【作者】 张良力；

【导师】 严新平；

【作者基本信息】 武汉理工大学 ， 智能交通工程， 2011， 博士

【摘要】 利用先进的车辆传感与检测技术获取车辆驾驶行为信息及车外驾驶环境信息,通过信息融合及模式识别方法对道路交通危险态势进行评估与决策,是实现机动车驾驶安全预警系统功能的主要技术手段。然而,现有的驾驶安全预警系统对当前道路危险态势评估结论的准确度较低,产生此现象的原因是驾驶安全预警系统仅以车辆状态和车外环境信息作为当前道路危险态势评估的先决条件,忽略驾驶员的意图及其变化趋势在主动安全控制中起到的关键性作用,故而容易对当前道路危险态势做出错误估计。开展面向安全预警的机动车驾驶意图识别研究,对提高已有机动车驾驶安全预警系统对当前道路危险态势评估的准确度具有现实意义。本论文所探讨驾驶意图识别研究将从驾驶员行为特征角度进行分析,研究内容包括驾驶行为与驾驶意图机理,使用自行设计的驾驶行为试验平台开展实车驾驶试验；根据驾驶行为特征构建不同条件下的驾驶意图识别模型,以及使用迭代算法对建立的模型参数进行学习与优化等。具体研究工作内容如下：(1)分析驾驶员产生驾驶意图的机理。从机理层面分析驾驶员的驾驶行为与意图之间的关联,得出驾驶意图在整个驾驶过程中产生的环节及其作用,确立本论文研究的驾驶意图类型识别及其适用范围。当车辆处于道路直线路段时,驾驶员所持有并为系统所识别的驾驶意图类型为“换道意图”和“跟车意图”中的1种；在交叉口(无信号灯控制)路段内,驾驶意图类型为“直行意图”和“转弯意图”,转弯意图可细分为“左转意图”和“右转意图”。根据驾驶行为与意图之间的不确定性,本论文选定动态贝叶斯网络(DBN)和隐马尔科夫模型(HMM)理论与方法作为研究驾驶意图识别的主要途径。(2)构建驾驶意图识别建模数据实车信息采集系统。较详细地介绍了系统实现过程,包括系统设备选型与制备、传感器信息采集设计与实现、采集信息数据上传协议以及数据管理软件等。信息采集系统需求分析是以驾驶意图识别研究对象为依据,明确实车驾驶试验采集数据对象包括驾驶员手脚部驾驶动作、车辆速度和车辆地理位置等。根据实车传动机构特点以及车辆已有传感器,选用适配的传感与检测技术实现。采集的信息数据采用自定义传输协议报文格式,并与GPS上传信息同时被上位机读取。数据管理软件采用ADO组件和VisualC++6.0实现,数据存储文件为Microsoft Office Access系统文件。(3)道路不同地段内驾驶意图识别建模研究。针对车辆处于道路不同位置时驾驶员所具有驾驶意图类型不同的机理特征,分别对直线路段内和交叉口路段内的驾驶意图识别开展研究。研究工作包括：,实施驾驶意图建模数据采集实车试验；从驾驶意图类型特征角度对试验数据开展驾驶行为特征研究,得出与驾驶意图具有映射关系的驾驶行为参数及驾驶动作序列；使用图模型及DBN相关理论构建了驾驶行为与意图HMM网络架构,并提出以时段T和交叉口与车辆之间距离s更新HMM中隐状态转移概率和观测状态概率,以此使HMM网络结构具有动态性。构建基于驾驶行为与意图HMM和Viterbi算法的驾驶意图识别模型；探讨基于驾驶意图与可观测驾驶动作序列的驾驶行为预测原理,建立了基于驾驶意图识别的驾驶行为预测模型。实例说明了建立的HMM网络结构能够描述驾驶员行为与意图之间的动态关系；使用Viterbi算法及判定法则可实现驾驶意图识别及驾驶行为预测。(4)驾驶意图识别HMM模型参数学习研究。首先针对驾驶意图识别HMM建模与使用过程中可能存在的误差进行分析并探讨模型参数学习原理。其次,介绍了基于Baum-Welch算法的驾驶意图识别HMM模型参数学习过程,学习过程包括前向概率推算、后向概率推算、最大期望概率比值计算,并根据计算结果优化驾驶行为与意图HMM模型参数中各矩阵概率值。最后,给出某驾驶意图识别HMM模型参数学习实例,结果可使已知驾驶动作序列出现概率增大,表明Baum-Welch算法对驾驶意图识别HMM模型参数学习有效。更多还原

【Abstract】 The main technical features of vehicle driving safety precaution system (VDSPS) are as followed:driver’s behavior and traffic environment information are collected by advanced sensing and detection technology, and risk assessment of road traffic safety situation and decision-making operation are based on data fusion and pattern recognition method. However, the conclusion accuracy of risk assessment was low in current VDSPS. What result in that phenomenon were those the VDSPS made decision only relied on vehicle statement and road traffic environment information, but ignoring the key factor of driver’s intention and its variation tendency in active safety control. Accordingly, the function validity of VDSPS would be interfered driver’s normal judgments. It is obvious that making research on motorist’s intention recognition for traffic safety precaution is useful to improve the risk assessment conclusion accuracy of current VDSPS. In this paper, driver’s intention is analyzed through characters of driver’s driving behavior. The works include mechanism analysis between driving behaviors and intentions, using self-designed information collecting platform to design and implement real vehicle driving experiments. According to the characters of driving behavior, intention recognition models are building in different conditions. Finally, an iterative algorithm is used for model parameters on learning and optimization. The detail researching works are as follow:(1) The association between driver’s behavior and intentions is analyzed on level of mechanism. How and where driver’s intentions generated, and what effects it made in process of driving are discussed. The types of driver’s intention recognized in this paper and their scope of application will be fixed after mechanism analysis. When driver’s vehicle on straight going road section (be far from the intersection in front), the types of driver’s intention for recognition would be defined as "lane-change" or "car-follow". When the vehicle on road section which was near to the intersection (without traffic control signals) in front, the types of driver’s intention for recognition would be defined as "straight-going" or "turning". While the "turning" could be sub-defined as "left-turning" and "right-turning". According to non-determinacy between driver’s behavior and intentions, the method to recognize driver’s intention in this paper would be based on theory and algorithms of dynamic Bayesian networks and hidden Markov model.(2) A real vehicle information collecting system is established for driver’s intention modeling data collection. In this paper, the design and implement processes of the information collecting system are introduced in detail. Processes include vehicle-embedded equipments selection, detection design for different sensors in actuating mechanism, protocol for collected information up transmission, data management software. Information demand analysis is based on objects of driver’s intention recognition in different condition. The data types collected by information collecting system include driver’s operation with hands and feet, vehicle velocity and geographical position. According to characters of real vehicle actuating mechanism and sensors, suitable sensing and detection technologies are selected for remote information collecting. All collected information utilizes self-defined protocol to make up message format and be transmitted to host computer with GPS information. The data management software is implemented by ADO component and Visual C++, and data are stored in Microsoft(?) office access file.(3) According to the mechanism characters of driver’s intention held in different road sections, the intention recognition models are built toward straight-going road section and intersection road section. The detail researching works are as followed. Experimental data of driver’s behaviors for driver’s intention recognition modeling is collected by information collecting system. Study features of driver’s behaviors data in the aspect of intention type. The probability statistics between behaviors and intention mapped each other are calculated, and corresponding operation sequences of driver are constituted. Using graphic model and dynamic Bayesian networks (DBN) theory to establish the hidden Markov model (HMM) network structure which describing relationship between driver’s intention and behavior. The network structure is dynamic. Because the probabilities of hidden statement transfer and observable statement mapped with hidden statement could be varying with real-time variable. The intention recognition method is based on HMM of driver’s intention and behavior and Viterbi algorithm. And driver’s behavior prediction is based on those as well.(4) HMM model parameters learning for driver’s intention recognition is discussed in this part. Firstly, errors existed in HMM modeling or applications are analyzed and the principles for model parameters learning are explored. Secondly, process of HMM model parameters learning based on Baum-Welch algorithm is introduced. The process includes forward probability calculating for operation sequences, backward probability calculating, and expectation maximization for operation sequences. The results calculated are the foundation for optimization of probabilities matrixes in HMM of driver’s intention and behavior. Finally, an example of HMM model parameters learning is given. The result of model parameters learning makes the emerge probability of a known operation sequence increased. That means the Baum-Welch algorithm is effective to HMM model parameters learning for driver’s intention recognition.更多还原