Trajectory data-driven risk assessment of freeway accidents

DONG Chunjiao; XU Bo; LI Penghui; ZHUANG Yan; YANG Miaoyan

doi:10.11860/j.issn.1673-0291.20240006

Abstract：To address the challenges posed by fully enclosed freeway segments,high vehicle speeds and the substantial damage caused by traffic accidents,this study proposes a freeway accident risk as-sessment method that integrates the Random Forest(RF)algorithm for feature selection with the eX-treme Gradient Boosting(XGBoost)algorithm.First,by filtering private vehicle trajectory data from freeway accident segments,a data foundation for accident risk assessment is established under four dif-ferent spatiotemporal conditions(30 km upstream and 30 minutes before the accident,10 km upstream and 15 minutes before the accident,10 km upstream and 10 minutes before the accident,and 10 km upstream and 5 minutes before the accident).Next,a combined accident risk assessment method based on the RF and XGBoost is constructed.It evaluates accident risk after selecting various operational in-dicators for vehicles on the freeway.Finally,the algorithm's performance is assessed using five met-rics:accuracy,precision,recall,balanced F Score(F1),and Area Under Curve(AUC).Results indi-cate that the RF-XGBoost combination algorithm outperforms the Decision Tree(DT),Support Vec-tor Machine(SVM),and traditional XGBoost algorithms in accident risk assessment.Compared to the traditional XGBoost algorithm,the average accuracy of the RF-XGBoost algorithm is increased by 11.1%,the average precision is increased by 8.9%,and the average recall rate is increased by 7.625%.Under the spatiotemporal condition of 10 km upstream and 10 minutes before the accident,the algorithm achieves an accuracy of 80%,demonstrating optimal overall assessment performance.These findings provide theoretical and methodological support for freeway accident risk assessment and dynamic warnings for private vehicles.