目前，我国高速公路总里程逾13.6万公里，其三大要素在于安全、绿色、畅通。对江苏而言，影响交通安全的主要因素为能见度、“黑冰”等不良天气。充分利用沿线已有的监控视频，实时监测解析，间接预测交通事件，是预防事故、保障安全、提升管理的重要手段。鉴于高速公路具有区域广、范围大、管理要素复杂的特点，云计算、大数据、分层动态地图技术具有“用武之地”。为此，论文结合多项省部级科研项目，基于既有监控视频，专题研究高速公路运行环境、预测交通事件、构建智能交通集成应用示范。 论文描述了国内外高速公路监测的现状，分析了高速公路事故的成因，研究了道路能见度视频监测、雨雪冰实时监测、交通事件间接检测等方法，解析了实验结果；在阐述算法机理的基础上，给出了ELM-AID交通事件检测算法和H-ELM交通事故预测算法及其技术实现；丰富了LDM³(Local Dynamic Map/ Multimedia/Management)区域分布·动态·地图/多媒体/管理这一 ITS智能交通集成体系架构。其中： 在研究已有能见度检测算法的基础上，为缓解局部最优解和噪声敏感问题，提出了迹范数道路能见度检测算法，经对比实验，所提算法具有较好的鲁棒性和检测精度； 考虑到单一技术手段或方法在道路结冰检测方面存在检测精度、稳定性不够好的问题，采用综合光、热流、温度和电容检测等多种技术与手段的路表面结冰检测方法和装置； 综合采用室内模拟场景和现场实际场景的摩擦系数实验，研究了轮胎与沥青路面在不同相态的水作用下的接触摩擦系数，并给出了两者之间的阶梯式下降关系； 鉴于ELM算法具有极限运算能力和良好的泛化性能，研究了ELM-AID交通事件检测算法，并通过ELM-AID与BP检测模型的比较研究，验证了ELM-AID在检测准确度和运算速度方面的优势； 针对单隐层ELM可能会出现过拟合的问题，研究了多层架构的H-ELM算法，并构建了相应的预测模型，通过两种交通事件预测比较，表明H-ELM在预测效率方面具有较大的优越性； 鉴于大数据和云计算技术发展趋于成熟，研究了信息汇聚的的多时空综合路网数据处理方法。采用Hadoop和Spark等主流大数据平台，构建相应的智能交通大数据存储模型和分布式存储系统、以及分布并行计算系统平台，形成了较为完整的智能交通信息系统整体构架和系统解决方案。 论文的创新点或特点在于： ·提出了迹范数道路能见度检测算法。其缓解了局部最优解和噪声敏感问题，简化了参考物标定，实验表明，该算法相对于视频亮度特征点算法和路面固有亮度估算算法的检测准确率平均提高了近10%，计算用时却降低了逾20%，检测效率更高； ·提出了一种道路结冰综合检测方法。其综合采用光、热流、温度和电容检测手段，解决了单一技术手段检测精度不够高、稳定性不够好等问题，提高了结冰检测的精度和准确度； ·提出了ELM-AID事件间接检测模型及多层架构的H-ELM事故预测模型。前者相对于传统的BP神经网络具有良好的分类性能和极限的计算能力，运算速度提高了百倍，而后者兼具无监督特征学习和有监督的特征类，更加适用于海量数据和云计算环境； ·优化了LDM³智能交通集成系统架构。其采用Hadoop和Spark等主流大数据平台，以交通大数据分布式存储与并行计算，丰富了信息汇聚层内涵，形成了较为完整的智能交通信息系统整体构架和系统解决方案。 关键词：智能交通，视频监测，ELM-AID，H-ELM，云计算，LDM³

 At present, the total mileage of expressways in China has exceeded 136,000 kilometers, of which "safe, green and smooth" are the three main elements. For Jiangsu Province, the main factors affecting traffic safety are visibility, "black ice" and other unfavorable weather conditions. Relying on best use of the existing monitoring videos, real-time monitoring analysis and indirect prediction of traffic events have been important measures to predict traffic accidents, assure safety as well as to improve management. In view of expressway's characteristic in broad region, wide range, and sophisticated management factors, cloud computing, big data, and hierarchical dynamic map technology have finally found a use for themselves. Therefore, combined with a number of provincial-level scientific research projects, based on the existing monitoring videos, dedicated research has been carried out for the environment conditions on expressway, prediction of traffic accidents, and construction of demo for integrated application of intelligent traffic. In this dissertation, the status of expressway monitoring at home and abroad is described and the causes of highway accidents are analyzed. The monitoring methods of visibility in videos, the real time monitoring ways of rain &snow ice, and indirect detection means of traffic events are studied. Then the experimental results are analyzed. Based on interpretation of the algorithm mechanism, the ELM-AID traffic incident detection algorithm and H-ELM traffic accident prediction algorithm as well as its technical implementation are shown, which enriches the content of LDM3 (Local Dynamic Map/ Multimedia/ Management), an intelligent traffic integrated system architecture of ITS. Out of which: On the basis of research on the existing visibility measurement algorithm, a visibility detection algorithm based on trace norm is proposed to alleviate local optimal solution and noise sensitive problem. According to comparison experiments, the proposed algorithm has got better robustness and detection precision. Considering the problems of poor accuracy and lack of stability for the single technical method or means in detection of road icing, a method and device combined by various technology and means of optical, heat, temperature and capacitance is used for road surface icing detection. Based on the friction coefficient experiment of the indoor simulated scene and the actual scene, the friction coefficient of tire and asphalt pavement under the action of different phase water is studied, with illustration of the step-style descending relationship between them. In view of the extreme learning ability and good generalization performance of the ELM algorithm, ELM-AID traffic incident detection algorithm is studied, and advantages of ELM-AID in detection accuracy and operation speed are verified by comparison of the ELM-AID and the BP detection models. As to the over fitting problem single hidden layer ELM might run into, the H-ELM algorithm of multilayer architecture is studied and the corresponding prediction model is constructed. By means of comparing the two traffic event predictions, it is shown that the H-ELM has obtained comparatively greater superiority in the prediction efficiency. In view of the maturing technological development of big data and cloud computing, study work has been carried out for the multi space-time integrated road network with information convergences well as the data processing method. Based on the main data platforms of Hadoop, Spark and so on, corresponding big data storage model and distributed storage system as well as a distributed parallel computing system platform have been built-up, with fairly complete framework and solution of intelligent transportation information system coming into being. The innovative points or characteristics of the dissertation lie in: A road visibility detection algorithm based on trace-norm is proposed, which has alleviated the local optimal solution and noise sensitive problem, and simplified the calibration of reference objects. Experiments show that, the proposed method has better measurement performance. The measurement accuracy of this method increased by nearly 10% but calculation time decreased by more than 20%, compared to the method based on intensity feature points in the video and the method using the inherent intensity of the road surface. A comprehensive method for road icing detection is proposed to improve the accuracy, which has combined light, heat flow, temperature as well as capacitance detection methods to solve problems of poor detection accuracy, lack of stability when using single detection technology. Indirect event detection model for ELM-AID event and H-ELM model of multi-tier architecture is proposed. The former has obtained favorable classification performance and extreme learning ability, and more than 100 times speed acceleration than the traditional BP detection model, and the latter has non-supervision feature learning and supervised feature classification simultaneously, which is more suitable for mass data and cloud computing. The intelligent traffic integrated system framework of LDM3 has been optimized. With adoption of the mainstream big data platform such as Hadoop, Spark and so on, through means of distributed storage & parallel computing of traffic big data, the connotation of information convergence layer has got further enriched, and the integral structure and systematic solution of intelligent traffic information system has been developed. Keywords: Intelligent Traffic, Video Surveillance, ELM-AID, H-ELM, Cloud Computing, LDM³