Scholarlink

<p>With the continuous growth of the demand for container transportation, the proportion of container trucks passing through ports and surrounding roads has increased significantly. Due to their large size and poor maneuverability, once a truck accident occurs, it is often accompanied by serious casualties and property losses. Current research on traffic conflicts for container trucks is limited by the lack of high-quality data: first, publicly available container truck trajectory datasets are extremely rare; second, although drones have the ability to collect data over a large range, their shooting data have problems such as limited accuracy and discontinuous trajectories, which make it difficult to meet the high requirements of micro-modeling for data quality. This problem directly restricts the accuracy of conflict prediction and the credibility of causal analysis. To improve the accuracy and completeness of trajectory data, we introduce a trajectory reconstruction method to repair and complete the original trajectory. The experimental results show that the reconstructed trajectory is significantly better than the original data in terms of continuity and rationality. On this basis, a two-dimensional time to collision (2D-TTC) indicator was constructed to identify side-swipe conflict events, and based on the extraction of micro-behavior features, sample sets of sideswipe and rear-end conflicts were constructed, and a variety of machine learning models were introduced to carry out conflict prediction analysis. The results show that the gradient boosting decision tree (GBDT) model performs best in side-swipe conflict prediction, and the extreme gradient boosting (XGBoost) model in rear-end conflict prediction. By introducing the Shapley additive explanation (SHAP) method to improve the interpretability of the model, our analysis shows that the key factors influencing side-swipe conflict are the lateral speed and the average longitudinal speed within 5 s. The lateral speed reflects the lateral deviation of the vehicle, and the average longitudinal speed within 5 s reflects the driving stability and acceleration trend in a short time. The two together determine the lateral controllability of the vehicle in the dynamic process. Rear-end conflict is affected mainly by the change in longitudinal acceleration, revealing the instability of the vehicle during braking and the lack of control over the distance between the vehicle and the preceding vehicle. Finally, the model performance was optimized through feature ablation experiments, where, in the prediction of side-swipe conflicts, the GBDT achieved an accuracy of 0.911 and an area under the receiver operating characteristic curve (AUC) of 0.953.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lateral risk prediction and influencing factor analysis of container trucks based on trajectory reconstruction data

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Zhihao Zhu,
  • Hexuan Liu,
  • Rongjun Cheng

摘要

With the continuous growth of the demand for container transportation, the proportion of container trucks passing through ports and surrounding roads has increased significantly. Due to their large size and poor maneuverability, once a truck accident occurs, it is often accompanied by serious casualties and property losses. Current research on traffic conflicts for container trucks is limited by the lack of high-quality data: first, publicly available container truck trajectory datasets are extremely rare; second, although drones have the ability to collect data over a large range, their shooting data have problems such as limited accuracy and discontinuous trajectories, which make it difficult to meet the high requirements of micro-modeling for data quality. This problem directly restricts the accuracy of conflict prediction and the credibility of causal analysis. To improve the accuracy and completeness of trajectory data, we introduce a trajectory reconstruction method to repair and complete the original trajectory. The experimental results show that the reconstructed trajectory is significantly better than the original data in terms of continuity and rationality. On this basis, a two-dimensional time to collision (2D-TTC) indicator was constructed to identify side-swipe conflict events, and based on the extraction of micro-behavior features, sample sets of sideswipe and rear-end conflicts were constructed, and a variety of machine learning models were introduced to carry out conflict prediction analysis. The results show that the gradient boosting decision tree (GBDT) model performs best in side-swipe conflict prediction, and the extreme gradient boosting (XGBoost) model in rear-end conflict prediction. By introducing the Shapley additive explanation (SHAP) method to improve the interpretability of the model, our analysis shows that the key factors influencing side-swipe conflict are the lateral speed and the average longitudinal speed within 5 s. The lateral speed reflects the lateral deviation of the vehicle, and the average longitudinal speed within 5 s reflects the driving stability and acceleration trend in a short time. The two together determine the lateral controllability of the vehicle in the dynamic process. Rear-end conflict is affected mainly by the change in longitudinal acceleration, revealing the instability of the vehicle during braking and the lack of control over the distance between the vehicle and the preceding vehicle. Finally, the model performance was optimized through feature ablation experiments, where, in the prediction of side-swipe conflicts, the GBDT achieved an accuracy of 0.911 and an area under the receiver operating characteristic curve (AUC) of 0.953.