基于图神经网络轨迹预测的合流区交通冲突预测方法
A traffic conflict prediction method for merging areas based on trajectory pre-diction with graph neural network
赵涛 1张宁 2王小超 3马川义 2田源 1张圣涛 2杨梓梁1
作者信息
- 1. 山东大学齐鲁交通学院,山东济南 250002
- 2. 山东高速集团有限公司,山东济南 250014
- 3. 济南新旧动能转换起步区管理委员会,山东济南 250000
- 折叠
摘要
为保证高速公路合流区路段的交通安全,减少交通冲突,提出一种基于图神经网络轨迹预测的合流区交通冲突预测方法.该方法包括基于时空图卷积神经网络的轨迹预测方法以及基于预测轨迹的交通冲突预测方法.利用Mirror-Traffic数据集进行交通冲突指标阈值计算,并通过一定的轨迹数据处理方法得到适用的数据,进行网络模型训练和验证.结果表明,严重冲突的后侵入时间(post-encroachment time,PET)阈值为2.0 s,轻微冲突的PET阈值为5.36 s.该轨迹预测方法的平均位移误差为1.5 m,最终位移误差为2.1 m,时间成本为0.59 s,与其他4种方法相比,本研究方法的轨迹预测整体效果最好.在交通冲突预测方面,采用准确率、精确率、召回率和F1 评价交通冲突预测模型,结果表明交通冲突预测效果较好.本研究提出的方法保证了预测的正确性,增强了预警系统下行车的安全性,提高了预警系统下的合流区通行效率.
Abstract
To ensure the traffic safety and reduce traffic conflicts in this section of the highway,a traffic conflict prediction method was proposed for the merging area based on the trajectory prediction with graph neural network.The method included a trajectory prediction method based on spatio-temporal graph convolutional neural network and a traffic conflict prediction method based on pre-dicted trajectories.The Mirror-Traffic dataset was utilized for traffic conflict indicator threshold calculation.Applicable data were ob-tained through certain trajectory data processing methods for network model training and validation.The results showed that the PET threshold for severe conflicts was 2.0 s and that for minor conflicts was 5.36 s.The trajectory prediction method had an average dis-placement error of 1.5 m,a final displacement error of 2.1 m,and a time cost of 0.59 s.Compared with the other methods,the trajec-tory prediction of the proposed method had the best overall effect.For traffic conflict prediction,accuracy,precision,recall and F1 were used to evaluate the traffic conflict prediction model,and the results showed that the traffic conflict prediction was effective.The proposed method ensured the correctness of the prediction,enhanced the safety of traveling under the warning system,and improved the efficiency of the merging area under the warning system.
关键词
图神经网络/轨迹预测/交通冲突预测/合流区/交通安全Key words
graph neural network/trajectory prediction/traffic conflict prediction/merging area/traffic safety引用本文复制引用
基金项目
国家自然科学基金(52002224)
山东省重点研发计划重大科技创新工程项目(2020CXGC010118)
出版年
2024
NETL
NSTL
万方数据