[1]阮鸿柱,黄小弟,王金宝,等.面向高速公路事故风险预测的深度学习方法[J].计算机技术与发展,2023,33(11):189-195.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 028]
　RUAN Hong-zhu,HUANG Xiao-di,WANG Jin-bao,et al.A Deep Learning Approach for Highway Accident Risk Prediction[J].,2023,33(11):189-195.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 028]

点击复制

面向高速公路事故风险预测的深度学习方法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年11期

页码:

189-195

栏目:

人工智能

出版日期:

2023-11-10

文章信息/Info

Title:

A Deep Learning Approach for Highway Accident Risk Prediction

文章编号:

1673-629X(2023)11--0189-07

作者:

阮鸿柱¹ ; 黄小弟¹ ; 王金宝¹ ; 杜梦辉²

1. 云南省综合交通发展中心,云南 昆明 650031;

2. 北京交通大学 计算机与信息技术学院,北京 100044

Author(s):

RUAN Hong-zhu1 ; HUANG Xiao-di1 ; WANG Jin-bao1 ; DU Meng-hui2

1. Yunnan Transportation Development Center,Kunming 650031,China;

2. School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China

关键词:

智能交通; 交通事故风险预测; 对比学习; 自适应图神经网络; 数据增广

Keywords:

intelligent transportation; traffic accident risk prediction; contrastive learning; adaptive graph convolutional network; data augmentation

分类号:

TP391. 4

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 11. 028

摘要:

高速公路的交通事故风险预测对智能交通和公共安全具有重要意义。 现有方法通过挖掘历史事故的时空特征预测交通事故风险。 但是,在高速公路事故风险预测中仍存在以下两个挑战。 首先,事故具有不均衡的空间分布,相邻路段的事故分布差异可能较大,而相隔较远却具有相似拓扑连接关系路段的事故分布可能较相似。 另外,由于事故的偶发性,其在时间维的分布非常稀疏,因此在捕获事故影响因素时缺乏足够的样本。 针对第一个挑战,使用自适应图卷积网络以数据驱动的方式学习路段间的空间相关性;此外,根据 Mixup 策略进行数据

增广以生成足够多的事故风险样本解决事故数量稀疏的问题,然后用对比学习方法以更好地区分风险与非风险样本,以实现更准确的事故风险预测。 基于桂林市高速公路网真实交通数据集的实验结果表明,相比于最优方法,该方法的平均绝对误差指标降低了 18. 3% ,平均准确率、召回率指标分别提升了 8. 1% 、6. 9% ,因此,该方法可以更准确地预测高速公路事故风险。

Abstract:

Highway traffic accident risk prediction is vital to intelligent transportation and public safety. Existing approaches predict trafficaccident risk by mining the spatio - temporal characteristics of historical accidents. However,there are still two challenges in highwayaccident risk prediction. Firstly,the accidents have uneven spatial distribution. The difference in accident distribution between adjacentroads may be large,while the accident distribution of distant roads with similar topological connection relationships may?

be similar. In addition,due to the contingency of the accident, its distribution in the time dimension is quite sparse,so there is not enough sample tocapture the influence factors?

of traffic accidents. For the first challenge,we use an adaptive graph convolutional network to learn thespatial correlation between roads in a data-driven way. In addition,we?

adopt data augmentation based on a Mixup strategy to generateenough accident risk samples to solve the problem of data sparsity and then use contrastive learning to better distinguish risk and non-risksamples so as to achieve more accurate accident risk prediction. The experimental results based on the real traffic dataset of the Guilin expressway network show that compared with other optimal methods,the mean absolute error of the proposed method is reduced by 18. 3% ,and the average accuracy and the recall are increased by 8. 1% and 6. 9% ,respectively. Therefore, the proposed method can predicthighway accident risk more accurately.

常用功能