车辆换道行为风险识别研究

发布时间：2018-05-30 00:47

本文选题：换道行为 + 支持向量机　；参考：《昆明理工大学》2017年硕士论文

【摘要】：换道行为是车辆在行驶过程中的随机动作,该行为有可能产生交通冲突点,甚至导致不同程度的交通事故发生。一般根据换道过程平顺或猛烈,大致将其分为安全性换道行为、风险性换道行为两类,换道行为过于猛烈将增加交通事故风险。伴随车辆换道预警系统的不断发展,车辆换道预警系统的出现,有望在车辆进行换道的起始阶段准确识别出两类不同的换道行为,并自动对存在风险的换道操作给出相应预警或干预。本文运用模式识别技术中的支持向量机方法,尝试建立一种能有效识别区分安全换道和风险换道两种驾驶行为的识别模型。借助KMRTDS驾驶模拟仿真平台开展安全换道和风险换道模拟实验,用采集到的驾驶操纵数据和车辆运行数据提取训练及测试识别模型的样本。本文研究意义在于,当识别模型判断出某一换道行为,因操作过于猛烈而超出一般安全换道行为阈值时,换道预警系统第一时间报警提示驾驶员,以避免事故发生。分类模型的识别效果主要受识别时间窗大小、特征参数、模型自身参数等方面的影响,本文运用ROC.运行结果,综合AUC值和模型的分类准确率对识别效果进行分析,通过“最优时间窗确定、最优特征参数提取、最优算法寻找模型参数”三个步骤,建立最优识别模型。首先,将定义的换道行为起始时刻作为时间窗中点,向前向后分别取相同时间段(0.5s、1s、1.5s)形成3个时间窗(1s、2s、3s),并经过对比分析三个时间窗下的模型识别效果,确定了最优时间窗为2s。其次,运用逐步回归分析、因子分析、多维偏好分析三种方法对原有的特征参数进行降维处理。其中,经逐步回归分析法提取的特征参数训练后的分类器性能最好,故将此方法提取的参数作为最优特征参数。最后,运用了枚举算法、粒子群算法、遗传算法等三种算法进行参数寻优。其中,遗传算法的分类准确率最低,枚举算法和粒子群算法分类准确率较为接近,但后者的AUC值为0.992,接近原始数据下的0.996,较好的弥补了因数据降维处理而损失的信息,所以选取粒子群算法为最优算法。在最优时间窗、最优特征参数、最优模型参数确定后,借助LIBSVM算法,在MATLAB中训练、验证识别模型,得到模型最终总体识别率为92.55%,基本能准确识别出车道保持、安全换道、风险换道三种行为,达到了运用较小样本量建立较高识别率模型的预期,取得较好的识别效果。
[Abstract]:The behavior of changing the road is the random action of the vehicle in the course of driving, which may produce traffic conflict points and even lead to traffic accidents of different degrees. Generally according to the smooth or violent course of changing the road, it can be roughly divided into two types of safe changing, and the risk of changing the road is two kinds, and the risk of traffic accident will be increased if the change of course is too violent. With the continuous development of the vehicle early warning system, the emergence of the vehicle early warning system, it is expected to accurately identify two different types of road change behavior in the initial phase of the vehicle change. And give the corresponding warning or intervention to the change operation of the risk automatically. In this paper, the support vector machine (SVM) method in pattern recognition technology is used to establish a recognition model which can effectively identify and distinguish two driving behaviors: safe change and risk change. With the help of the KMRTDS driving simulation platform, the simulation experiments of safe and risk change are carried out, and the samples of training and testing identification model are extracted from the collected driving control data and vehicle running data. The research significance of this paper is that when the identification model judges a certain changing behavior and exceeds the threshold of the general safe changing behavior because of the heavy operation, the early warning system of changing the channel will alert the driver in the first time to avoid the accident. The recognition effect of the classification model is mainly affected by the size of the time window, the characteristic parameters and the model's own parameters. The result of the operation is based on the analysis of the AUC value and the classification accuracy of the model. The optimal recognition model is established through the three steps of "determining the optimal time window, extracting the optimal feature parameters and finding the parameters of the model by the optimal algorithm". Firstly, the starting time of the change behavior is taken as the midpoint of the time window, and the same time interval is taken forward and backward, respectively.) three time windows are formed, which are 1 sm ~ 2 s ~ 3 s ~ (3), and the optimal time window is determined to be 2 s by comparing and analyzing the model recognition effect under the three time windows. Secondly, stepwise regression analysis, factor analysis and multidimensional preference analysis are used to reduce the dimension of the original characteristic parameters. Among them, the performance of the classifier trained by stepwise regression analysis is the best, so the parameters extracted by this method are regarded as the optimal feature parameters. Finally, three kinds of algorithms, enumeration algorithm, particle swarm optimization algorithm and genetic algorithm, are used to optimize the parameters. The classification accuracy of genetic algorithm is the lowest, the classification accuracy of enumeration algorithm and particle swarm optimization algorithm is close, but the AUC value of the latter is 0.992, which is close to 0.996 under the original data, which makes up for the loss of information due to the dimensionality reduction processing. So the particle swarm optimization algorithm is chosen as the optimal algorithm. After the optimal time window, the optimal feature parameter and the optimal model parameter are determined, the recognition model is trained in MATLAB with the help of LIBSVM algorithm, and the overall recognition rate of the model is 92.55. The three behaviors of risk changing reach the expectation of establishing a higher recognition rate model by using smaller sample size and obtain better recognition effect.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491

【参考文献】