Modeling Human Vehicle Driving by Model Predictive Online Optimization

A driver model is designed which relates the driver's action to his perception, driving experience, and preferences over a wide range of possible traffic situations. The basic idea behind the work is that the human uses his sensory perception and his expert knowledge to predict the vehicle's future behavior for the next few seconds (prediction model). At a certain sampling rate the vehicle's future motion is optimized using this prediction model, in order to meet certain objectives. The human tries to follow this optimal behavior using a compensatory controller. Based on this hypothesis, human vehicle driving is modeled by a hierarchical controller. A repetitive nonlinear optimization is employed to plan the vehicle's future motion (trajectory planning task), using an SQP algorithm. This is combined with a PID tracking control to minimize its deviations. The trajectory planning scheme is experimentally verified for undisturbed driving situations employing various objectives, namely ride comfort, lane keeping, and minimized speed variation. The driver model is then applied to study path planning during curve negotiation under various preferences. A highly dynamic avoidance maneuver (standardized ISO double lane change) is then simulated to investigate the overall stability of the closed loop vehicle/driver system.     

面向前车的驾驶行为感知与意图识别算法研究

感知周围车辆的驾驶行为并识别其意图将成为新一代高级驾驶辅助系统的重要组成部分。针对现有方法只考虑单一驾驶行为且可扩展性和可伸缩性差,提出一种基于稀疏表示理论的驾驶行为感知字典模型(Driving Behavior Perception Dictionary Model, DBPDM)。将车辆行驶状态视为时间序列,设计基于自回归积分移动平均(Autoregressive Integrated Moving Average,ARIMA)结合在线梯度下降(Online Gradient Descent, OGD)优化器的在线预测模型,提出基于驾驶行为预测的意图识别构架(Intention Recognition Framework, IRF)。首先,采用图Lasso方法估计典型驾驶行为的稀疏逆协方差矩阵构建驾驶行为字典库,并采用Logdet散度方法计算各逆协方差矩阵的差异获得行为感知字典模型。然后,基于在线预测模型对目标车辆的行驶轨迹和运动状态进行预测,结合主车车辆的行驶状态作为稀疏表示的观测信号,以获取预测时域内的目标车辆意图。最后,采用NGSIM (Next Generation SIMulation)真实驾驶数据对模型进行开发和测试。研究结果表明:所提出的行为感知模型能对6种典型驾驶行为构建行为字典,在分类准确率上与现有方法相比有明显提升,对换道和转向行为样本的平均识别准确率分别达到99.1%和92.9%;该模型能够在相对早期阶段准确地识别出车辆行为;在线预测算法能较好预测出目标车辆的行驶轨迹和运动状态,从而间接地反映出其在预测时域内的驾驶意图;IRF可在换道和转向行为开始前的1.5 s较为准确地识别出目标车辆的意图,平均识别准确率超过80%。     

On model-based longitudinal feed-forward motion control for low velocities on known road profiles

ABSTRACT

So far, longitudinal motion control has focused on situations like highway driving, where disturbances of the road profile can be neglected. In this paper, we show how the Two Point Tire Model can be used to derive a novel feed-forward control law for a vehicle's longitudinal motion that considers the effects of the road profile and can complement existing control approaches. For this purpose, we recapitulate the basic model assumptions and equations and briefly discuss how it can be used on arbitrary road profiles. Two approaches for implementation in a real vehicle are presented. Comparisons of these approaches in simulation and to a human driver of an experimental vehicle show that the controller can deal with stepped obstacles of up to 14?cm in height. However, the control performance is essentially limited by the actuator delay and human drivers outperform the controller due to their ability of sensing subtle vehicle motions. The results indicate that the control performance can be further improved by using a preview on the necessary drive torque, which can be provided by the solution that we propose.     

Steering disturbance rejection using a physics-based neuromusculoskeletal driver model

The aim of this work is to develop a comprehensive yet practical driver model to be used in studying driver–vehicle interactions. Drivers interact with their vehicle and the road through the steering wheel. This interaction forms a closed-loop coupled human–machine system, which influences the driver's steering feel and control performance. A hierarchical approach is proposed here to capture the complexity of the driver's neuromuscular dynamics and the central nervous system in the coordination of the driver's upper extremity activities, especially in the presence of external disturbance. The proposed motor control framework has three layers: the first (or the path planning) plans a desired vehicle trajectory and the required steering angles to perform the desired trajectory; the second (or the musculoskeletal controller) actuates the musculoskeletal arm to rotate the steering wheel accordingly; and the final layer ensures the precision control and disturbance rejection of the motor control units. The physics-based driver model presented here can also provide insights into vehicle control in relaxed and tensed driving conditions, which are simulated by adjusting the driver model parameters such as cognition delay and muscle co-contraction dynamics.     

人机协作系统中车辆轨迹规划与轨迹跟踪控制研究

自动驾驶系统需具备响应驾驶人意图且有效执行驾驶人意图的能力，以解决人机协作系统中存在的人机冲突、人机优势融合等问题。提出决策层“以人为主”、执行层“以机为首”的人机协作关系，构建包含驾驶人意图识别模块、基于意图识别的轨迹规划模块与轨迹跟踪控制模块的人机协作一体化控制系统框架，并重点对轨迹规划模块与轨迹跟踪控制模块开展研究。首先，结合双向长短期记忆神经网络（Bi-directional Long Short Term Memory，Bi-LSTM）与注意力机制模型建立换道轨迹规划模型；在改进人工势场算法中引入模型预测控制并建立避险轨迹规划模型。其次，通过开展驾驶模拟器试验建立换道与避险驾驶行为数据集，为拟人化模型训练和模型参数确定提供支撑。然后，综合考虑车辆状态变量、控制输入与输出以及道路结构参数等约束条件，构建基于最优转向前轮输入的线性时变模型预测轨迹跟踪控制器，实现对规划轨迹的精准跟踪。最后，基于驾驶模拟器搭建人机协作系统硬件在环测试平台，对轨迹规划模块与轨迹跟踪控制模块开展硬件在环测试与验证。结果表明：换道与避险规划轨迹光滑且平稳，轨迹跟踪控制过程中，车辆航向角与前轮转角变化平稳；所构建的轨迹规划与轨迹跟踪控制模块在确保安全性前提下可实现不同场景中的车辆运动控制需求。     

A novel integrated chassis controller for full drive-by-wire vehicles

In this paper, a systematic design with multiple hierarchical layers is adopted in the integrated chassis controller for full drive-by-wire vehicles. A reference model and the optimal preview acceleration driver model are utilised in the driver control layer to describe and realise the driver's anticipation of the vehicle's handling characteristics, respectively. Both the sliding mode control and terminal sliding mode control techniques are employed in the vehicle motion control (MC) layer to determine the MC efforts such that better tracking performance can be attained. In the tyre force allocation layer, a polygonal simplification method is proposed to deal with the constraints of the tyre adhesive limits efficiently and effectively, whereby the load transfer due to both roll and pitch is also taken into account which directly affects the constraints. By calculating the motor torque and steering angle of each wheel in the executive layer, the total workload of four wheels is minimised during normal driving, whereas the MC efforts are maximised in extreme handling conditions. The proposed controller is validated through simulation to improve vehicle stability and handling performance in both open- and closed-loop manoeuvres.     

Motion cueing in high-performance vehicle simulators

The majority of motion cueing algorithms have been developed for passenger car applications, with correspondingly less research dedicated to race-car and high-performance vehicle simulators. In the high-performance context, the focus is on cueing the vehicle's behavioural and handling characteristics, particularly when driving near the limits of performance. In race-car simulators the cueing requirements are therefore quite different, with the problem made all the more challenging by the presence of large accelerations. To understand the drivers' cueing needs, the vehicle's stability and handling response characteristics must be examined near the performance boundary. Frozen-time eigenvalue analyses are used to determine stability and response characteristics across all vehicle operating conditions, including accelerating and braking under cornering, with the results used to determine motion cueing algorithm requirements. Lateral acceleration and yaw cueing filters are designed in order to retain information critical to understanding the vehicle's behaviour on its performance boundary. Cueing filters are then tested, with the help of a professional race-car driver, and are found to provide the cues necessary for the driver to control the vehicle on the limit of performance.     

面向全矢量线控车辆极限运动的分层式协同控制

装配四轮分布式驱动-转向（4WID-4WIS）底盘的全矢量线控车辆具备多可控自由度、高速稳定性强的特点,是极限工况稳定裕度和安全性较高的理想车型。为了解决全矢量线控车辆在极限工况下纵横向控制冲突危害行车安全的问题,提出一种基于模型预测控制 （MPC） 的分层式车辆纵向和横向运动协同控制方法。建立基于单轨模型的期望运动状态识别方法,设计模型预测控制器转换动力学目标,采用泰勒展开和前向欧拉方法对预测模型进行线性离散化处理;设计基于负荷率的轮胎力优化分配方法,利用反正切轮胎逆模型求解控制执行量。仿真结果表明,协同控制方法能显著提高车辆在不同路面下的极限运动稳定性,更精准地跟踪期望运动状态,扩大稳定裕度,保障行车安全。     

基于前车轨迹预测的高速智能车运动规划

针对现有运动规划算法大多只考虑障碍车当前状态,本文中提出一种基于前车运动轨迹预测的高速车辆运动规划算法。首先,融合考虑驾驶意图与基于车辆运动模型的方法对前车轨迹进行预测;然后,采用贝塞尔曲线(Bezier)规划主车运动轨迹,结合避撞过程中与前车碰撞风险概率,高速避撞车辆速度变化特点以及车辆运动稳定性等因素建立目标函数,并考虑车辆动力学与运动学约束,使用序列二次规划(SQP)方法对Bezier曲线的控制点和主车运动目标点位置进行优化求解,得到最优避撞运动轨迹;最后,以前车直行和换道两种工况为例,对主车的避撞运动轨迹进行规划,分析不同工况下主车避撞过程中的运动状态变化以及与前车碰撞风险概率变化。结果表明,所提出的运动规划算法能够保证车辆的避撞安全性与运动稳定性。     

From robotics to automotive: Lane-keeping and collision avoidance based on elastic bands

In the near future, drivers will more and more share vehicle guidance with assistance systems. This contribution provides a potential field-based approach to the underlying motion planning problem. In doing so, the concept of elastic bands, known from robotics, is extended to automotive applications. Contrary to robotic applications, extrapolation routines anticipating the motion of the surrounding traffic are incorporated in the motion planning. New in this paper is the distinction of different types of obstacles such as traffic staying in its lane and traffic intending to depart from it. Beyond that, the motion planning adapts to the driver's commands. The driver can be included in the overall control loop by means of a haptic interface generating a torque that depends on the difference of the actual steering angle and the steering angle necessary to follow the planned trajectory. However, this contribution focuses only on the underlying motion planning procedure.     

智能网联汽车运动规划方法研究综述（双语出版）

分析了近年来智能网联汽车（Intelligent Connected Vehicle，ICV）运动规划方法的研究，根据规划时空尺度和任务目标，将ICV运动规划细分为路径规划、路线规划、动作规划和轨迹规划4级子任务，回顾了各级子任务中智能网联技术的研究和应用现状；探讨了ICV中驾驶人行为特性及其对运动规划结果的影响；从技术背景、研究场景、算法流程和应用理论4个方面，提出ICV运动规划方法研究的未来发展方向。结果表明：由于ICV主要依赖车辆网联信息规划运动路径，而路网中同时存在不同网联等级的ICV，这将增加路径规划问题的求解难度；现有ICV路线规划模型较少考虑周边多车运动状态以及路段车道设置情况，将现有算法与微观交通流模型相结合有助于解决此问题；ICV中人机协同及任务切换领域已出现诸多研究热点，如城市道路上换道与转弯动作规划、ICV引导非网联车辆行驶等问题；借鉴驾驶人行为模式规划ICV运动轨迹已成为研究共识，但是车-车、车-路网联信息在此领域的应用仍然有限；采用反馈-迭代的方法进行ICV运动路线和动作协同规划、运动规划和轨迹跟踪控制有助于获得全局最优的运动规划结果和车辆控制策略；根据具体规划任务特点选择构建ICV运动规划模型的基础理论，有助于发挥各类理论的优势，提升规划算法的灵活性和适用性。     

基于状态空间采样的高速公路智能网联车辆轨迹动态规划

为实现智能网联车辆在高速公路动态行车环境下的轨迹实时规划，提出一种基于状态空间采样的轨迹动态规划方法。首先，以安全性为原则选取主车当前行驶的理想车道。基于Frenet坐标与笛卡尔坐标的转换关系，建立车辆运动横、纵向解耦的独立积分系统。将高速公路常见的行驶状态分为车道保持与定速巡航、变道以及前车跟随3类，预测主车行驶车道并针对3类行驶状态分别设计轨迹终端的目标配置方法。然后，利用多项式函数生成连接初始配置和目标配置的多条待选轨迹。构建考虑轨迹偏离理想车道程度、始末速度变化、规划周期和轨迹舒适性的综合损失函数，结合速度、加速度、曲率检查来评价各条待选轨迹的成本并进行排序。最后，预测车辆的横、纵向运动轨迹并构建一种胶囊形的车辆虚拟安全边界，通过碰撞检测，确定主车的最优轨迹，设置动态规划触发条件及时更新最优轨迹并避免过度规划浪费资源。研究结果表明：提出的算法能满足高速公路场景的动态规划需求；通过对轨迹规划周期、虚拟安全边界、动态规划时间间隔等关键参数的分析与优化，主车的横摆角速度范围稳定在-0.1~0.15 (°)·s^-1，横向加速度范围稳定在-0.16~0.32 m·s^-2，跟踪参考轨迹的最大误差不超过0.022 m，提出的算法能规划出具有高安全性、稳定性和舒适性的轨迹。     

Modelling of a driver's behaviour considering roll motion

This paper presents an anlysis of the control behaviour of a driver during curves and lane changes. We model the driver's behaviour taking the roll motion of the vehicle into consideration. Using this model with constraints on the roll angle, it is possible to model lane change maneuvers without specifying a path. The validity of the model is investigated through a comparison between computer simulation and experimentation using a driving simulator system.     

New level of vehicle comfort and vehicle stability via utilisation of the suspensions anti-dive and anti-squat geometry

In this article a novel vehicle dynamics control concept is designed for a vehicle equipped with wheel individual electric traction machines, electronically controlled brakes and semi-active suspensions. The suspension's cross-couplings between traction forces and vertical forces via anti-dive and anti-squat geometry is utilised in the control concept to improve driving comfort and driving stability. The control concept is divided into one main and two cascaded branches. The main controller consists of a multivariable vehicle dynamics controller and a control allocation scheme to improve the vehicle's driving comfort. The cascaded feedback loops maintain the vehicle's stability according to wheel slip and vehicle sideslip. The performance of the combined vehicle dynamics controller is compared to a standard approach in simulation. It can be stated that the controller piloting semi-active suspensions together with brake and traction devices enables a superior performance regarding comfort and stability.     

基于注意力Seq2Seq网络的高速公路交织区车辆变道轨迹预测

针对交通状态复杂的高速公路交织区域,经验丰富的驾驶人能够通过正确地推断周围车辆的未来运动进行及时的车道变换,这对于实现安全高效的自动驾驶至关重要,然而目前的自动驾驶车辆往往缺乏这种预测能力。为此,基于深度学习理论,提出了一种结合注意力机制和编-解码器结构的交织区车辆强制性变道轨迹预测方法,利用Next Generation Simulation（NGSIM）数据集提取车辆变道过程中的关键特征,并引入碰撞时间（Time to Collision,TTC）和避免碰撞减速度（Deceleration Rate to Avoid a Crash,DRAC）2种风险指标,将变道车辆及其周围车辆视为一个整体状态单元,同时补全状态单元内部不同车辆在横向和纵向上的时空状态特征,从而更有效地刻画车辆间的动态交互行为;然后将不同观测车辆的连续窗口序列输入基于长短期记忆网络（Long Short-term Memory,LSTM）的编-解码器,预测交织区车辆变道的未来运动轨迹,通过添加软注意力模块,使模型能够集中聚焦于影响车辆在不同时刻下位置变化的关键信息,再现了真实交通场景下车辆的变道行为。试验验证表明：基于注意力机制的编-解码器模型与当前流行的卷积长短期记忆网络、极限梯度提升树等模型相比具有更高的轨迹预测精度,在长时域的变道轨迹拟合上有显著的优越性,为辅助和自动驾驶领域的发展提供了新思路。     

基于LSTM网络的驾驶意图识别及车辆轨迹预测

自动驾驶汽车需具备预测周围车辆轨迹的能力,以便做出合理的决策规划,提高行驶安全性和乘坐舒适性。运用深度学习方法,设计了一种基于长短时记忆（LSTM）网络的驾驶意图识别及车辆轨迹预测模型,该模型由意图识别模块和轨迹输出模块组成。意图识别模块负责识别驾驶意图,其利用Softmax函数计算出驾驶意图分别为向左换道、直线行驶、向右换道的概率;轨迹输出模块由编码器-解码器结构和混合密度网络（MDN）层组成,其中的编码器将历史轨迹信息编码为上下文向量,解码器结合上下文向量和已识别的驾驶意图信息预测未来轨迹;引入MDN层的目的是利用概率分布来表示车辆未来位置,而非仅仅预测一条确定的轨迹,以提高预测结果的可靠性和模型的鲁棒性。此外,将被预测车辆及其周围车辆组成的整体视为研究对象,使模型能够理解车-车间的交互式行为,响应交通环境的变化,动态地预测车辆位置。使用基于真实路况信息的NGSIM（Next Generation SIMulation）数据集对模型进行训练、验证与测试。研究结果表明：与传统的基于模型的方法相比,基于LSTM网络的轨迹预测方法在预测长时域轨迹上具有明显的优势,考虑交互式信息的意图识别模块具备更高的预判性和准确率,且基于意图识别的轨迹预测能降低预测轨迹与真实轨迹间的均方根误差,显著提高轨迹预测精度。     

网联车辆并线预测与巡航控制的研究

为检测旁车道车辆驾驶员的并线意图,提升网联车辆巡航跟车的主动安全性,提出了一种基于NAR神经网络学习的迭代循环预测算法。NAR神经网络的训练样本由实际交通环境中的车辆并线数据获得,通过训练的网络预测未来一段时间内旁车的横向行驶轨迹,并根据划定的监控区域计算旁车的切入概率。同时,提出了一种考虑并线概率的跟车距离策略,并应用到网联车辆CACC系统中。结果表明,所提出的并线预测算法能精确计算出旁车的横向换道轨迹,所提出的跟车策略可提升车辆的跟车安全性。     

基于车辆动力学模型的换道轨迹规划研究

针对自动驾驶车辆换道轨迹规划时的操纵稳定性问题，基于CarSim/Simulink仿真平台建立了车辆动力学模型，构建了轨迹规划系统框架，通过轨迹信息后处理并提出了目标函数设计，进行了横向控制序列采样以保证车辆的稳定与极限性能，完成了算法对轨迹的综合评价选优。随后开展了仿真试验，对比分析了轨迹跟踪控制系统下的实际轨迹、最优规划方法所规划的换道轨迹。仿真结果表明，该轨迹规划系统框架及算法模型能有效提高车辆的操纵稳定性，可实现冰雪路面等极端工况下自动驾驶车辆换道轨迹规划。     

Investigation of a passenger car's dynamic response due to a flywheel-based kinetic energy recovery system

With the advent of flywheel-based kinetic energy recovery systems in automotive applications new safety issues arise as a consequence of the flywheel's high rotational speed. While the special structural safety requirements of the components are well discussed in the literature, there is still little research on the influence of gyroscopic effects on vehicle dynamics. The aim of this paper is to investigate the influence of a typical high-speed flywheel on the driving dynamics of an average passenger car. To this end the equations of motion of a gyroscope are derived, which relate the vehicle's roll, pitch and yaw rate with the transverse torque acting on the flywheel. These equations are implemented in a commercial vehicle dynamics simulation program in order to determine the reaction torques acting on the vehicle within a representative range of driving situations. Numerical simulations indicate that the gyroscopic effect can be considered insignificant in standard driving situations.     

考虑驾驶人风格的换道轨迹规划与控制（双语出版）

基于自动换道控制技术中融合个性化驾驶人风格的研究,建立考虑驾驶人风格的车辆换道轨迹规划及控制模型以提高换道规划控制模型对不同风格驾驶人的适用性,在保证安全性的基础上进一步满足驾驶人的个性化需求。首先通过问卷调查的方式采集得到了212份驾驶人风格量表数据,采用主成分分析法和K均值（K-means）聚类分析法将驾驶人按驾驶风格分为激进型、普通型和谨慎型,并通过驾驶模拟器试验采集不同风格驾驶人分别在自车道前车、目标车道前车和目标车道后车影响下的换道行为数据。然后对椭圆车辆模型进行改进,以描述不同风格驾驶人的行车安全区域,并据此构建3种典型工况下不同风格驾驶人的换道最小安全距离模型,结合驾驶舒适性约束、车辆几何位置约束以及不同风格驾驶人的换道行为数据,以换道纵向位移最短为目标,实现适应驾驶人风格的换道轨迹规划。最后以基于预瞄的路径跟踪模型作为前馈量,设计基于动力学的线性二次型最优（LQR）反馈控制器,通过调节控制权重矩阵实现3种工况下不同驾驶人风格的换道轨迹跟踪。PreScan和MATLAB/Simulink联合仿真结果表明：所设计的考虑驾驶人风格的换道轨迹规划及跟踪控制模型能够实现不同驾驶风格的自动换道轨迹规划及跟踪控制,可满足驾驶人个性化换道需求。