交通运输系统工程与信息

交通运输系统工程与信息 ›› 2023, Vol. 23 ›› Issue (5): 67-74.DOI: 10.16097/j.cnki.1009-6744.2023.05.007

• 智能交通系统与信息技术 • 上一篇    下一篇

基于时间图注意力的交叉口交通状态识别及关联度研究

李鹏程，董宝田*，李思贤   

  1. 北京交通大学，交通运输学院，北京 100044
  • 收稿日期:2023-06-02 修回日期:2023-06-25 接受日期:2023-07-17 出版日期:2023-10-25 发布日期:2023-10-22
  • 作者简介:李鹏程(1991- )，男，山东潍坊人，博士生。
  • 基金资助:
    国家自然科学基金(61772065)。

Intersection Traffic State Recognition and Correlation Degree Study Based on Temporal Graph Attention

LI Peng-cheng,DONG Bao-tian*,LI Si-xian   

  1. School of Traffic and Transportation, Beijing Jiaotong University, Beijing 100044, China
  • Received:2023-06-02 Revised:2023-06-25 Accepted:2023-07-17 Online:2023-10-25 Published:2023-10-22
  • Supported by:
    National Natural Science Foundation of China (61772065)。

PDF

摘要： 为实时获取交叉口交通状态及其邻居交叉口关联度，本文利用实际交通调查数据建立VISSIM仿真环境，基于仿真数据提出包含交叉口交通流特征和物理特征的交叉口交通特征矩阵，并根据车流量、流向比和有效绿灯时间得到交叉口的交互时间矩阵。提出基于时间图注意力(Temporal Graph Attention, TGAT)网络的交叉口交通状态识别模型，将上述矩阵和交通数据的初始标签输入模型中，得到目标交叉口在畅行、平稳、拥挤、阻塞这4种状态下的分类准确率，并得到邻居权重矩阵作为交叉口之间的关联度。选取本文方法和多层感知机、长短期记忆网络、支持向量机分别进行交叉口交通状态识别，准确率分别为93.38%、90.00%、92.03%、82.84%，本文方法的精确率、召回率、F1度量指标均优于其他方法。最后，提出基于关联度和车流量的权重评判系数，定量描述关联度的可靠性，选取主干路上11个交叉口的数据进行验证。结果表明，对于流量分布相对均匀的非孤立交叉口，本文获得的关联度与对应流量正相关，具有有效性和可解释性。

关键词: 智能交通, 交叉口状态识别, 时间图注意力网络, 交叉口关联度

Abstract: To analyze the intersection real-time traffic status and its correlation with adjacent intersections, this paper established the VISSIM simulation environment for the study based on actual traffic survey data. An intersection traffic characteristics matrix is proposed to describe traffic flow and physical characteristics. An intersection interaction time matrix is also derived by the traffic volume, traffic flow ratio, and effective green time of the traffic signals. This paper then proposes a temporal graph attention (TGAT) network based intersection traffic state recognition model. The matrix and initial labels of traffic data are inputted into the model, and the classification accuracy of the study intersection is obtained under "smooth", "stable", "congested", and "blocked" statuses. The weight matrix of neighbors can be calculated and used to represent the degree of association between intersections by computing the distances between neighbors. Additionally, the neighbor weight matrix can be obtained by computing the similarity between neighbors to describe the association between intersections. The TGAT, multilayer perceptron, long and short-term memory network, and support vector machine were selected for intersection traffic status recognition, and the accuracy rates were respectively 93.38%, 90.00%, 92.03%, and 82.84%. The precision, recall, and F1-score of TGAT were higher than the compared methods. At last, a weight judging factor based on correlation and traffic volume is proposed to quantitatively describe the reliability of the correlation, and data from 11 intersections on main roads are selected for validation. The results showed that for non-isolated intersections with relatively uniform traffic distribution, the correlation obtained in this paper was positively correlated with the corresponding traffic and was valid and interpretive.

Key words: intelligent transportation, intersection state recognition, temporal graph attention network, association degree of intersection

中图分类号: