摘要：

目前自动驾驶技术重点是关注如何主动避免碰撞，然而在面对其他交通参与者入侵而导致不可避免的碰撞事故场景时，预测车辆在不同行驶模式下的碰撞严重程度来降低事故严重程度的研究却很少。为此，提出一种双层Stacking事故严重程度预测模型。基于真实交通事故数据集NASS-CDS完成训练，模型输入为车辆传感器可感知得到的事故相关特征，输出为车内乘员最高受伤级别。在第1层中，通过实验对不同学习器组合进行训练，最终综合考虑预测性能以及耗时挑选K近邻、自适应提升树、极度梯度提升树作为基学习器；在第2层中，为降低过拟合，采用逻辑回归作为元学习器。实验结果表明，该方法准确率达到85.01%，在精确率、召回率和F1值方面优于其他个体模型和集成模型，该预测结果可作为智能车辆决策规划模块先验信息，帮助车辆做出正确的决策，减缓事故损害。最后阐述了模型在L₂辅助驾驶与L₄自动驾驶车辆中的应用，在常规车辆安全防护的基础上进一步提升车辆的安全性。

关键词: 交通安全, 交通事故严重程度预测, 智能车辆, 集成学习, K近邻, 自适应提升树, 极度梯度提升树, 逻辑回归

Abstract:

At present, autonomous driving technology focuses on how to proactively avoid collisions. However, when it comes to scenarios of inevitable collisions caused by the intrusion of other traffic participants, few studies have explored how to reduce the severity of the accidents by predicting such severity under different vehicle driving modes. To address this issue, a two-layer stacking accident severity prediction model is proposed; it was trained on the NASS-CDS dataset, which is composed of real-world accident data. The model takes accident-related features that can be perceived by vehicle sensors as input and outputs the highest level of injury to passengers. In the first layer, different learner combinations are trained through experiments, and K-Nearest Neighbor(KNN), adaptive boosting, and extreme gradient boosting are eventually selected as base learners considering both predictive performance and algorithm time consumption. In the second layer, logistic regression is selected as a meta learner to reduce overfitting. Experimental results show that the classification accuracy of the proposed model reaches 85.01%, which is superior to other individual and integrated models in terms of precision, recall, and F1 value. The predicted results can be used as a priori information for the decision-planning module of intelligent vehicles to help vehicles make correct decisions and mitigate accident damage. Finally, the application of the model in L₂ assisted driving and L₄ autonomous driving vehicles is reported in detail. The proposed approach further improves the safety of vehicles based on conventional protection.

Key words: traffic safety, traffic accident severity prediction, intelligent vehicle, ensemble learning, K-Nearest Neighbor(KNN), Adaptive Boosting(AdaBoost) tree, eXtreme Gradient Boosting(XGBoost) tree, logistic regression