【作者】 陈超；

【导师】 程明； 范晓亮；

【作者基本信息】 厦门大学 ， 计算机技术， 2019， 硕士

【摘要】 交通事故被广泛认为是加剧城市交通拥堵的重要因素之一。近年来,大量研究工作关注于交通事故的致因因素分析研究,并以此来预测交通事故风险和时空影响。交通事故的主要致因因素包括:驾驶员特征、交通流量、天气因素和道路网络等。然而上述致因因素在城市交通环境中存在复杂性和不确定性,进而对交通事故风险预测带来了两方面挑战:第一,现有工作大多关注预测某一封闭道路(如高速路、快速路等)的事故,而面向城市复杂路网环境的交通事故风险预测的研究鲜有报道;第二,现有工作大多采用传统的机器学习方法,而机器学习方法缺乏有效挖掘多维数据高维特征的能力。总之,交通事故风险预测研究迫切需要综合考虑城市道路环境的复杂性和动态性,以及事故致因因素之间的高维依赖性等因素,相关研究工作具有重大理论和应用价值。针对上述问题,本文面向海量城市交通事故数据,并融合了实时交通流量和天气等多源数据,提出了一种城市级别的交通事故风险预测算法,称为“栈式降噪卷积自编码算法(SDCAE)”。SDCAE算法旨在及时有效地预测城市区域交通事故发生的风险,主要研究内容包括:第一,结合城市路网、交通事故和车牌识别设备的分布情况对城市进行相应网格区域划分;第二,数据预处理并整合城市网格区域内的交通事故数和交通流量;第三,构建栈式降噪自编码网络模型来学习多源输入数据的高维特征并进行风险预测。最后,本文利用某市2016年1月1日至8月31日约7万条交通事故数据,以及9.4亿条车牌自动识别记录,同时结合降雨量和能见度等天气数据开展大规模实验,验证了本文算法的有效性,比6种对比算法(线性回归、逻辑回归、决策树、随机森林、支持向量机和栈式降噪自编码)的预测误差降低了6%～20%。本文提出的城市交通事故风险预测方法,一方面有利于交通指挥管理部门在事故风险高的区域提前部署警力,以提高交通事故快速处理效率,缓解因事故而产生的交通拥堵;另一方面,有利于提高道路交通事故的实时感知能力,为驾驶人的出行提供风险预判和路径规划,并引导其在事故风险高的区域绕行。更多还原

【Abstract】 Traffic accidents are widely regarded as one of the important factors that aggravate urban traffic congestion.In recent years,a lot of studies have focused on analyzing causation factors of accident occurrences,and used it to predict the risk and spatio-temporal influences of traffic accidents.The major causation factors of accident occurrences include driver characteristics,traffic flows,weather and road networks.However,the aforementioned factors are complex and uncertain in the urban traffic environment,which brings two challenges to predict accident risks.First,most of the existing works focus on predicting a closed road(i.e.,highways,expressways)rather than accident risk prediction on complicated urban road environment.Second,many works adopt traditional machine learning algorithms,which is believed to be lack of the ability to effectively uncover high-dimensional features among multi-dimensional data.In general,it is urgent to investigate the accident risk prediction issues,combining the complexity dynamics of the urban road environment with the high-dimensional dependencies among causation factors of accident occurrences.The relevant research works have promising values in both theoretical and application area.To address challenges above,this paper is oriented to massive urban traffic accident data,fused with multi-source data such as real-time traffic flows and weather.We propose a city-level traffic accident risk prediction algorithm named SDCAE(Stack Denoise Convolution AutoEncoder).The SDCAE algorithm aims to predict the risk of city-level traffic accidents in a timely and effective manner.The main research contents of this paper include:1)we divide the urban networks into grids,fused with the distribution of road network,accident occurrences and license plate recognition equipment;2)we pre-process the aforementioned datasets and integrate traffic accidents with traffic flows in each grid;and 3)we construct a stack denoise convolution autoencoder network model to learn high-dimensional features of multi-source input data and then conduct risk prediction.Finally,based on about 70,000 accident occurrences,940 million traffic flows records,and weather data such as rainfall and visibility during January 1st,2016 to August 31st,2016,we validate the effectiveness of the proposed algorithm.Intensive experiments and evaluations illustrate that SDCAE could reduce the prediction error by 6%-20%,compared with 6 baseline algorithms(linear regression,logistic regression,decision tree,random forest,support vector machine and stack denoise autoencoder).The significant of this work is twofold.On one hand,it benefits the traffic control and management department to deploy traffic polices in high-risk accident areas in prior,which could both increase the efficiency of immediate processing on accidents,and ease the traffic congestions caused by accidents.On the other hand,it enhances the real-time awareness of road accidents,which could provide the prognosis of accident risks and route planning for drivers,and then guide them to make a detour on high-risk regions.更多还原