【作者】 朱焱；

【导师】 余立建；

【作者基本信息】 西南交通大学 ， 控制理论与控制工程， 2012， 硕士

【摘要】 高速公路的建设与发展,给人们的生活出行带来了方便和快捷,同时它所引发的安全问题也引起了人们的重视。高速公路上发生交通拥挤、交通事故等各类交通事件,不仅降低了高速公路的运输效率和运营效益,还给人们的生命和财产安全带来严重的损失。如何快速检测并及时处理交通事件,有效降低交通事件的危害成了人们高度关注的问题。交通事件检测技术的核心就是检测算法,检测算法的好坏直接影响事件的检测效率。因此,对检测算法研究有着十分重要的意义。大多数交通事件检测算法都是通过分析交通流数据找到交通事件和交通流参数的对应关系。因此,本文首先分析交通事件和交通流参数的特性；接着学习研究BP(Back Propagation)神经网络和其学习算法,仿真分析各种优化BP学习算法后选择LM(Levenberg Marquard)算法作为本文的BP神经网络学习算法,并给出N·W(N guyen-Widrow)离散度选取法优化LMBP[4260]神经网络初始权值。在此基础上,设计基于优化后的LMBP神经网络的交通事件检测算法,算法建模过程中就LMBP神经网络结构和参数设置进行详细的设计,然后利用构建好的算法模型检测交通事件。最后,根据弱分类器差异越大集成效果越好的原则,利用交通流参数的不同组合形式构成不同网络结构的LMBP弱分类器,用AdaBoost方法集成得到LMBP模型集成系统,并建模检测交通事件。本文采用I-880交通流数据,运用MATLAB7.10工具进行仿真,通过仿真结果分析得知,利用N-W离散度选取法优化网络初始权值提高了LMBP神经网络的收敛速度,缩短了LMBP模型检测交通事件的时间,结合AdaBoost方法又进一步提高了检测精度,获得了良好的综合检测性能。更多还原

【Abstract】 The construction and development of the freeway has brought much convenience and efficiency to people, however, at the same time we need pay attention to the security issues caused due to it. Traffic congestion, traffic accidents and other types of traffic incidents occur on the freeway, not only reduce the transport efficiency and operational effectiveness of the freeway, and also cause plenty of serious losses of people’s lives and properties. How to quickly detect and deal with traffic incidents, effectively reduce the traffic incident hazards has become a highly topical issue. The detection algorithm is the core of the traffic incident detection technology which directly impact on the performance of traffic incident detection. Therefore, the research of the detection algorithm is of great significance.Most of the detection algorithms find the corresponding relationship of traffic incidents and traffic flow parameters by analyzing traffic flow data. Therefore, in this paper we firstly analyze the traffic incidents and traffic flow parameters’characteristics, then study the BP neural network and its learning algorithm. Finally we choose the LM algorithm in my paper after analysis and comparison of various BP algorithms and use the N-W dispersion select method to optimize LMBP neural network’s initial weights. Based on those, we design traffic incident detection algorithm due to the improved LMBP neural network, and make detailed designation of LMBP network structure and parameters in the model building process, then use the built algorithm model to detect traffic incident. At last, according to the principle of "the more the difference of weak classifiers, the better the integration result", we use different combinations of traffic flow parameters to constitute LMBP weak classifiers in different network structures, and put in use of the AdaBoost method to integrate improved LMBP model, and build model to detect traffic incident.This paper use1-880data in simulation, the result show that the use of this N-W dispersion select method optimization LMBP’s initial weights greatly raise the convergence speed of the detection algorithm, further improve the detection precision in combination with the AdaBoost method and obtain a good detection performance in the end.更多还原