基于模式转移和操控特性的驾驶风格评测研究

发布时间：2018-06-17 15:01

本文选题：驾驶风格 + 驾驶模式　；参考：《清华大学》2016年博士论文

【摘要】：我国道路交通安全形势严峻,不良驾驶风格充斥于日常驾驶行为之中。在车联网技术迅速发展的大背景下,研究并实现对险态驾驶风格的有效监测,并以离线反馈教育或在线危险预警的方式进行干预,对提升道路行车安全性具有重要意义。当前针对驾驶风格的研究普遍存在综合评测维度局限性和驾驶操作评测维度不完整性的问题。为解决这些问题,本课题构建了考虑人认知特性和信息优化表达的驾驶模式分解与辨识方法体系,从驾驶模式转移和驾驶操作控制两个维度出发,设计了相应的驾驶风格险态评测方法,拓展了驾驶风格评测的方法体系。针对高速工况下驾驶模式辨识存在的动态时变、多维耦合和局部相似的问题,提出了以制动减速度、跟驰时距和跟驰时距变化率作为纵向驾驶模式的辨识指标;针对横向驾驶模式的辨识,采用了2s时窗内的方向盘转角香农熵、2s时窗内的横向加速度均方根、5s时窗内的横摆角速度标准差和4s时窗内的速度香农熵四个特征参数作为随机森林分类器的输入,以分类器的输出概率作为横向驾驶模式的辨识指标。验证结果显示,本课题提出的驾驶模式辨识方法体系可实现对各驾驶模式的有效辨识,辨识精度可达86~98%。为实现对驾驶风格在模式转移维度(频度)的有效评测,以驾驶模式转移概率为基础,优选出了五种可表征驾驶风格模式转移特性的典型驾驶模式转移形态分别为:近距离跟驰→受限右换道,受限右换道→受限左换道,受限左换道→迫近,迫近→受限右换道和受限左换道→自由直行。以这五种形态发生的条件概率值作为随机森林分类器的输入,以输出的隶属于不同风格类型的概率值作为判别标准,实现了对驾驶风格险态频度的有效评测。交叉验证结果显示,该方法的辨识精度可达93%,比基于传统方法对驾驶风格险态频度表现的评测精度高出18%。为实现对驾驶风格在操作控制维度(强度)的有效评测,提出了以加速度的幂指数在时间序列上的积分量化表征人感知到的险态强度,通过对人在加速、制动、车距控制、换道控制和转弯控制五个维度上的险态强度感知进行加权综合,得到驾驶操作激进指数作为驾驶风格险态强度表现的评测指标。基于事后视频和现场评价的验证试验结果显示,该算法的辨识精度可达85~92%。综合驾驶风格在险态频度和强度两个维度的表现,构建了决策树模型对综合驾驶风格进行评测。验证试验表明,该决策树模型的辨识精度可达89%。
[Abstract]:Our country road traffic safety situation is grim, the bad driving style is flooded in the daily driving behavior. Under the background of the rapid development of vehicle networking technology, it is of great significance to study and realize the effective monitoring of dangerous driving style and to intervene in offline feedback education or online hazard warning. At present, there are some problems in the study of driving style, such as the limitation of comprehensive evaluation dimension and the incompleteness of driving operation evaluation dimension. In order to solve these problems, this paper constructs a driving pattern decomposition and identification method system, which takes into account the cognitive characteristics of human beings and the optimal expression of information, starting from the two dimensions of driving mode transfer and driving operation control. The method of dangerous driving style evaluation is designed, and the method system of driving style evaluation is expanded. Aiming at the problems of dynamic time-varying, multi-dimensional coupling and local similarity in driving mode identification under high-speed operating conditions, this paper puts forward the identification index of longitudinal driving mode based on braking deceleration, following time distance and changing rate of following driving time distance. For lateral driving mode identification, In this paper, four characteristic parameters of steering wheel angle Shannon entropy in 2s window and transverse acceleration root-square velocity standard deviation in 5s window and velocity Shannon entropy in 4s window are used as input of random forest classifier. The output probability of the classifier is used as the identification index of the lateral driving mode. The verification results show that the driving mode identification system presented in this paper can effectively identify each driving mode, and the identification accuracy can reach 860.98%. In order to realize the effective evaluation of driving style in the dimension of mode transfer (frequency), it is based on the probability of driving mode transfer. Five typical driving mode transfer patterns which can characterize the characteristics of driving style mode transfer are selected as follows: short distance following, restricted right change, restricted left approach. The approach is restricted to the right and the restricted left to go straight and free. The conditional probabilistic values of these five morphogenesis are taken as the input of the random forest classifier and the probabilistic values of the output which belong to different style types are taken as the criterion to realize the effective evaluation of the dangerous frequency of driving style. The results of cross validation show that the accuracy of this method can reach 933, which is 18.5% higher than that of the traditional method in evaluating the dangerous frequency of driving style. In order to effectively evaluate driving style in the dimension of operation control (intensity), this paper presents a method of quantifying the perceived dangerous state intensity by integrating the power exponent of acceleration in time series, which is controlled by acceleration, braking and distance control. On the basis of weighted synthesis of risk intensity perception in five dimensions of change control and turn control, the radical index of driving operation is obtained as the evaluation index of dangerous intensity performance of driving style. The experimental results based on post video and field evaluation show that the identification accuracy of the algorithm can reach 85 / 92. The decision tree model is constructed to evaluate the comprehensive driving style in the two dimensions of dangerous frequency and intensity. The experimental results show that the precision of the decision tree model is up to 89.
【学位授予单位】：清华大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：B842

【相似文献】