实测数据驱动的自动驾驶道路测试驾驶模式辨别方法

自动驾驶道路测试中车企驾驶模式数据具有一定保密性，导致自动驾驶能力难以被客观评估。为此，提出了实测数据驱动的自动驾驶道路测试驾驶模式辨别方法。首先选取数据特征值构建K近邻估计、支持向量机、决策树、随机森林和BP神经网络5种机器学习监督分类模型；其次通过非参数秩和显著性检验确定驾驶模式持续时长阈值，持续时长大于阈值的数据段记录为准确的驾驶模式数据，小于等于阈值的数据段则为驾驶模式待分类数据集；随机选取70%记录准确的驾驶模式数据作为监督分类模型训练数据集，剩余30%作为测试数据集；最后利用正确率、精确率和召回率3个指标评价5种监督分类模型，并选取表现最佳的分类模型用于待分类数据的驾驶模式辨别。基于上海市城市道路和快速路2个道路测试场景共约43.6万条数据，验证驾驶模式辨别方法的有效性。结果表明：随机森林监督分类模型辨别道路测试驾驶模式的效果最佳；城市道路场景和快速路场景待分类数据驾驶模式记录有误率分别达到42.3%和39.4%。实测数据驱动的驾驶模式的辨别与修复，可显著提升评估自动驾驶道路测试驾驶能力的准确度。     

基于自然驾驶数据的分心驾驶行为识别方法

大量证据表明，驾驶人分心是导致交通事故的主要原因之一。当前基于侵入式（如脑电波等）或半侵入式（如视频等）检测驾驶人分心的方法，不仅对驾驶任务造成一定干扰，且受多种环境因素的制约，误报率较高。基于此，只考虑非侵入式车辆运动特征，提出一种基于深度学习的驾驶人分心状态识别方法：首先，从自然驾驶数据集中获得大量的跟驰片段，采用态势感知方法，提取典型的分心驾驶片段，并建立仅包含车辆运动学特征的分心判别指标集；其次，利用梯度提升决策树-递归特征消除算法（GBDT-RFE）和随机森林-递归特征消除算法（RF-RFE）对特征进行重要度排序，得到重要度较高的分心监测指标；最后，采用长短时记忆神经网络（LSTM-NN）实现分心驾驶的分类识别，并与支持向量机和AdaBoost的模型结果进行对比。研究结果表明：LSTM-NN在判别分心或正常状态时F₁分别为89%、91%，高于SVM和AdaBoost对应二分类结果；进行多分类任务时，判别分心情景的平均F₁较SVM和AdaBoost分别提升了12%和7%，不同类别分心识别的误报率在15%以下，说明LSTM-NN能够有效学习分心序列的前后信息，有利于准确估计驾驶人的状态。研究结果可为车辆分心预警系统和驾驶风险倾向性评估提供方法基础。     

考虑驾驶风格的电动公交车能耗灰色关联投影-随机森林预测模型

为探索驾驶员驾驶行为与电动公交车能耗之间的关系，采用随机森林算法建立电动公交车能耗预测模型。为克服驾驶行为特征参数和样本数据的随机性对电动公交车能耗预测模型的负面影响，运用灰色关联投影法计算各驾驶行为特征参数的灰色关联度以及各样本数据的投影值，筛选出与能耗具有高关联性的驾驶行为特征参数作为模型的输入变量，以及相似度较高的样本数据作为训练集和测试集。同时，引入了与能耗具有显著相关性的驾驶风格变量以进一步提升模型的预测能力，运用K-means聚类方法将驾驶风格分类并得到驾驶风格标签。将驾驶风格标签和筛选后驾驶行为特征参数作为输入变量，单位里程能耗作为输出变量，基于筛选后的数据集建立了考虑驾驶风格的电动公交车能耗灰色关联投影-随机森林（GRP-RF）预测模型。基于广州市某线路电动公交车运营数据对模型进行检验，并运用该模型分析加速、制动和运行3种典型场景下相应驾驶行为特征参数对电动公交车能耗的影响。结果表明:该模型预测能耗的均方根误差（RMSE）和平均绝对百分比误差（MAPE）分别为0.001 8 kW·h/km和3.42%。相比于不考虑驾驶风格的GRP-RF模型和随机森林模型，该模型的RMSE分别降低了35.71%和48.57%，MAPE分别降低了38.82%和46.81%。研究结果表明:加速、制动和运行阶段的平均能耗分别为1.066，0.903 7，0.955 2 kW·h/km；为使各阶段能耗在相应均值以下，加速阶段应控制加速踏板开度在55%以内；制动阶段应控制制动踏板开度在25%以内；运行阶段应控制车速在40 km/h以内。      

基于多分类监督学习的驾驶风格特征指标筛选

交通事故与驾驶风格具有强烈的相关性,而驾驶风格的直观体现是驾驶行为.为深入分析驾驶行为与驾驶风格的关联性,探索不同驾驶风格群体之间的差异,筛选驾驶风格分类与识别影响因素,建立驾驶风格识别模型并验证有效性.依托车联网实验数据,利用K-means++算法对驾驶员样本数据集进行驾驶风格聚类,设计支持向量机-递归特征消除(SV...     

Driving segment simulation for determination of the most effective driving features for HEV intelligent control

This paper presents a methodological approach for determination of the most effective driving features for hybrid electric vehicle intelligent control, using the driving segment simulation. In this approach, driving data gathering is first performed in real traffic conditions using Advanced Vehicle Locator systems. The vehicle's speed time series are then divided into small segments. Subsequently, 19 driving features are defined for each driving segment, and the influence of the driving features on the vehicle's fuel consumption (FC) and exhaust emissions is investigated, using driving the driving segment simulation. The simulation approach is also verified by experimental test. Finally, the driving features are ranked by a new approach based on the definition of an effectiveness index and a correlation analysis. The results demonstrate that the velocity-dependent driving features such as ‘energy’, ‘mean of velocity’, ‘displacement’ and ‘maximum velocity’ are more effective on vehicle's FC and exhaust emissions. However, because of high dependency between these features, this study suggests independent driving features among the most effective driving features.     

基于随机森林算法的内河船舶油耗预测模型

准确的船舶油耗预测模型是船舶实现各项航行优化措施的基础.以长江干线某旅游船为研究对象,通过安装信息采集系统获得了大量的船舶实时营运数据.通过理论分析得出影响船舶油耗的主要因素为风速、风向、水深、水流速度和船舶航速;改进了随机森林建模时参数的设置方法,提出一种变量的重要性测度方法;对去噪处理后数据进行系统抽样并进行归一化处理,得到建模的样本数据;把样本数据按0.7∶0.3的比例随机分为训练样本和测试样本,对训练样本采用随机森林(RF)算法建立油耗预测模型;通过模型预测测试样本的油耗值,与实测数据对比,结果显示预测误差低于6.8%,优于BP神经网络与支持向量机(SVM)的预测结果;分析模型中各变量的重要性顺序为:航速>水流速度>水深>风速>风向,利用偏相关分析得到了单个因素与油耗间的定量关系.      

A similarity-based neuro-fuzzy modeling for driving behavior recognition applying fusion of smartphone sensors

Drivers’ behavior evaluation is one of the most important problems in intelligent transportation systems and driver assistant systems. It has a great influence on driving safety and fuel consumption. One of the challenges in this regard is the modeling perspective to treat with uncertainty in judgments about driving behaviors. Really, assessing a single maneuver with a rigid threshold leads to a weak judgment for driving evaluation. To fill this gap, a novel neuro-fuzzy system is proposed to classify the driving behaviors based on their similarities to fuzzy patterns when all of the various maneuvers are stated with some fuzzy numbers. These patterns are also fuzzy numbers and they are extracted from statistical analysis on the smartphone sensors data. Our driving evaluation system consists of three processes. Firstly, it detects the type of all of the maneuvers through the driving period, by using a multi-layer perceptron neural network. Secondly, it extracts a new feature based on the acceleration and assigns three fuzzy numbers to driver’s lane change, turn and U-turn maneuvers. Thirdly, it determines the similarity between these three fuzzy numbers and the fuzzy patterns to evaluate the safe and the aggressive driving scores. To validate this model, Driver’s Angry Score (DAS) questionnaires are used. Results show that the fusion of Inertial Measurement Unit (IMU) sensors of smartphones is enough for the proposed driving evaluation system. Accuracy of this system is 87% without using GPS and GIS data and this system is independent of smartphones and vehicles types.     

考虑电动公交在途特性的电池状态梯次划分

电动公交电池容量衰减造成里程焦虑增加、服务可靠性降低、电池资源浪费等问题。因此,评估和发现电动公交实际运营过程中影响电池健康状态的关键因素并划分电池状态尤为重要。基于电动公交长时间实际行驶过程中的充放电数据,结合安时积分法与最小二乘拟合建立电池容量估计模型,并据此计算各充放电片段的电池健康状态。进一步考虑电动公交在途特性,从电池组充放电属性、车辆行驶工况、公交营运状态3个角度提取可能影响电池健康状态的相关因素,并采用因子分析法将影响因素组合为12个影响因子,使用随机森林回归构建电池健康状态预测模型,从而根据预测结果的准确性反推获得各影响因子的重要度。最后考虑不同影响因素的重要度,利用加权聚类算法梯次划分电动公交电池健康状态为4个类别,下降梯度分别为-0.013 6、-0.011 9、-0.003 4、-0.002 8,并通过对比研究发现了同一条线路不同梯次的车辆电池组在放电深度、速度标准差、最大加速度和刹车次数等影响因素上的差异。研究结果表明:车辆荷载、电池电流释放情况、车辆行驶中速度的变化、电池的使用时间、线路拥挤状况以及电池充电深度大小对于电池健康状态的影响程度较大,而在公交营运状态相同条件下,驾驶人的行为对电池健康状态衰减程度有着较大影响。     

基于多源数据的公交专用道效能评价方法与影响模型

施画公交专用道是落实公交优先策略、改善公交服务质量,提升交通系统效率的重要策略.近年来,如何提高公交专用道设置的科学性、提升服务效能受到广泛关注.为科学定量评价公交专用道设置的正、负效能,建立基于多维运行影响的公交专用道效能评价指标体系,提出基于主观层次分析法和客观离差最大化法相结合的组合评价方法,实现对公交专用道效能...     

Development of E-rickshaw driving cycle (ERDC) based on micro-trip segments using random selection and K-means clustering techniques

In India, auto rickshaws are the most convenient and cheapest mode of near-to-door transport in both rural and urban areas. Such vehicles powered with internal combustion engines (ICEs) are one of the main sources of pollutants on urban corridors. One way to minimize tail-pipe emissions is to use electric motors in place of ICE. To evaluate the vehicle performance, energy consumption, driving behavior, optimal design and management of such electric vehicles, driving cycle is an important tool. So far, only limited studies exist on the development of a driving cycle for e-rickshaw. Moreover, these studies are concentrated in urban traffic environment and research accounting rural and urban environment together remain unexplored. In this study, real world driving data for 100 trips of e-rickshaw are collected on a road stretch passing through rural and urban setting. A high-end GPS data logger was used to collect vehicle kinematics such as continuous speed profile, acceleration/deceleration, heading, and vehicle position coordinates. Nine different driving characteristics representing actual traffic conditions are identified and used for developing e-rickshaw driving cycle (ERDC). Two approaches, random selection and k-means clustering are explored to arrive at best representative ERDC using micro-trips technique. The analysis results revealed that k-means clustering outperforms the random selection method with additional benefit of accounting traffic conditions systematically. The insights from this study can be used to understand and model the performance of e-rickshaw, in terms of energy consumption and driving characteristics, compared to other fossil-fuel driven automobiles.     

Detection of Cognitive and Visual Distraction Using Radial Basis Probabilistic Neural Networks

This paper suggests a real-time method for detecting a driver’s cognitive and visual distraction using lateral driving performance measures. The algorithm adopts radial basis probabilistic neural networks (RBPNNs) to construct classification models. In this study, combinations of two driving performance data measures, including the standard deviation of lane position (SDLP) and steering wheel reversal rate (SRR), were considered as measures of distraction. Data for training and testing the RBPNN models were collected under simulated conditions in which fifteen participants drove on a highway. While driving, they were asked to complete auditory recall tasks or arrow search tasks to create cognitively or visually distracted driving periods. As a result, the best performing model could detect distraction with an average accuracy of 78.0 %, which is a relatively high accuracy in the human factors domain. The results demonstrated that the RBPNN model using SDLP and SRR could be an effective distraction detector with easy-to-obtain and inexpensive inputs.     

模拟驾驶环境下驾驶人分心状态判别

为了探寻驾驶人分心判别方法,构建了驾驶人分心状态判别模型。首先设计分心模拟驾驶试验,采集正常驾驶和发送语音信息过程中的驾驶绩效特征和驾驶人眼动特征数据,建立驾驶人分心状态判别指标备选集;其次,采用基因选择算法对备选指标进行筛选,得到29个备选指标的重要度排序;然后,依次选取重要度较高的部分指标作为BP神经网络的输入指标,利用遗传算法（GA）全局搜索的性能优化BP神经网络的初始权值和阈值,将优化后的GA-BP神经网络作为弱分类器,再将多个弱分类器组合成Adaboost强分类器,建立基于Adaboost-GA-BP组合算法的驾驶人分心状态判别模型;最后,利用模拟驾驶器试验平台采集的数据计算不同判别指标数量下模型的性能,从而确定最优判别指标,并对模型进行验证和评价。结果表明：模型最优判别指标为重要度排序中前14个指标;模型能够准确识别驾驶人分心状态,判别精度为95.09%;与BP神经网络算法、GA-BP神经网络算法和Adaboost-BP神经网络算法相比,Adaboost-GA-BP组合算法在准确率、精准率、召回率、F₁值和ROC曲线等模型性能方面均最优。建立的模型能够有效判别驾驶人分心状态,可为驾驶人分心预警系统和分心控制策略提供依据。     

基于Simulink车辆舒适性和行驶安全性仿真分析

根据国标GB／T7031—2005机械振动道路路面谱测量数据报告,在MatLab中编写了随机路面激励谱仿真程序;利用拉格朗日方程建立了1／2车辆动力学模型,并用Simulink对其进行了仿真;以不同等级路面和不同车速下的随机路面激励谱作为输入,分析了车辆在不同等级路面、不同车速下的车身加速度均方根值和后轮的动载荷均方根值。这对满足汽车行驶舒适性和行驶安全性的情况下优化悬架参数具有重要意义。     

基于 AHP 和物元分析的愤怒驾驶状态辨识研究

驾驶人在愤怒情绪下的驾驶行为是影响车辆行驶安全性的重要因素之一,愤怒驾驶情绪的产生及其程度受到驾驶人自身和道路交通环境中多因素的影响。文中综合考虑驾驶人自身因素和行车环境对驾驶状态的影响,提出了愤怒驾驶状态的辨识方法。文中筛选了与愤怒驾驶行为相关的驾驶人因素和道路环境因素,构建了1个驾驶人愤怒状态辨识的层次分析模型,并根据相关因素之间对愤怒驾驶行为影响的重要程度构造判断矩阵,求出各相关因素对愤怒驾驶行为的影响权值。应用综合权重的物元多属性决策方法辨识驾驶人的愤怒驾驶状态及程度。应用所提出的方法对22组实车试验中出现的愤怒驾驶状态进行辨识,结果表明,72.7%的结果与实车实验所得的结果相符,因此,该方法可对愤怒驾驶行为进行识别。文中所提出的方法能够融合驾驶人因素和环境因素对愤怒驾驶行为的影响,有效的辨识出驾驶人的愤怒驾驶状况及程度。      

Cluster-based correlation of severe driving events with time and location

In this article, a systematic strategy is proposed to identify severe driving events occurrence correlation with time and location. The proposed approach, which is constructed based on batch clustering and real-time clustering techniques, incorporates historical and real-time data to predict the time and location of severe driving events. Batch clustering is implemented with the combination of subtractive clustering and fuzzy c-means clustering to generate clusters representing the initial correlation patterns. Real-time clustering is then developed to create and update real-time correlation patterns on the foundation of the batch clustering using the evolving Gustafson–Kessel like (eGKL) algorithm. In both clustering processes, the correlation of the events within time domain is identified first, and then two different levels of accurate correlations are conducted for the location domain. Real-time data of operating vehicles each equipped with a data acquisition and wireless communication platform are used to validate the proposed strategy. Batch clustering results reveal the severe braking events distribution and concentration at daytime and nighttime. Real-time clustering provides and updates the variation of the correlations/intercorrelation of different regions. Drivers can be notified of the potential severe driving locations through maps showing the driving routes. Through the variation of the correlations, drivers can recognize the events occurrence at different times and locations. The generated time series can be potentially used to develop spatial-time models for regions to model and forecast the events occurrence.     

Normal and risky driving patterns identification in clear and rainy weather on freeway segments using vehicle kinematics trajectories and time series cluster analysis

Enhancing traffic safety on freeways is the main goal for all transportation agencies. However, to achieve this goal, many analysis protocols of network screening models need to be improved through considering human factors while analyzing traffic data. This paper introduces one on the new analysis protocol of identifying and discriminating between normal and risky driving in clear and rainy weather. The introduced analysis protocol will consider the effect of human factors on updating the networking screening process of identifying hotspots of crash risk. This paper employs the Second Strategic Highway Research Program (SHRP2) Naturalistic Driving Study (NDS) data to investigate the behavior of normal and risky driving under both rainy and clear weather conditions. Near-crash events on freeways, which were used as Surrogate Measure of Safety (SMoS) for crash risk, were identified based on the changes in vehicle kinematics, including speed, longitudinal and lateral acceleration and deceleration rates, and yaw rates. Through a trajectory-level data analysis, there were significant differences in driving patterns between rainy and clear weather conditions; factors that affected crash risk mainly included driver reaction and response time, their evasive maneuvers such as changes in acceleration rates and yaw rates, and lane-changing maneuvers. A cluster analysis method was employed to classify driving patterns into two clusters: normal and risky driving condition patterns, respectively. Statistical results showed that risky driving patterns started on average one second earlier in rainy weather conditions than in clear weather conditions. Furthermore, risky driving patterns extended in average three seconds in rainy weather conditions, while it was two seconds in clear weather conditions. The identification of these patterns is considered as a primary step towards an automated development that would distinguish between different driving patterns in a Connected Vehicle CV environment using Basic Safety Messages (BSM) and to enhance the network screening analysis for increased crash risk hotspots.     

Drowsy behavior detection based on driving information

Drowsy behavior is more likely to occur in sleep-deprived drivers. Individuals’ drowsy behavior detection technology should be developed to prevent drowsiness related crashes. Driving information such as acceleration, steering angle and velocity, and physiological signals of drivers such as electroencephalogram (EEG), and eye tracking are adopted in present drowsy behavior detection technologies. However, it is difficult to measure physiological signal, and eye tracking requires complex experiment equipment. As a result, driving information is adopted for drowsy driving detection. In order to achieve this purpose, driving experiment is performed for obtaining driving information through driving simulator. Moreover, this paper investigates effects of using different input parameter combinations, which is consisted of lateral acceleration, longitudinal acceleration, and steering angles with different time window sizes (i.e. 4 s, 10 s, 20 s, 30 s, 60 s), on drowsy driving detection using random forest algorithm. 20 s-size datasets using parameter combination of accelerations in lateral and longitudinal directions, compared to the other combination cases of driving information such as steering angles combined with lateral and longitudinal acceleration, steering angles only, longitudinal acceleration only, and lateral acceleration only, is considered the most effective information for drivers’ drowsy behavior detection. Moreover, comparing to ANN algorithm, RF algorithm performs better on processing complex input data for drowsy behavior detection. The results, which reveal high accuracy 84.8 % on drowsy driving behavior detection, can be applied on condition of operating real vehicles.     

Emission-robust operation of diesel HEV considering transient emissions

In this paper, we consider a method to create an engine emission simulation model for cycle and customer driving of a vehicle. The emission model results from an empiric approach, also taking into account the effects of engine dynamics on emissions. We analysed transient engine emissions in driving cycles and during representative customer driving profiles and created emission meta models. The analysis showed a significantly higher correlation in emissions when simulating realistic customer driving profiles using the created verified meta models (< 1 % model error) compared to static approaches, which are commonly used for vehicle simulation. Therefore, a transient modelling approach is conducted, which shows a great increase in accuracy in customer driving operation.     

An adaptive lateral preview driver model

Successful modelling and simulation of driver behaviour is important for the current industrial thrust of computer-based vehicle development. The main contribution of this paper is the development of an adaptive lateral preview human driver model. This driver model template has a few parameters that can be adjusted to simulate steering actions of human drivers with different driving styles. In other words, this model template can be used in the design process of vehicles and active safety systems to assess their performance under average drivers as well as atypical drivers. We assume that the drivers, regardless of their style, have driven the vehicle long enough to establish an accurate internal model of the vehicle. The proposed driver model is developed using the adaptive predictive control (APC) framework. Three key features are included in the APC framework: use of preview information, internal model identification and weight adjustment to simulate different driving styles. The driver uses predicted vehicle information in a future window to determine the optimal steering action. A tunable parameter is defined to assign relative importance of lateral displacement and yaw error in the cost function to be optimized. The model is tuned to fit three representative drivers obtained from driving simulator data taken from 22 human drivers.     

基于主成分分析和学习向量量化神经网络的制动工况路面识别与验证

开展车辆制动时路面类型识别的研究，提出一种基于主成分分析-学习向量量化神经网络 （Principal Component Analysis - Learning Vector Quantization，PCA-LVQ） 的制动工况路面识别方法。利用主成分分析对多维度驾驶数据降维处理，提取能表征路面特征的主要成分，采用学习向量量化神经网络对降维处理后的驾驶数据进行训练，并用于路面特征分类，使用制动工况下实车试验数据和硬件在环仿真数据进行验证。结果表明，所提出的 PCA-LVQ算法能准确识别路面类型特征，路面识别的精度达到 97%，与传统 BP神经网络的路面类型特征识别精度提升 7%；同时，在不同车速下，基于PCA-LVQ算法也能较准确地识别路面类型特征。