【作者】 夏创文；

【导师】 徐建闽；

【作者基本信息】 华南理工大学 ， 控制理论与控制工程， 2013， 博士

【摘要】 近年来我国各省加快高速公路建设进程，高速公路网已基本形成，随着车流量日益增长，路网的交通拥堵问题也越来越突出。因此，作为路网管理数据支撑基本条件的高速公路网运行状态监测工作的重要性不言而喻。本文围绕高速公路网运行监测信息获取与加工，力争实现路网运行监测信息的“可视”与“可测”相结合这一主题，对基于高速公路监控视频的交通参数获取、车型分类与雾霾检测，以及高速公路网多源交通信息融合处理等关键问题展开研究，针对现有技术方法的不足，提出了若干具有创新性和实用性的解决方案和算法，并通过丰富的在实际路网运行中的实验数据加以验证，同时通过本人以攻博期间亲自主持设计实施的广东省高速公路网运行监测与服务信息平台建设的工程实践，实现了部分技术研究成果的转化，取得了良好成效，也为未来系统升级完善发展奠定了坚定基础和指明了方向。完成的主要科研工作和研究成果如下：（1）针对固定单摄像机非标定的检测系统，综合研究了现有检测方法在交通场景中最常出现的影响检测鲁棒性的各种因素，分析原因并提出提高可靠性的约束条件。为消除主要因素的影响，提出一种适合在交通检测情形下应用的背景参考模型——基于区域的码本模型。并围绕此背景参考模型，提出一系列可裁剪、可扩展、全时段且适应复杂应用环境的交通流参数鲁棒性检测方法。通过对实际交通场景不同时段下的检测结果与传统检测方法进行对比，验证了检测方法的有效性和可靠性。（2）不同于现有的利用大气学物理模型、基于能见度检测方法的思想，本文提出了一种基于高速公路监控视频内容的自动雾霾级别检测方法，无需对摄像机标定以计算能见度，而是利用雾天图像的视觉特性进行检测。本文利用边缘检测提取出粗略的疑似雾区，通过梯度计算和距离变换区分天空与真实雾霾区并分割出真实雾霾区，提取出真实雾霾区内的颜色直方图作为特征向量，同时为了增强特征的表达能力，对一系列视频帧序列的特征进行融合，最后，通过SVM支持向量机分类器进行雾霾级别分类。实验表明能够针对各种有雾和无雾天气，以及雾霾的级别进行检测，具有较高的正确率。（3）提出了一种新颖的基于高速公路监控视频的车型分类方法。该方法通过提取车型图像的纹理特征，然后使用在线学习方式得到特征的编码字典，通过稀疏编码的稀疏性，挑选出最适合表达该车型的特征。稀疏编码能够更好地挑选适合的特征，与传统的量化相比能够防止过拟合的情况。同时为了排除车型图像提取中多车重叠等情况的干扰，使用低秩表示技术，通过对训练集中无干扰的图片进行矩阵分解，保留矩阵的低秩部分，去除稀疏（噪声）部分。得到表达矩阵与之前稀疏编码系数和字典进行连乘，从而得到车型图像中的主要成分，大大减轻了车型分类中引入的误差，得到优良的分类结果。最后使用BP神经网络结合改进的BOW分类结果、基于稀疏和低秩去噪的分类结果、深信网络分类结果，进行决策数据融合，得到最终结果。通过对实际数据的分析实验表明该方法进行得到了较高的分类准确度。（4）提出了一种基于模糊聚类方法（FCM）的高速公路交通拥挤判别方法。研究了高速公路交通信息数据融合处理技术，结合广东省佛开高速真实高速公路历史数据对本章提出的交通拥挤判别方法进行验证，实验结果表明该方法能够很好的应用于高速公路监控系统中。以排队论为基础，对高速公路交通拥挤扩散特性进行了研究，包括分析了传统累计到达-离去模型对估计最大排队长队估计存在的缺陷，并建立了改进的最大排队长度估算模型及其求解流程。仿真实验结果表明模型能够很好的与实际交通流相匹配，在高速公路管理中具有较高的应用价值。（5）提出了一种基于有限混合模型的高速公路行程时间估计模型。分析车牌识别数据或收费数据处理方法，分析不同车型的行程时间概率分布规律，提出高速公路路段行程时间概率分布是由若干车型的行程时间组成的混合概率分布模型，进而对行程时间进行估计。将高速公路路径行程时间划分为主干道和出入匝道的行程时间组成，提出一种对高速公路路径行程时间的估计方法。提出了一种基于卡尔曼滤波的高速公路网行程时间的融合估计与预测模型。在考虑高速公路网中有限的、低覆盖率的交通信息采集，整合、融合区域高速公路网监控系统中的基于现有交通信息资源的行程时间估计值，并充分利用海量的行程时间历史数据，构建基于卡尔曼滤波模型的行程时间估计模型，完成对高速公路网行程时间的融合估计与预测。实例分析表明该方法的可行性和有效性。最后，在上述研究成果的背景基础上，本文最后以广东省高速公路网运行监测与服务信息共享平台实际工程为例，介绍了其系统的设计与实现。从平台的总体功能定位和系统功能结构设计开始，简要介绍了八个核心子系统——联网设施与设备管理子系统、交通流数据采集分析子系统、联网视频监控子系统、高清卡口车牌识别子系统、应急与救援子系统、交通地理信息子系统、出行信息服务子系统、基于移动互联网的“广东高速通”移动应用子系统的实现。并在已实现的系统功能基础上，结合未来智能交通系统技术的发展趋势，提出了系统未来技术升级研究发展的展望。更多还原

【Abstract】 In recent years, freeways in our country have witnessed a rapid development and thenetwork of freeways in provinces has been formed gradually. At the same time, trafficcongestion in freeways has been increasingly serious with the development of society andeconomy. Therefore, it’s very important to obtain traffic condition data for monitoring trafficconditions in freeway network.In this paper, we focus on acquiring the traffic condition data and processingmulti-source data to achieve the goal of the combination of “visualization" and"measurability". Therefore, we study on the acquiring of traffic parameters based on freewaysurveillance video, models classification, detection of haze and multi-source trafficinformation fusion processing about freeways and other key issues. Concerning theshortcomings of the existing methods, it puts forward some innovative and practical solutionsand algorithms, which have been verified through abundant experimental data in actual roadnetwork operations. At the same time, through the engineering practice personally presidedover by the author himself about the construction of information platform about freewaysystem operation monitoring and service in Guangdong province, part of the implementationof technical research has been realized and good results has been achieved. A firm foundationhas been laid for the future development and pointing out the direction.The main work and achievements are summarized as follows:1. For the calibration of test system of single fixed cameras we comprehensively studiedthe existing test methods of testing the influence of the traffic most commonly occurring inthe scene of robustness of various factors, analyzed related reasons and improved constraintsabout reliability. To eliminate the influence of main factors, a kind of suitable application inthe situation of traffic detection background reference model based on the area code isestablished. And around the background reference model, a series of cutting, extensible, themethods of robustness test about traffic flow parameters adapting to the complex applicationenvironment as well as in full time were proposed. Through the comparison between actualtraffic scene detection results under different times and the traditional detection method, thevalidity and reliability of the test method have been verified. 2. On the basis of the analysis of the monitoring vehicle in the video image extractionmethod, three methods about the model classification have been proposed. The first one isbased on improved bag-of-word (BOW) models of the classification method, the second oneis based on sparse coding and low rank denoising algorithm, the third one is based on the edgeof the conviction in the network model. On this basis, the surveillance video is proposed inthis paper based on the freeway of multimodal fusion model classification method.Experimental results have verified the effectiveness and practicability of these models.3. This paper proposes a new fog haze level detecting method based on freewaysurveillance video. This method is different from the existing atmospheric physics modelbased on the idea of visibility detecting method. By using of the characteristics of imagecontrast, innovative statistical histogram is proposed using HSV color space as a featurevector, using the SVM classifier to detect fog haze level.The method does not need tocompute the visibility of camera calibration, therefore it is not limited to the scene referenceand be compatible with the original freeway network video monitoring system, making fulluse of the traffic camera video images. The cost is low. Through the experiment, the methodfunctions well in all kinds of foggy and fogless weather, as well as different fog haze levels.4. A novel freeway traffic congestion identification method based on the fuzzy clusteringmethod (FCM) has been put forward. Freeway traffic information data fusion processingtechnology is studied. With the real historical running data of Foshan-Kaiping Freeway inGuangdong Province,the traffic congestion identification method of this paper has beenvalidated. The experimental results show that this method can be applied to freewaymonitoring system very well. Based on queuing theory, the dispersion characteristics offreeway traffic are studied, including the analysis of the traditional cumulative reach-leavemodel to estimate the defects, and to estimate the maximum queue. And maximum queuelength estimation model and its solving process has been established and improved.Simulation experiment results show that the model can well match the actual traffic flow, andhas higher application value in the freeway management.5. A novel travel time estimation model based on limited mixture distribution wasproposed. It’s usually difficult to obtain the accurate travel time in freeway network as trafficdetector is limited. To integrate or fuse the travel time in freeway network which is estimated from single-point detector (such as microwave traffic detector), point-to-point detector(license plate recognition/toll data) and the floating car data, a novel travel time fusingestimation or prediction algorithm based on Kalman filtering is proposed to obtain moreaccurate and complete travel time in freeway network.On the basis of the above research results, in the final this paper takes the real project ofGuangdong provincial expressway network monitoring system platform as an example toanalyze the design, the overall orientation and functional structure of the system. The eightcore subsystems are introduced one by one: traffic flow detection subsystem, entrance ramplimiting subsystem, the main road control subsystem, exit ramp information subsystem, trafficillegal catch beat system, CCTV system and traffic information issuing system. This papercombines with the development trend of freeway monitoring technology and proposes thefuture technical development prospect of the system based on the realized system function.更多还原