为基础,提取多种类别的交通参数以构建包含多类交...

Real-time Crash-risk Prediction Model That Distinguishes Collision Types

Using traffic data to establish a real-time crash-risk prediction model (RTCPM) is the basis of active traffic-safety management. Vehicle data and surrogate safety measures (SSMs) extracted from roadside sensors have potential value in the RTCPM field. This study used high-precision roadside data to generate SSMs as inputs, and proposed an RTCPM for road sections that distinguishes collision types. Based on the traffic data of the studied road section, the traffic parameters were extracted to construct a detailed database. These parameters included the vehicle-motion parameters and the SSMs. A traffic-conflict extraction method, based on vehicle-avoidance behavior and spatiotemporal proximity, was developed to obtain lateral and longitudinal traffic conflicts. These labeled traffic-conflict events were used as the type labels for samples in the modeling. The extreme gradient boosting algorithm (XGBoost) was used for RTCPM. The edited nearest neighbor (ENN) method was adopted to eliminate sample-size imbalances, and the Shapley additive explanations (SHAP) method was employed to explain the contribution of the model features. The traffic parameters before a traffic conflict occurred were aggregated with a 30 s-time window as the sample feature, and input into the XGBoost model for training and testing. The established XGBoost model predicts the collision risk and type 30 s before the collision. The model achieves an overall accuracy of 97.4%, predicting 93.0% of longitudinal conflicts with a false-positive rate of 0.13%, and 61.8% of lateral conflicts with a false-positive rate of 0.12%. The interpretation results of the SHAP model show that SSMs play an important role in prediction. The 5%-quantile 1/modified time to collision (MTTC) has the greatest impact on the longitudinal-conflict prediction, while the average traffic flow and acceleration are the most important features for lateral-conflict prediction. The proposed model framework can provide a basis for active traffic management in the affected area of an interchange.