考虑道路平曲线设计的驾驶疲劳对车道保持能力影响研究

不同的道路平面线形几何设计对于驾驶人车道保持能力的需求是有差异的,驾驶人受疲劳程度影响也会呈现车道保持能力下降的趋势,当前的研究未综合考虑以上2个因素:线形和疲劳程度对驾驶横向表现的交互影响.邀请41位被试者分别开展550 km的实车实验,获取车辆位置信息GPS以匹配道路线形类型,基于问卷调查方法获取驾驶过程疲劳等级.分析不同疲劳程度、不同平面线形类型以及弯道半径条件下的车道偏离标准差参数,构建了多元线性回归模型.数据分析结果表明,相同疲劳程度下驾驶人在圆曲线段驾驶的偏离值要超过直线段以及缓和曲线段;当弯道半径超过5 500 m时,曲线段弯道半径越大,车道偏离差值越高.同时,考虑了线形影响的多元线性回归模型对疲劳程度的预测精度要高于未考虑线形因素的模型,进一步说明在针对驾驶疲劳行为表现开展研究时,有必要对道路设计参数加以考虑以提高疲劳辨识精度.      

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。     

基于人体姿态估计的分心驾驶行为检测

为弥补现有驾驶特征提取方法的不足,提高分心驾驶行为检测的准确性和鲁棒性,将2D/3D人体姿态估计应用于驾驶人行为检测,提出一种适用于驾驶舱环境下的驾驶特征提取方法。首先通过将2D姿态估计网络Simple Baseline和分类网络ResNet进行融合,构建基于2D姿态估计的分心驾驶行为检测模型,并在分心驾驶数据集State Farm上分析不同数据增强方法、不同超参数、不同分类网络对模型性能的影响。其次,融合3D密集姿态估计网络DensePose与分类网络ResNet,构建基于3D姿态估计的分心驾驶行为检测模型。接着,在State Farm数据集上,针对模型的实时性和泛化能力,对比分析基于原始图像和基于2D/3D姿态的分心驾驶行为检测模型。最后,针对效果更优的基于2D姿态估计的分心驾驶行为检测模型,在分心驾驶数据集State Farm上,对使用不同姿态估计算法和分类网络的分心驾驶行为检测模型做了交叉试验,对比分析4个不同检测模型的优缺点。进一步地,将基于2D姿态估计的分心驾驶行为检测模型应用于实际采集的驾驶图片,对模型的泛化能力和有效性进行了测试验证。研究结果表明:与基于原始图像的检测模型相比,基于2D和3D姿态的检测模型都能显著提高分心驾驶行为的检测准确率;基于3D姿态的检测模型在检测精度方面略优,但基于2D姿态的检测实时性更好,检测效率是基于3D姿态检测的4倍;在驾驶舱单一环境下,基于2D姿态估计的分心驾驶行为检测模型能够满足分心驾驶行为检测的需求,在分心驾驶行为检测方面具有重要应用价值。     

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。     

分心对驾驶人交通冲突反应时间的影响(双语出版)

为了研究分心对交通冲突状态下驾驶人反应时间的影响,采用驾驶模拟器构建城市道路交通环境下2种典型冲突形态：侧向行人冲突和纵向追尾冲突,设计认知、视觉以及发短信（认知+视觉复合分心）3种分心任务,在不同行驶车速、跟车时距、前车减速度等紧迫度条件下,采集30名驾驶人应对交通冲突的制动反应时间,分别采用重复测量一般线性模型及线性混合模型进行统计分析。研究结果表明：认知分心使驾驶人应对侧向行人冲突的制动反应时间增加0.09 s,但未观察到其对纵向追尾冲突反应时间的显著性影响;视觉分心与发短信都会延缓驾驶人应对侧向行人（分别增加0.31 s和0.27 s）以及纵向追尾冲突（分别增加0.47 s和0.38 s）的制动反应时间;此外,在纵向追尾冲突中,随着冲突紧迫度提高（前车减速度增大、车头时距减小以及自车速度增大）,驾驶人制动反应时间显著减小。表明驾驶分心延长了驾驶人应对交通冲突的反应时间,容易导致事故的发生,具体而言,认知分心主要延长驾驶人应对侧向冲突的反应时间,涉及视觉的分心同时延长驾驶人应对侧向及纵向冲突的反应时间;视觉分心对驾驶人反应时间的延长显著性高于认知分心,说明视觉分心对行车安全影响更大。     

认知分心对车辆跟驰过程操控安全性的影响

为了研究车辆跟驰过程中驾驶人认知分心与驾驶安全的关系,采用驾驶模拟器构建城市道路车辆跟驰场景,并设计3种难度等级的认知分心次任务,采集35名被试驾驶人在试验过程中的方向盘转角、油门开度、制动踏板力等操作参数,以及车辆位置、速度、加速度等车辆运动参数。采用重复测量一般线性模型,分析不同等级认知分心对上述参数的影响。研究结果表明：在横向操控方面,随着认知分心程度增高,方向盘回转率增大,但车辆横向位置标准差减小,表明驾驶人处于认知分心时,采取频繁修正方向盘的补偿方式,降低车辆横向位置波动,过度补偿车辆横向安全性,且该补偿行为与认知分心程度正相关;在纵向操控方面,认知分心时,油门开度、制动踏板位置方差增大,且制动踏板位置均值增大,同时车头间距及时距未观察到显著性变化,表明认知分心时驾驶人采取频繁操作油门、制动踏板,增大制动幅度等方式进行补偿,使车头间距及车头时距等表征车辆纵向跟车安全性参数处于正常驾驶水平,但加速度标准差增大,表明跟车稳定性降低。研究结果为涉及分心的人车交互装置优化设计及考虑分心状态的驾驶人状态管理系统开发提供了一定的理论依据。     

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...     

基于注意力机制的多模态自动驾驶行为决策模型

针对现有端到端自动驾驶模型未考虑驾驶场景中不同区域的重要性和不同语义类别之间的关系而导致预测准确率低的问题，受驾驶人注意力机制和现有端到端自动驾驶模型的启发，充分考虑驾驶场景的动态变化、驾驶场景的语义信息和深度信息对驾驶行为决策的影响，以连续多帧驾驶场景的RGB图像为输入，构建一种基于注意力机制的多模态自动驾驶行为预测模型，实现对方向盘转角和车速的准确预测。首先，通过语义分割模型和单目深度估计模型分别获取RGB图像的语义图像和深度图像；其次，为剔除与驾驶行为决策无关信息，以神经科学和空间抑制理论为基础，设计一种拟人化注意力机制作为能量函数来计算驾驶场景中不同区域的重要度；为学习语义图像中与驾驶行为决策最为相关类别之间的关系，采用图注意力网络(Graph Attention Network,GAT)对驾驶场景的语义图像进行特征提取；然后，以保留RGB特征为原则对提取的驾驶场景的图像特征、语义特征和深度特征进行融合，采用卷积长短期记忆网络(Convolutional Long Short Term Memory,ConvLSTM)实现融合特征在连续多帧之间的传递，进而实现下一帧驾驶场景对应驾驶行为的预测；最后，与其他模型的对比试验、消融试验、泛化试验和特征可视化试验来充分验证所提出自动驾驶行为预测模型的性能。试验结果表明：与其他驾驶行为预测模型相比，所提出模型的训练误差为0.021 2，预测准确率为86.97%，均方误差为0.031 5，其驾驶行为的预测性能优于其他模型；连续多帧的语义图像和深度图像、拟人化注意力机制和面向语义特征提取的GAT有助于提升驾驶行为预测的性能；该模型具有较好的泛化能力,其做出驾驶行为预测所依赖的特征与经验丰富的驾驶人所关注的特征基本一致。     

Highly reliable driving workload analysis using driver electroencephalogram (EEG) activities during driving

Traffic accidents are caused by various factors, which can be classified into human factors, vehicle factors and environmental factors. Recently, human factors have been drawing particular attention as efforts are being made to enhance the safety performance of vehicles and improve road conditions. Driving distraction caused by an increased driving workload is a representative human factor. Various studies in the past have attempted to quantify the driving workload by using EEG activities. However, they have failed to consider vibration properties generated from vehicle engines. A number of noise signals were included in brainwave signal processing, which resulted in a failure to obtain reliable outcomes. Thus, this study suggests driver EEG activities free of vehicle engine secondary vibration in order to develop a method that analyzes the driving workload with high statistical reliability. By using the analytical method developed in this study, standard values of driving workload for straight and left-turn driving that has statistical significance could be calculated. The analytical method for driving workload created by this study can be applied to HVI and road design.     

On-road assessment of in-vehicle driving workload for older drivers: Design guidelines for intelligent vehicles

There has been recent interest in intelligent vehicle technologies, such as advanced driver assistance systems (ADASs) or in-vehicle information systems (IVISs), that offer a significant enhancement of safety and convenience to drivers and passengers. However, the use of ADAS- and IVIS-based information devices may increase driver distraction and workload, which in turn can increase the chance of traffic accidents. The number of traffic accidents involving older drivers that are due to distraction, misjudgment, and delayed detection of danger, all of which are related to the drivers’ declining physical and cognitive capabilities, has increased. Because the death rate in traffic accidents is higher when older drivers are involved, finding ways to reduce the distraction and workload of older drivers is important. This paper generalizes driver information device operations and assesses the workload while driving by means of experiments involving 40 drivers in real cars under actual road conditions. Five driving tasks (manual only, manual primarily, visual only, visual primarily, and visual-manual) and three age groups (younger (20–29 years of age), middle-aged (40–49 years of age), and older (60–69 years of age)) were considered in investigating the effect of age-related workload difference. Data were collected from 40 drivers who drove in a real car under actual road conditions. The experimental results showed that age influences driver workload while performing in-vehicle tasks.     

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.     

Assessment of risky driving caused by the amount of time focused on an in-vehicle display system

Korea is currently experiencing a rapidly increasing distribution rate of in-vehicle display devices, such as navigation or DMB displays, owing to remarkable advances in IT. At the same time, the number of traffic accidents and traffic violations is increasing due to the distraction of drivers’ attention by such devices. In particular, in-vehicle display devices such as navigation systems temporarily distract drivers’ visual or cognitive attention when they perform a unit task. Accordingly, it is necessary to prepare adequate standards to regulate in-vehicle display devices, especially in Korea. There are few empirical studies that have employed experiments to support such regulation. In this study, an experiment was conducted using a driving simulator to establish the proper standards regarding the maximum distraction time per unit task that can be allowed without causing any disturbance in safe driving. A total of 25 participants participated in the experiment. The distraction time was controlled by asking participants to perform the two tasks at once: while participants were driving as a primary task, they performed secondary task that count the number of intersections between the start point and the arrival point displayed on the screen. The results showed that the 2.0 second condition differed from the controlled condition in the deviation in the distance from the preceding vehicle, speed, and steering wheel movement, whereas there were no differences between the controlled condition and the 1.0 or 1.5 second condition. Finally, the limitations of the study and the implications of the findings with regard to future studies and application of the Korean version of guidelines for in-vehicle display devices are discussed.     

Which factors lead to driving errors? A structural equation model analysis through a driving simulator experiment

As driving error is a main contributory factor of road accidents, its causes and consequences are of great interest in the road safety decision making process. This paper investigates several factors (including driver distraction, driver characteristics and road environment) that affect overall driving error behaviour and estimates a new unobserved variable which underlines driving errors. This estimation is performed with data obtained from a driving simulation experiment in which 95 participants covering all ages were asked to drive under different types of distraction (no distraction, conversation with passenger, cell phone use) in rural and urban road environment, as well as in both low and high traffic conditions. Driving error was then modeled as a latent variable based on several individual driving simulator parameters. Subsequently, the impact of several risk factors such as distraction, driver characteristics as well as road environment on driving error were estimated directly. The results of this complex model reveal that the impact of driver characteristics and area type are the only statistically significant factors affecting the probability of driving errors. Interestingly, neither conversing with a passenger nor talking on the cell phone have a statistically significant impact on driving error behaviour which highlights the importance of the present analysis and more specifically the development of a measure that represents overall driving error behaviour instead of individual driving errors variables.     

平直高速公路低交通量下L2自动驾驶对驾驶人心理负荷的影响

为研究驾驶人在L2自动驾驶模式下的心理负荷特性,设计了正常驾驶和次任务驾驶2种状态,进行实车高速道路试验,采集21名被试驾驶人在2种驾驶状态下分别选择手动驾驶和自动驾驶模式的眼动数据、次任务绩效和主观评价数据.采用重复测量一般线性模型,分析不同驾驶模式对上述参数的影响,从客观和主观两方面分析驾驶人的心理负荷变化.结果表...     

Driver behavior formulation in intersection dilemma zones with phone use distraction via a logit-Bayesian network hybrid approach

Use of cellular phone while driving is one of the top contributing factors that induce traffic crashes, resulting in significant loss of life and property. A dilemma zone is a circumstance near signalized intersections where drivers hesitate when making decisions related to their driving behaviors. Therefore, the dilemma zone has been identified as an area with high crash potential. This article utilizes a logit-based Bayesian network (BN) hybrid approach to investigate drivers' decision patterns in a dilemma zone with phone use, based on experimental data from driving simulations from the National Advanced Driving Simulator (NADS). Using a logit regression model, five variables were found to be significant in predicting drivers' decisions in a dilemma zone with distractive phone tasks: older drivers (50–60 years old), yellow signal length, time to stop line, handheld phone tasks, and driver gender. The identified significant variables were then used to train a BN model to predict drivers' decisions at a dilemma zone and examine probabilistic impacts of these variables on drivers' decisions. The analysis results indicate that the trained BN model was effective in driver decision prediction and variable influence extraction. It was found that older drivers, a short yellow signal, a short time to stop line, nonhandheld phone tasks, and female drivers are factors that tend to result in drivers proceeding through intersections in a dilemma zone with phone use distraction. These research findings provide insight in understanding driver behavior patterns in a dilemma zone with distractive phone tasks.     

Application of a driving simulator to the development of in-vehicle human-machine-interfaces

The use of a driving simulator in the development of human-machine-interfaces (HMI) such as a navigation, information or entertainment system is discussed. Such use addresses the need to study and evaluate the characteristics of a candidate HMI early in the R&D and design stage to ensure that it is likely to meet various objectives and requirements, and to revise the HMI as may be necessary. Those HMI requirements include such things as usability, driver comfort, and an acceptable level of attentional demand in dual task conditions (driving while using an HMI). Typically, such an HMI involves an information display to the driver, and a means for driver input to the HMI. Corresponding simulator requirements are discussed, along with typical simulator features and components. The latter include a cab, control feel systems, visual image generator, real time scenario control (task definitions), a motion system (if provided), and data acquisition. Both fixed and moving base systems are described, together with associated benefits and tradeoffs. Considerations in the design of the evaluation experiment are discussed, including definition of primary and secondary tasks, and number of driver subjects (experimental participants). Possible response and performance measures for the primary and secondary tasks are noted, together with subjective measures such as task difficulty and ease of using the HMI. The advantages of using a driving simulator to support R&D are summarized. Some typical and example simulator uses are noted.     

基于车道线虚线角点检测的行车安全视距测算模型

开展行车视距调查对于营运期公路安全评价至关重要,这对车载条件下行车视距检测提出了要求.针对现有基于车道线图像特征点所构建的视距模型精确度不高的问题,提出了1种以车道线虚线角点为关键特征的行车安全视距测算模型.在车载设备获取的图像预处理基础上,采用轮廓跟踪法对车道线虚线轮廓进行提取,通过设定轮廓尖锐度阈值以实现对车道线虚...     

小样本下基于迁移学习与LSTM的雾天高速公路车辆跟驰模型

由于在现实生活中能够采集到的不同雾天等级的高速公路车辆跟驰样本有限,导致雾天跟驰模型精度不佳,为此在长短时记忆神经网络（long short-term memory,LSTM）跟驰模型的基础上,采用迁移学习（transfer learning,TL）方法来提升雾天跟驰模型的性能。利用驾驶模拟实验平台搭建高速公路雾天与正常天气2种实验场景进行驾驶模拟实验,获得296组正常天气下（源域）的跟驰样本与100组雾天下（目标域）的跟驰样本。提出了基于最长公共子序列（longest common sequence solution,LCSS）的迁移样本选择方法,从源域中选出100个样本迁移至目标域中,通过扩大训练样本提升LSTM从源域、目标域特征到目标域输出的端对端泛化学习能力,得到雾天高速公路车辆跟驰模型。为对比所提样本迁移方法对LSTM模型的效用,将LSTM-TL模型与训练样本全部来源于源域的LSTM-S模型和训练样本全部来源于目标域的LSTM-T模型进行对比,LSTM-TL模型的均方误差、均方根误差和平均绝对误差比LSTM-S模型分别减小47.5%、27.7%和46.5%,比LSTM-T模型...     

eTS fuzzy driver model for simultaneous longitudinal and lateral vehicle control

In this paper, evolving Takagi-Sugeno (eTS) fuzzy driver model is proposed for simultaneous lateral and longitudinal control of a vehicle in a test track closed to traffic. The developed eTS fuzzy driver model can capture human operator’s driving expertise for generating desired steering angle, throttle angle and brake pedal command values by processing only information which can be supplied by the vehicle’s on-board control systems in real time. Apart from other fuzzy rule based (FRB) models requiring human expert knowledge or off-line clustering, the developed eTS driver model can adapt itself automatically, even ‘from scratch’, by an on-line learning process using eTS algorithm while human driver is supervising the vehicle. Proposed eTS fuzzy driver model’s on-line human driver identification capability and autonomous vehicle driving performance were evaluated on real road profiles created by digitizing two different intercity express ways of Turkey in IPG© CarMaker® software. The training and validation simulation results demonstrated that eTS fuzzy driver model can be used in product development phase to speed up different tests via realistic simulations. Furthermore eTS fuzzy driver model has an application potential in the field of autonomous driving.     

Hierarchical map representation using vector maps and geometrical maps for self-localization

Map-based self-localization estimates the pose of the self-driving vehicle in an environment, becoming an essential part of autonomous driving tasks. Generally, maps used in self-localization have detailed geometric information on an environment in formats such as point cloud maps and Gaussian mixture model (GMM) maps. As other maps are widely developed for autonomous driving, vector maps store more object-focused information, such as buildings and road facilities, for navigation and scene understanding in autonomous driving tasks. However, it is not compatible with self-localization due to the lack of detailed geometric information. The two different map formats of vector maps and maps for self-localization complicate the management, preventing the development of the area where a self-driving vehicle can drive stably. This paper proposes a unified map format with a hierarchical structure that enables both vector maps and self-localization maps (i.e., GMM maps) to be managed more easily. Because proposed maps can be treated as vector maps at the high-level layer, various tasks related to navigation and scene understanding in autonomous driving can utilize. A GMM map is stored at the low-level layer associated with a vector map component, enabling accurate self-localization in an environment. The proposed map format is compatible with vector maps widely developed by mapping companies on the surface and facilitates future map management. The experimental results of self-localization in urban areas showed that the proposed map gives the competitive self-localization accuracy compared with the GMM map even with fewer cells that link to vector components. The proposed maps enable self-localization with sufficient accuracy for safe autonomous driving operations.