【作者】 秦岭；

【导师】 秦岭； 邓又源；

【作者基本信息】 武汉理工大学 ， 车辆工程（专业学位）， 2024， 硕士

【摘要】 随着社会和经济的快速发展,公共交通体系的建设,尤其是城市交通路口安全问题,已成为各级政府和社会各界普遍关注的热点和焦点。本文基于英国交通部道路安全数据集,对典型代表地区的交通事故数据采用机器学习算法,挖掘城市交通路口处的风险与成因,为相关职能部门提供了治理参考。具体工作内容如下:第一,将事故多发点拓展至事故多发区域、事故多发路口与事故多发时空路口。考虑到传统聚类方法中参数选择困难的问题,采用了一种核密度估计与DBSCAN密度聚类结合的参数选择方法,基于最优搜索半径(窗宽)初步鉴别事故多发区域与事故多发路口,发现不同类型事故路口的成因;考虑到传统聚类算法局限于二维空间的问题,使用ST-OPTICS密度聚类算法将模型拓展至三维时空,鉴别出两类事故多发时空路口。研究结果表明,典型代表地区事故点呈现聚集态势,且本文提出的参数选择方法较传统6)-空间距离曲线法的检验指标更加优秀,ST-OPTICS密度聚类算法较传统算法运行时间更短。第二,针对传统机器学习算法预测精度低的问题,基于LightGBM回归集成学习算法对典型代表地区路口处风险特征进行学习,得到对交通路口事故发生影响最大的30个风险参数。按照时间、地点、人、车、路、环境的分类对风险特征进行统计与可视化,并按风险特征重要性降序输出,留待后续场景识别使用。研究结果表明,本文使用的LightGBM模型较传统算法的R2、MAE、MSE、RMSE、皮尔逊相关系数等检验指标表现更优秀。第三,将路口事故点的高维分布视作图数据结构,求解其邻接矩阵。针对传统AP聚类算法聚类效果较差的问题,基于图聚类中的谱聚类,使用对事故影响最大的30个风险特征对道路安全数据集交通事故进行高维聚类,根据聚类簇中心发现典型风险场景并分析后,从驾驶员、交通管理部门、城市规划部门的角度提出相应治理建议。研究结果表明,谱聚类较传统AP聚类的轮廓系数、CH分数、DBI值等检验指标表现更加有效。综上,本文多方位地找出了事故多发路口的分布,发现了交通路口的风险来源,并提供了相应的解决策略,为公共交通网络的治理做出了一定的贡献。更多还原

【Abstract】 With the rapid development of society and economy,the construction of the public transportation system,especially the safety issues at urban traffic intersections,has become a hot topic and focus of attention for governments at all levels and various sectors of society.Based on the Road Safety Data Set of the UK Department for Transport,this paper employs machine learning algorithms to analyze traffic accident data from typical representative areas,uncovering the risks and causes at urban traffic intersections,and providing governance references for relevant functional departments.The specific work is as follows:First,expanding accident-prone spots to accident-prone areas,accident-prone intersections,and accident-prone spatiotemporal intersections.Considering the difficulty in selecting parameters in traditional clustering methods,a parameter selection method combining kernel density estimation with DBSCAN density clustering is adopted.Based on the optimal search radius(window width),accident-prone areas and accident-prone intersections are preliminarily identified,and the causes of different types of accident intersections are discovered.Considering the limitation of traditional clustering algorithms to two-dimensional space,the ST-OPTICS density clustering algorithm is used to extend the model to three-dimensional spatiotemporal space,identifying two types of accident-prone spatiotemporal intersections.The research results show that accident spots in typical representative areas are clustered,and the parameter selection method proposed in this paper is superior to the traditional k-space distance curve method in terms of evaluation indicators.The ST-OPTICS density clustering algorithm has a shorter runtime compared to traditional algorithms.Second,addressing the problem of low prediction accuracy of traditional machine learning algorithms,the LightGBM regression ensemble learning algorithm is employed to learn the accident characteristics at intersections in typical representative areas,obtaining the 30 risk parameters that have the greatest impact on traffic intersection accidents.Risk features are statistically analyzed and visualized according to the classification of time,location,people,vehicles,roads,and environment,and output in descending order of importance for subsequent scene recognition use.The research results indicate that the LightGBM model used in this paper outperforms traditional algorithms in terms of evaluation indicators such as R2,MAE,MSE,RMSE,and Pearson correlation coefficient.Third,treating the high-dimensional distribution of intersection accident spots as graph data structure and solving its adjacency matrix.Addressing the poor clustering effect of traditional AP clustering algorithm,spectral clustering in graph clustering is utilized.Using the 30 risk features with the greatest impact on accidents to conduct high-dimensional clustering on the road safety data set of traffic accidents,typical risk scenarios are discovered based on the cluster centers,and corresponding governance suggestions are proposed from the perspectives of drivers,traffic management departments,and urban planning departments.The research results show that spectral clustering is more effective than traditional AP clustering in terms of evaluation indicators such as silhouette coefficient,CH score,and DBI value.In summary,this paper comprehensively identifies the distribution of accident-prone intersections,uncovers the sources of risks at traffic intersections,and provides corresponding solutions,making certain contributions to the governance of public transportation networks.更多还原