基于驾驶行为分析的醉酒驾驶检测方法研究

发布时间：2018-10-24 10:53

【摘要】：随着社会经济的不断发展,机动车保有量的持续增加,道路交通安全问题日益严峻。交通事故逐渐成为造成人类伤亡的主要原因之一,而醉酒驾驶是其中一个很大的原因。因此,对醉酒驾驶进行准确识别意义重大。本文主要利用驾驶模拟舱进行驾驶实验,采集了醉酒驾驶和正常驾驶的驾驶行为参数,对不同状态驾驶行为进行了分析,提取了能显著表征醉酒驾驶的特征参数,并基于提取的特征参数对驾驶员状态进行了识别。主要研究内容如下:首先,基于驾驶模拟平台,设计实验采集驾驶行为数据。本文对国内外研究现状进行了广泛调研,在总结前人研究的基础上,利用驾驶模拟平台开展驾驶实验,采集了25名驾驶员的驾驶行为数据,并对数据进行了整理与筛选,建立了样本数据库。其次,分析了醉酒对驾驶行为的影响,进行了特征选择。本文运用统计分析方法对醉酒状态和正常状态的驾驶行为进行了对比分析,明确了醉酒对车速、加速度、方向盘转角、制动踏板踩踏深度、加速踏板踩踏深度、偏离中心线距离等驾驶行为的影响规律,最终选取方向盘转角作为识别特征。再次,采用滑动数据窗进行特征提取。为减轻采用总体数据长度统计均值和标准差对数据特征的损失,本文基于滑动数据窗分别求取方向盘转角均值序列和标准差序列构建特征参数,并研究了不同数据窗长度对特征提取的影响和不同数据窗提取特征的有效性,结果表明滑动数据窗可以提取明显表征醉酒驾驶的特征参数。第四,采用近似熵和样本熵进行特征提取。首先对近似熵和样本熵的参数选择进行了分析,搭建了最优熵值计算模型;其次基于提取的最优熵值构建特征参数,并采用ROC曲线对特征参数的判别性能进行了对比分析,结果表明方向盘转角近似熵作为判别特征的性能优于样本熵。最后,搭建了基于驾驶行为特征的驾驶员状态检测模型。分别采用KNN和SVM构建基于单特征参数、多特征参数加权融合的醉酒驾驶检测模型,对比分析了各醉酒驾驶检测模型的识别准确度与运行效率。结果表明,SVM检测模型识别效果优于KNN检测模型,加权融合可以提高检测模型的识别准确度;SVM在寻找最优参数时需要耗费比较多的时间,运行效率低于KNN。
[Abstract]:With the development of social economy and the increasing number of motor vehicles, the problem of road traffic safety is becoming more and more serious. Traffic accidents have gradually become one of the main causes of human casualties, and drunken driving is one of the major causes. Therefore, it is of great significance to identify drunken driving accurately. In this paper, the driving behavior parameters of drunken driving and normal driving were collected, and the driving behavior of different states were analyzed, and the characteristic parameters which could represent drunken driving were extracted. The driver state is recognized based on the extracted feature parameters. The main research contents are as follows: firstly, based on the driving simulation platform, the experiment is designed to collect driving behavior data. This paper has carried on the extensive investigation to the domestic and foreign research present situation, has carried on the driving experiment using the driving simulation platform, has collected 25 drivers' driving behavior data, and has carried on the arrangement and the screening to the data. The sample database is established. Secondly, the influence of drunkenness on driving behavior is analyzed, and the characteristic selection is carried out. In this paper, the statistical analysis method is used to compare and analyze the driving behavior of drunken state and normal state, and it is clear that drunkenness affects speed, acceleration, steering wheel angle, tread depth of brake pedal, stampede depth of accelerator pedal, and so on. Finally, the steering angle is selected as the recognition feature. Thirdly, a sliding data window is used for feature extraction. In order to reduce the loss caused by the statistical mean and standard deviation of total data length, based on the sliding data window, the characteristic parameters of the steering wheel angle mean series and the standard deviation sequence are obtained respectively in this paper. The influence of different length of data window on feature extraction and the effectiveness of different data window in feature extraction are studied. The results show that sliding data window can extract the characteristic parameters that represent drunken driving obviously. Fourthly, approximate entropy and sample entropy are used for feature extraction. First, the parameter selection of approximate entropy and sample entropy is analyzed, and the calculation model of optimal entropy value is built. Secondly, the characteristic parameters are constructed based on the extracted optimal entropy value, and the discriminant performance of feature parameters is compared and analyzed by using ROC curve. The results show that the approximate entropy of steering wheel angle is superior to the sample entropy as the distinguishing feature. Finally, a driver state detection model based on driving behavior characteristics is built. KNN and SVM are used to construct the drunk driving detection model based on single feature parameter and multi-feature parameter weighted fusion. The recognition accuracy and running efficiency of each drunk driving detection model are compared and analyzed. The results show that the recognition effect of SVM detection model is better than that of KNN detection model, and the recognition accuracy of the detection model can be improved by weighted fusion, and SVM needs more time to find the best parameters, and its running efficiency is lower than that of KNN..
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U463.6

【相似文献】