【作者】 郭孜政；

【导师】 张殿业；

【作者基本信息】 西南交通大学 ， 交通运输规划与管理， 2009， 博士

【摘要】 随着机动化水平的不断提高,交通安全问题日益突显。驾驶员作为道路交通系统的核心因素,其驾驶行为状态很大程度决定了道路交通系统的安全水平,因此对驾驶行为状态进行研究成为道路交通安全的关键性课题。现有研究大多集中于事故与非事故组间驾驶员生理、心理状态指标的静态差异性对比,特定因素对驾驶行为的干扰性分析及驾驶员生理、心理、性格等因素与行为的相关性探讨。未能从时间序列的角度,对实时驾驶行为状态适宜性予以研究,分析不同时刻、不同状态水平与事故率间相关关系,确定实时系统安全水平,为系统安全性控制提供依据。针对上述问题,本文构建了一套系统的驾驶行为险态辨识理论与方法,并对相关模型予以研究,以期实现驾驶行为状态的过程辨识,动态确定驾驶行为危险水平,为预防和减少道路交通事故提供理论依据。本文研究内容可涵盖为以下要点:基于认知心理学,结合信息获取、信息处理、操纵输出这一信息处理链,探究了驾驶行为的形成机制,客观界定了驾驶行为能力指标集;通过道路交通系统任务需求能力与实际驾驶能力耦合匹配性分析,探讨危险的形成机理,并将其归结为突变、同步渐变、异步渐变三类基本模式;提出交通事故形成机理模型,该模型解释了危险、事故的产生原理及转换规律;通过数学推导证明随驾驶时间延长,驾驶行为状态水平逐步下降、系统危险性逐步增大,该结论与实际情况相一致。提出了三类驾驶行为险态量化分级方法。(1)基于风险分析的险态分级:首先将风险分析理论引入到驾驶行为危险性研究,提出了驾驶行为风险的概念,基于风险的本质化定义,确定了驾驶行为风险中事故概率与事故损失的合成关系,并就事故概率与事故损失的确定方法予以探讨,然后基于ALARP原则,将驾驶行为险态划分为可忽略、可容忍、不可容忍三级,并采用信息融合算法实现了各险级划分点安全边际成本收益(MP)值的确定;(2)基于状态指标变化性的险态分级:首先对驾驶行为状态各指标数据予以平均及指数平滑去糙化预处理,消除随机误差的干扰,显现各状态指标数据变化的本质性趋势,然后计算各指标数据的二次差分值,参照二次差分值确定各险级划分点:(3)基于行为状态相似性的险态分级,对驾驶行为状态指标测试数据予以时段平均化预处理,在设定驾驶行为险态分级数k的前提下,采用有序聚类算法,实现驾驶行为状态的优化分级。构建了驾驶行为险态辨识因子的量化析取方法。首先给出动视野、动视力、暗适应、听力、掩蔽听力、短时记忆力、判断能力、注意力、反应时、操纵能力等10项驾驶行为状态指标的测试方法及指标计算公式;其次在按性别、年龄、驾驶里程对驾驶员予以分组的前提下,进行12小时连续模拟驾驶测试,并每隔15分钟采集一组驾驶行为状态因子指标值:然后在对行为状态指标数据予以预先分级的前提下,采用单因子分析法对实验数据予以分析。分析结果表明反应时、注意力、判断能力三项指标在各分级间差异显著(p≤0.05),故可作为驾驶行为险态辨识主因子。构建了高负荷驾驶任务下驾驶员注意力状态概率模型。将注意力划分为集中、分散两态,应用更新过程构建连续短时注意力状态转换概率模型,采用数值分析方法给出模型的求解;随后针对状态转换循环点处难以满足齐次性的特征,采用时段分割对短时状态概率模型加以推广,得到长时连续驾驶员注意力状态转移概率模型。采用模型模拟结果与试验结果对比的方法验证了模型的合理性。构建了三类驾驶行为险态辨识模型。首先以贝叶斯决策理论为核心,以分类错判损失最小为目标函数,分别以经济损失量与状态差异性为基础设计了两类错判损失矩阵,构建了贝叶斯险态辨识模型;其次以模糊数学中的隶属函数理论为基础,以相似性分类误差平方和最小为目标函数,构建了FCM驾驶行为险态辨识模型,依据已有学习样本,采用循环迭代算法实现模型的训练;然后基于神经网络理论,设计了包含三个输入神经元,两个输出神经元的BP神经网络驾驶行为险态辨识模型。结合实际数据,以错判率为评定指标,对各类模型的辨识精度予以测定。最后对模型的适用范围进行了讨论,其中贝叶斯辨识模型适用于基于风险分析与状态相似性分级的驾驶行为险态辨识问题,FCM模型仅适用于基于状态相似性分级的驾驶行为险态辨识问题,BP神经网络智能算法辨识模型具有较强的通用性,可适用于本文所提出的三类驾驶行为险态分级辨识问题。开发了“驾驶行为险态辨识系统”软件。该系统实现了基础数据录入、处理,行为状态的量化分级,驾驶行为险态辨识三大功能。更多还原

【Abstract】 Transportation safety problems have become more and more obvious since urban motorized level has improved. Being a core factor of road traffic system, drivers’ driving status plays a rather important role in safety level of road traffic system. Thus it becomes a key issue of road traffic safety to study on driving behavior status. Most of researches at the present are focused on comparing the static differences between accident and non-accident groups on drivers’ physiological indexes and mental ones. And they have discussed the disturbance of some given factor, as well as some correlations among drivers’ physiology, psychology, characteristics and behaviors. The shortage is those cannot analyze the relationship among accidents rate, different status level and different time on the basis of time sequence, which can provide some evidences for system safety control and make real-time system safety levels. To solve the above problems, the dissertation constructs a series of theories and methods to identify driving behavior risk status and study on some related modules. The purposes are to realize process identify of driving behaviors and supply driving risk levels dynamically, which can control driving process and decrease or prevent traffic accidents on roads. The summary is as follow:The formation mechanism of driving behavior is discussed combining with a chain of information acquisition, information treatment and handle output, on the basis of cognitive psychology. It gives out an objective definition of driving ability indexes set and discusses how a risk can generated according to analyze the coupling matching ability between road traffic system tasks demand ability and actual driving ability. The formation mechanism is resolute as three basic modes which are mutation, synchronization and asynchronous gradual change. It also proposes a module of traffic accidents formation mechanism, which explain the production principles of risk and accidents as well as how they convert. Mathematical derivation has proved that driving status level will gradually decreased and system risk will increased along with the driving time passes. The result is consistent to actual situation. The paper puts forward to a quantity method to classify three types driving risk status. (1) Risk status classify based on economy: As risk analysis theory is introduced into the research on driving hazards, it gives out a concept of driving risk on the basis of risk essence and determines the composite relation between accidents probability and accidents lost in driving risk, and the method to find them out. On ALARP principle, driving risk statuses are divided into neglected class, tolerable class and intolerable class, after which safety margin cost profit for breakpoints for different classes with regards to information fusion algorithm. (2) Risk status classify based on status indexes variability: Exponents smoothing and average smoothing must be pretreated towards indexes data of driving status to eliminate random errors disturbance and show their essential trend out. Then calculate quadratic differential value of every index data, referring to which the breakpoints of every class can be determined. (3) Risk status classify based on similarity of behavior state: first to make time period averaged for driving status index test data, then set driving risk status class number k, classify optimization of driving status can be realizes if sequential cluster method is adopted.A way of quantization disjunction has been constructed here to find out driving risk discrimination factor. Test method and calculating formulas of ten indexes about driving status first are described here, which are dynamic visual field, dynamic vision, dark adaptation, hearing, masking hearing, short-time memory, thinking judgment ability, focus, reaction time and handling ability. Then we did an 12h’ continuous simulated driving test on grouped drivers by gender, age and driving mileage, collecting driving status factors values of one group every 15min.We have to classify driving status indexes data before analyze the collected data by single factor method. The result shows an obvious difference (p≤0.05) among reaction time, focus and judgment ability on different classes. Thus they can be used as main factors of driving risk status identification.A probability model of driver attention state under high-load driving task is porposed. It divided shifting state space of attention into concentration and disconvergence, the Markov process is introduced to this paper. The continuous short-term attention state for conversion probability model and a approximate solution were presented. Moreover, according to the long time driving the state conversion point hardly meeting the homogeneous, separated time sessions so the model was generalized. In order to verify the rationality of this model, combination of actual data simulated to test and compared the results with experiment.Here are three modules constructed to identify driving risk status. The core of first module is Bayesian Decision Theory. It aims at minimized classification risk, designing two matrixes to describe lost of wrong judgment based on economy losses and status differences, to obtain a Bayesian risk identification module. Depending on Membership Function Theory in fuzzy mathematics, minimized sum of squares of similar classify errors being objective function, the second module called FCM driving risk status identification is set up. According to owned study sample, the module can be trained by cyclic iteration algorithm. BP neural network driving risk status identification module consists of three input neurons and two output neurons, which is based on Neural Network Theory. Identification precision for each module is tested as mistaken judgment being the index and on the basis of actual data. Finally, some discussions are developed on the application scope of modules. Bayesian risk identification module is suitable for problems based on risk analysis and status similarity classify, while FCM driving risk status identification can only work when talking about problems on status similarity classify and BP neural network driving risk status identification module, which are more intelligent as well as universal, are applicable for the three types problems talked above.Also software of driving risk status identification is developed, which can help to input data, data processing and behaviors’ quantification and classification, identify driving risk status.更多还原