计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (13): 266-272.DOI: 10.3778/j.issn.1002-8331.2203-0361

• 工程与应用 • 上一篇    下一篇

时空图卷积网络下的路网交通事故风险预测

王庆荣,周禹潼,朱昌锋,吴玉玉   

  1. 1.兰州交通大学 电子与信息工程学院,兰州 730070
    2.兰州交通大学 交通运输学院,兰州 730070
  • 出版日期:2023-07-01 发布日期:2023-07-01

Road Network Traffic Accident Risk Prediction Based on Spatio-Temporal Graph Convolution Network

WANG Qingrong, ZHOU Yutong, ZHU Changfeng, WU Yuyu   

  1. 1.School of Electronic&Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
    2.School of Traffic&Transportation, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Online:2023-07-01 Published:2023-07-01

摘要: 路网交通事故预测是实现道路管控、路线规划的最重要方式之一。考虑到路网中各路段特征与环境因素的影响,建立基于图卷积神经网络(GCN)和门控循环单元(GRU)的时空门控图卷积(STGRGCN)模型预测交通事故风险。通过GCN提取出道路间的空间关联性,通过GRU提取出环境因素中的时间关联性,再通过GCN与GRU的复合模块提取出时空关联性。选取美国全国交通事故数据集中洛杉矶市和休斯顿市相关数据对模型进行检验,STGRGCN模型的均方根误差、平均绝对误差以及召回率在两个城市分别为4.09、2.14、0.714和5.79、3.24、0.683,优于已有统计模型、机器学习模型以及复合模型。设计该除各模块的消融实验,证明该模型各模块皆有助于提升预测性能。

关键词: 交通事故风险预测, 图卷积网络, 门控循环单元, 注意力机制, 深度学习

Abstract: Road network traffic accident prediction is one of the most important ways to realize road management and control and route planning. Considering the characteristics of each road section in the road network and the influence of environmental factors, a spatiotemporal gated graph convolution(STGRGCN) model based on graph convolution network(GCN) and gated recurrent unit(GRU) is established to predict the risk of traffic accidents. The spatial correlation between roads is extracted through GCN, the temporal correlation in environmental factors is extracted through GRU, and then the temporal and spatial correlation is extracted through the composite module of GCN and GRU. The relevant data of Los Angeles and Houston in the national traffic accident dataset of the United States are selected to test the model. The root mean square error, average absolute error and recall rate of STGRGCN model are 4.09, 2.14, 0.714 and 5.79, 3.24, 0.683 respectively in the two cities, which are better than the existing statistical model, machine learning model and composite model. The ablation experiment of removing each module is designed to prove that each module of this model is helpful to improve the prediction performance.

Key words: traffic accident risk prediction, graph convolution network(GCN), gated recurrent unit(GRU), attention mechanism, deep learning