Traffic flow simulation|10,000 words summary

1 Introduction

Virtualized traffic based on various simulation models and real traffic data is an ideal method to reconstruct traffic flow. Virtual transportation has great benefits for video games, virtual reality, traffic engineering, autonomous driving, and more.

This article first discusses three different levels of traffic simulation models; secondly introduces data-driven virtual traffic construction technology ; again discusses how to apply traffic simulation to the training and testing of autonomous vehicles ; and finally discusses the The current status of the study is studied, and future research directions are proposed.

1.1 Why study traffic simulation?

In recent years, visual traffic has received more and more attention. Among them, a large number of vehicles will inevitably be involved when constructing urban scenes. To control the movement of a single vehicle, a simple solution is to use keyframe methods. However, when simulating traffic congestion, frequent lane changes, and pedestrian-vehicle interactions in large-scale traffic scenes, the keyframe method not only requires designers to perform complex designs and repeated adjustments, but the generated vehicle motion generally does not conform to physics. law. Therefore, effectively simulating large-scale traffic flow has become an important topic in computer graphics. In addition, due to the popularity of road network visualization tools such as OpenStreetMap, ESRI, and Google Maps, it has become crucial to integrate real-time traffic flows into virtual road networks . However, it is very difficult to obtain the actual trajectories of vehicles in real time and integrate them into virtual applications. These trends have promoted research efforts in data-driven traffic simulation.

In addition to the above applications, traffic simulation has a wide range of applications in traffic research.

• Driving training programs based on virtual reality can help new drivers improve their driving skills by generating realistic traffic environments.

• Traffic simulation can also serve as an effective tool for generating various traffic conditions for training and testing autonomous vehicles.

• The growing traffic volume and complex road network have led to many traffic-related problems, such as traffic congestion, accident management, signal control, and network design optimization.

These problems are difficult to solve with traditional tools based on analytical models. Therefore, people try to use advanced computing technologies to model, simulate, and visualize traffic to analyze traffic conditions for traffic management, or to help traffic reconstruction in urban development.

1.2 What issues need to be studied

A major focus of traffic simulation is to answer the following question: given a road network, behavioral model, and initial vehicle states, how will traffic evolve?

There are a large number of mathematical descriptions for the modeling and simulation of traffic flow, which can be roughly divided into macro models, micro models, and meso models . The macro model treats the collection of vehicles as a continuous flow, and the micro model simulates the dynamics of each vehicle under the influence of the vehicles around it. Mesoscale models combine the advantages of micro and macro models to simulate different levels of traffic details. In addition, the generation and representation of road networks is also a fundamental issue in traffic simulation.

The previously mentioned traffic models can effectively capture the appearance of traffic flow, but the resulting simulation results are often unrealistic. With the development of sensor hardware and computer vision technology, more and more traffic flow data exists in the form of video, lidar and GPS sensors. This phenomenon gave rise to data-driven traffic animation technology. Examples include: reconstructing spatiotemporal data acquired from existing road traffic flow sensors; synthesizing new traffic flows from limited sample trajectories; and learning behavioral patterns and independent features from traffic monitoring data sets to generate traffic flows.

There is also insufficient research on how to measure the realism of simulated traffic . In order to solve these problems, the current mainstream methods include using subjective user evaluation methods and incorporating objective evaluation indicators into the measurement.

Virtual traffic is also used in autonomous driving training. Currently, when testing autonomous driving performance, it is common to use only a single road user with predefined behavior such as a vehicle, pedestrian or bicycle (Translator's Note: single actor). By training in simulated traffic flows with rich interactions between different road users, autonomous vehicles have the potential to acquire the ability to handle complex traffic conditions in complex urban environments. Considering the importance of traffic simulation in autonomous driving research, this article also describes the latest progress in autonomous driving from three aspects: data collection, motion planning, and simulation testing.

1.3 Overall framework

• Section 2 introduces three model-based traffic simulation methods and provides different methods for process modeling and geometric representation of road networks.

• Section 3 introduces data-driven traffic simulation technology based on different data acquisition methods.

• Section 4 introduces methods for evaluating the realism of generated virtual traffic flows.

• Section 5 introduces the dataset, e2e motion planning algorithm and autonomous driving research using virtual traffic.

• Section 6 discusses the current situation and prospects.

overall framework

2 Model-based traffic flow modeling

In traffic simulation, an important task is to depict the motion of vehicles at different levels of detail. At present, traffic simulation technology is roughly divided into three types, namely macro, micro and meso.

Traffic flow can be viewed as a flow in which vehicles share similar goals and behavior rules, interacting with their neighbors while maintaining their respective driving characteristics. In computer graphics, group simulation has always been an important research area, which provides support for the study of collective behavior and dynamics. Crowd simulation can model the group as a whole at the expense of the real movements of individuals in a macroscopic way, or it can model the group as a collection of individual movements in a microscopic way.

2.1 Macroscopic approach

Macroscopic methods, also known as continuum methods, describe vehicle behavior and interactions at a lower level: traffic flow is represented by a continuum of speed, flow, density, etc. Macroscopic methods are primarily designed for efficiency and focus on reproducing aggregation behavior measured with collective quantities such as traffic density and traffic volume.

One of the early first-order macro models was the LWR model developed by Lighthill, Whitham, and Richards. Essentially, LWR models describe the movement of large-scale traffic flows in low-resolution detail. Its limitation is that it cannot simulate vehicle motion under non-equilibrium conditions.

Payn和Whitha提出了连续二阶交通流模型 Payne-Whitham (PW)模型 。一阶模型假设存在一个固定的平衡状态，二阶模型引入一个二阶微分方程来描述交通速度动态。但在PW模型中司机可以受到他们的后续车辆的影响。

Aw、Rascle和Zhang对PW模型进行了修正，以消除其非物理行为。Zhang同样提出了对PW模型动量方程的修正来处理向后传播的交通，构建了 Aw-Rascle-Zhang (ARZ)模型 。

为了生成详细的交通流三维动画，Sewall等提出了 连续交通仿真模型 以生成大规模道路网络上真实的交通流。他们将车道离散为多个单元格来进行仿真，并通过引入一种新的变道模型，使单车道的ARZ模型适应于处理多车道交通。为了更新每个单元的状态，LEVEQUE R. J使用有限体积法(FVM)进行空间离散化，并结合Riemann求解器对ARZ方程进行求解。为了对车道合并和变道行为进行建模，Sewall等人将连续动态与离散车辆信息相结合，将车辆表示为车辆粒子系统。这些粒子系统是由底层的连续流驱动的。

离散道路

综上，宏观模型是模拟大规模交通的有效工具。然而这些技术仅限于高速公路网络，但 因为城市交通包含了汽车之间丰富的交互行为，因而 不适合模拟城市交通。 此外，由于这些模型不模拟汇入等行为，因此无法处理换道过程中的密度传递。

2.2 微观方法

微观模型在高水平的细节上产生车辆运动，每辆车都被视为一个离散的代理，且满足一定的控制规则。 针对城市交通仿真已经开发了大量的微观模型，这是因为微观模型可以灵活地 建模代理的异构行为、不同的道路拓扑以及车辆之间的交互关系 。

早期的微观模型包括 元胞自动机模型和跟车模型 。元胞自动机模型中车辆的运动由预先指定的时间、空间和状态变量中的演化规则来描述。具体来说，道路被离散化为单元，模型决定车辆何时从当前单元移动到下一个单元。由于其简单性，元胞自动机模型计算效率高，可以模拟大型路网上的大量车辆。然而由于其离散性，因此只能再现有限数量的真实交通行为。

相比之下，车辆跟驰模型，可以生成真实的驾驶行为和详细的车辆特征，但其计算成本相对较高。模型假设交通流由分散的粒子组成，并对汽车间的相互作用进行了详细的建模。模型通过基于刺激-响应框架(Response = Sensitivity Stimulus)的连续时间微分方程来表示每辆车的位置和速度。

在过去的几十年里，跟驰模型发展很快。其中两个著名的例子是 最优速度模型(OVM) 和智能驾驶模型(IDM) 。OVM模型假设主车保持最优速度，它的加速度由它的速度和前车的最佳速度之差决定。IDM模型根据车辆当前速度和相对于前车的速度和位置计算车辆的加减速，特定参数使IDM模型能够模拟各种车辆类型和驾驶风格。

对于 多车道模拟 ，其中一个方法是使用改进的OVM，该模型用于模拟双车道高速公路和有入口匝道的单车道高速公路的交通。另一种方法是使用twolane交通模型，该模型用来模拟交通的横向效应。

Shen和Jin 提出了一种增强的结合连续换道技术的IDM，可以产生具有平滑的加减速策略和灵活换道行为的交通流。 该模型对原有的IDM模型进行了修正，使其更适合于城市路网。该模型将加速度过程分为自由道路加速项和减速项，加速项描述了驾驶员达到期望速度的意愿，减速项描述了驾驶员与附近车辆保持安全距离的意愿。另外，对于减速项增加了一个激活控制部分，使停车更加平稳。该模型将城市道路变道行为分为 自由变道和强制变道 两种情况，并为这两种情况提供了一个灵活的连续模型。自由变道行为出现在相对自由的道路条件下，由 双车道MOBIL模型 建模。强制换道则应用于主车因为一些必要的因素要求的换道行为,如到达车道终端或在十字路口转向,而主车及周边车辆之间的gap不支持自由换道。

LU等人扩展完整的 速度差异模型(FVDM) 以处理在农村交通仿真中的close-car-braking情况。并在交通仿真中引入了人格模型。

必须换道的情况

与单车道或多车道的交通仿真相比， 交叉口交通仿真 难度更大。Doniec等人提出了一种多智能体的交通仿真行为模型，将交叉口交通视为一个多智能体协调任务。具体来说，首先每辆车感知周围的交通情况并做出决策，其次提出了一种预测算法来模拟车辆的预测能力。

Wang等在交通仿真中引入了影子交通的概念，以统一的方式对交通异常进行建模。Chao等人设计了一个基于规则的流程来模拟混合交通仿真中车辆与行人的交互作用。

综上所述，微观交通模型的目的是描述特定的车辆行为，因此可以用来模拟连续车道和十字路口的交通情况。该模型的瓶颈通常是计算成本，尤其是在进行大规模仿真时。

2.3 混合方法

Sewall等人使用基于代理的模型来模拟感兴趣区域的交通，而其余区域使用连续体模型，提出了一种混合方法。 通过在两种建模方法之间动态和自动切换，进而可以根据用户偏好来模拟不同详细级别下的交通。

黄框内使用微观agent，其他区域使用宏观flow

2.4 路网生成

交通仿真是车辆与路网相互作用的一种形式。底层道路网络的获取和建模是重要且具有挑战性的。 真实世界道路网络的数字表示已经越来越有可用性，但这些数据往往不能直接用于模拟交通。对于道路网络的过程建模和几何表示，已经提出了许多方法：

CityEngine软件 ，采用基于L-system过程的方法来生成道路网络。它可以以地图图像为输入，生成一组公路和街道，将土地分割成地块，并在相应的地块上为建筑物构建合适的几何形状。在基于CityEngine的路网生成模型的基础上学者还陆续提出了具有更大的灵活性基于模板的路网生成模型，如 利用生成对抗网络(GAN)来合成的道路网络模型等 。 这些方法是为构建虚拟场景而设计的，但它们常常无法为交通仿真提供必要的信息，如车道到车道的连接和邻接 。

CityEnginge地图和WilKie地图

交通仿真软件MITSIM 使用node，link，segment和lane来描述道路网络的语义。在该模型中，segment表示具有相同几何线的lane集合，link表示segment集合，向量数据存储在segment的数据结构中，所存储的信息包括起始点/结束点和段弧的曲率，一个node用来描述一个交点，这里node必须作为输入数据提供给模型，并且仅用于描述link是否连接，不考虑交叉口各方向link之间的冲突关系。

有人 利用真实地理信息系统(real Geographic Information System, GIS)数据建立了一个包含拓扑交通信息、路面和街道对象的连贯的街道网络模型， 该系统可以提供车道和车道间的相互连接，以作为交通仿真所需的基本几何信息，然而，他们使用lane作为原子单位来定义和组织道路网络，而忽略了道路网络的矢量数据。值得一提的是，为了方便不同驾驶模拟器之间的数据交换，目前提出了一种开放数据格式 OpenDRIVE 来规范逻辑道路描述。

一种新的道路网络模型图能将低细节的GIS数据自动转换为高细节的功能道路网络进行仿真。 利用该模型可以生成区域中心拓扑结构和弧路表示。该模型以车道为基础定义交叉口，通过交通信号和预先确定的移动优先级，在模拟中对交叉路口进行管理，生成的道路网络库可以在http://gamma.cs.unc.edu/RoadLib/上找到。该模型激发了更多基于车道的模拟技术，如Mao等在Frenet框架下基于道路轴线的车道模型，以方便复杂的交通仿真。

（ 编者按： 为了方便不同驾驶模拟器之间的数据交换，针对自动驾驶也有专用的道路网络，目前也有专用的道路网络库，相对重要的是OpenDRIVE ， 相关文章很多，这里不再赘述，使用RoadRunner可以很高效地生成，它还可以配合OpenSCENARIO使用。还有一种开放数据格式是lanelet，在它们之间还有转换库，OpenDrive2Lanelet。）

一般情况下， 宏观交通仿真对路网的细节要求较少，主要是需要几何信息，以便对交通流密度和速度的传播进行建模 。相比之下， 微观交通仿真由于输出单个车辆的详细运动，通常需要更多关于道路网络的信息。 这些信息包括车道的分隔和连接、交通信号逻辑、在十字路口和坡道上移动优先级等。

3 数据驱动的交通流仿真

本节中主要探讨真实世界交通数据的获取方法和各种数据驱动的交通重建和合成方法。

3.1 真实数据收集

交通传感器有几种形式： 感应环探测器 通常被放置在高速公路和主要道路上，记录每辆经过的车辆的属性。 摄像机 作为另一种固定的传感器也可用于监控交通情况。 手机和GPS设备 作为移动传感器也被用来记录车辆的速度和位置。

除了单车数据外，许多研究还致力于收集联网车辆的交通数据。例如，2012年在美国密歇根州的安娜堡启动了安全试点模型部署SPMD计划。大约3000辆车辆装备了GPS天线和DSRC(专用短程通信)设备。每辆车都向附近的车辆和路边的单位广播基本的安全信息，包括它的位置和速度。 由于这种类型的数据可以在高频率下采样(10hz)，这可能会导致存储和通信系统的巨大成本 。

3.2 交通重建和合成

创建符合真实世界交通的数字表征被称为虚拟交通，它由Van Den Berg等人首先提出。 他们的工作利用交通传感器提供的时空数据重建和可视化一个连续的交通流。 如下图所示，传感器(点A、点B、点C)每隔200-400米放置在道路上。对于某一辆车,传感器提供了一个元组，包含ABC三点的数据, 每个数据点分别包含了车辆通过时间t，车道id，车速v。 任务是计算在给定车道上，在给定时间内，以给定速度启动和到达车辆的轨迹。 该方法首先离散可能的状态-时间-空间，并约束车辆的运动到预先计算的路线图。然后，在路线图中为每辆车寻找最优轨迹，使换道次数和加减速度最小化，并与其他车辆的距离最大化，以获得平滑、真实的运动。

对于多辆车，采用基于优先级的多机器人路径规划算法来计算车辆的轨迹。但基于优先级的多智能体路由规划算法耗时较长，使得该方法随着搜索空间离散化分辨率的提高而变得难以处理。

从道路中获取的时空数据进行交通重建

Wilkie等人将稀疏传感器测量的宏观状态估计与基于agent的交通仿真系统相结合，引入了一种实时技术来重建单个车辆的真实运动。 如下图所示，该方法具有一个交通状态估计阶段，在这个阶段，使用Kalman smoothers (EnKS)和一个连续交通仿真器来创建整个道路网络的速度和密度场的估计。然后利用状态估计来驱动一个基于agent的交通仿真模型，生成各个车辆的详细运动。最后输出与传感器测量的原始交通信号相一致的二维交通流。该方法具有更高的灵活性和更低的计算成本。然而这种估计方法除了个别车辆的匹配外，基本上还是一种宏观模型。

交通流重构算法流程

Li等提出了 一种利用GPS数据重建城市尺度交通的方法 。为了解决数据覆盖不足的问题，该方法以GIS地图和GPS数据为输入，采用双阶段过程重构城市尺度的交通。 在第一阶段，即交通重建过程中 ，利用统计学习与优化、地图匹配和行程时间估计技术相结合的方法，从稀疏的GPS数据中重建并逐步细化单个路段的交通条件。在 第二阶段，即动态数据补全过程中 ，引入了基于元模型的仿真优化，以有效地细化第一阶段的重建结果，同时引入了一个微观仿真器，在数据覆盖不足的区域动态补全缺失的数据。

上述交通重建技术主要用于 预测同一场景下具有稀疏输入数据的完整交通流， 而其他数据驱动的交通综合方法则旨在 从有限的交通轨迹样本中生成新的交通流。 Chao等人利用一组有限的车辆轨迹作为输入样本，通过 纹理合成（texture synthesis）和交通行为规则的融合 来合成新的车辆轨迹。如下图所示，输入车辆轨迹集包含多种考虑车道数和交通流密度的段，将交通流时空信息作为二维纹理，可以将新交通流的生成表示为纹理合成过程，通过最小化新开发的交通纹理能量度量来解决这一问题。

两车道车辆轨迹的纹理类比

另外一种方法是使用 机器学习算法来学习车辆的详细运动特征，包括纵向加减速和换道过程 。Chao等人提出了一种基于视频的方法，从交通动画视频中学习驾驶员的具体驾驶特性。该方法将每辆车独特驾驶习惯的估计问题转化为寻找微观驾驶模型的最优参数集的问题，并采用自适应遗传算法求解。Bi等人从车辆轨迹数据中学习变道特性，该方法首先从预先收集的车辆轨迹数据集中提取与换道任务最相关的特征，然后利用所提取的特征对换道决策过程进行建模，并对换道执行过程进行估计。

数据驱动换道模型的pipline示意图。预处理步骤从预采集的交通数据集中提取最相关的特征，然后决策模块推断出目标车辆是否需要变道，以及需要变到哪个目标车道/间隙。最后，执行模块计算所涉及车辆的详细轨迹，以完成变道任务

上述工作的重点是模拟高速公路或大型城市网络上的车辆。最近，Bi等提出了 一种基于深度学习的交叉口交通仿真 框 架 。为了描述车-环境相互作用的视觉感知效应，他们建立了一个称为网格地图的网格坐标系统，编码异源之间的车辆与行人混合的相互作用。如下图所示，在网格地图上滑动五个通道的窗口可以为每辆车生成一个环境矩阵。环境矩阵可以捕捉车辆和行人在车辆周围的速度和位置。除了环境矩阵外，基于所收集的交叉口交通数据集的车辆标识还被用来描述当前车辆状态，然后利用卷积神经网络和递归神经网络对交叉口处的车辆轨迹模式进行学习。除了模拟路口交通，它还可以通过提供车辆新的目的地和驾驶环境来改变现有的路口交通动画。

4 验证和评估

一般来说有两种交通真实度评估方法: 可视化方法和统计方法 。在可视化验证中，将真实交通和模拟交通的图形表示并排显示，以确定它们是否可以区分。但主观的用户研究会耗费大量时间且容易出错，使用定量和客观的度量进行统计验证不仅可以用来测量各种模拟交通流的真实性，还可以用以一致的方式客观地比较不同交通仿真模型的性能。 在交通仿真中，由于交通的随机性，通常不进行直接的轨迹比较，而是比较平均速度和流量随时间的变化等，此外，更详细的比如特定的运动参数，包括速度、加速度和车辆间隙也被用来验证交通仿真技术的有效性 。

5 在自动驾驶中的应用

本章介绍自动驾驶训练数据收集，基于深度学习的运动规划方法和自动驾驶仿真

5.1 自动驾驶数据集

自动驾驶数据集也可以用于交通仿真和动画。 首先，车辆轨迹可以用来校准交通仿真模型。其次，大规模的交通数据集丰富了数据驱动的交通综合方法。最后，各种真实的交通数据集也能对虚拟交通评估提供帮助 。

常用的数据集包括comma.ai数据集、Berkeley DeepDrive视频数据集(BDDV)、LiDAR-Video数据集(LiVi-Set) 、本田研究所(Honda Research Institute)的驾驶数据集(HDD)、 Drive360。

其他一些没有驾驶行为的数据集也有助于自动驾驶的视觉语义理解和基于视觉的控制。包括KITTI数据集、Cityscape数据集、牛津RobotCar数据、Udacity 数据集等。

5 .2 运动规划与决策

运动规划和决策对于智能体在其环境中导航至关重要 。（ 编者按：构建更真实的agent模型 ）这一节将回顾了几种基于学习的自动驾驶车辆和其他智能体的运动规划方法和决策算法。

Lenz et al. studied vehicle motion at highway entrances. Partially Observable Markov Decision Processes (POMDPs) are used to predict vehicle motion. Kuefler et al. used generative adversarial imitation learning (GAIL) to learn driving behavior. This method overcomes the problem of cascading errors and can produce real driving behavior. Hecker et al. integrated information from surrounding 360-degree cameras into a route planner. The network used in this method maps sensor output directly to low-level driving actions , including steering angle and speed. Kim et al. introduced an end-to-end, explainable self-driving approach that incorporates an introspection-based explanation model. The model consists of two parts: the first part is a CNN-based visual attention mechanism that maps images to driving behaviors; the second part is an attention-based video-text model for textual explanations of model actions. Yang et al. utilized virtual traffic data collected in CARLA and TORCS to predict vehicle behavior, namely DU-drive , as shown in the figure below.

In recent years, reinforcement learning has also been applied to autonomous driving . Abbeel et al. proposed an efficient algorithm to reconcile the trade-off between global navigation and local planning for generating vehicle trajectories. Silver et al. proposed a coupled cost function suitable for autonomous navigation systems to balance different preferences, including where the vehicle should be and how it should be driven. Lillicrap et al. use deep q-learning to implement a model-free system that learns strategies to guide the vehicle to stay on the track in a simulated driving environment. Kuderer et al. proposed a feature-based inverse reinforcement learning (IRL) method to learn individual driving styles for autonomous driving. Wolf et al. proposed a Deep Q-Networks (DQN) for guiding vehicles in three-dimensional physics simulations. Pan et al. used a novel reality translation network (VISRI) to train an autonomous driving model in a virtual environment and applied it to the real environment. Liang et al. proposed a general controlled imitation reinforcement learning (CIRL) method to alleviate the low exploration efficiency of large continuous action spaces, based on visual information input directly from the CARLA simulator.

In order to drive vehicles efficiently and safely in complex traffic environments, self-driving cars need to predict the movements of surrounding vehicles. The interaction between vehicles and pedestrians needs to be accurately represented. The task of trajectory prediction can be divided into several categories: physics-based, behavioral policy-based and interaction-based models. Furthermore, a large amount of work based on deep learning has been used for trajectory prediction. (Editor's note: Traffic modeling based on trajectory prediction method)

Lee et al. proposed a deep stochastic IOC RNN codec framework DESIRE to predict the future distance of agents in dynamic scenes, thereby generating accurate vehicle trajectories. Kim et al. proposed a probabilistic vehicle trajectory prediction method based on LSTM , which uses an occupancy grid map to characterize the driving environment. Deo and Trivedi adopted a convolutional social pooling network to predict vehicle trajectories on highways. The entire network includes an LSTM encoder, a convolutional social pooling layer and an operation-based decoder. Specifically, it first uses an LSTM encoder to learn vehicle dynamics based on tracking history, then, uses a convolutional social pooling layer to capture the interdependencies of all vehicle trajectories, and finally trains a maneuver-based LSTM decoder to predict future vehicle trajectories. Distribution

5.3 Autonomous driving simulation

The amount of data in the real world is not enough to cover many complex traffic scenarios, which restricts the self-driving system from learning different driving strategies. More importantly, self-driving cars always adopt the most conservative and least efficient strategies for safety reasons. decision making. As an effective alternative tool, the development of high-fidelity driving simulators can provide various types of traffic conditions for training autonomous vehicles.

In fact, simulation has been used to train driving models since the early days of autonomous driving research. Racing simulators are used to evaluate various driving methods. For example, Chen et al. used TORCS to evaluate perceptual models. Some researchers have also used GTAV to derive autonomous driving strategies and achieved performance comparable to control strategies generated with manually annotated real-world images.

CARLA is an open source simulator used to support the development, training and verification of urban autonomous driving models. The simulation platform enables flexible setup of sensor suites and provides data that can be used to train driving strategies. This data includes GPS coordinates, speed, acceleration/deceleration and collisions, etc. A wide range of environmental factors can be specified, including weather and time of day. With these settings, CARLA has been used to study the performance of many autonomous driving methods, including classic modular methods, end-to-end training models through imitation learning, and end-to-end training models through reinforcement learning.

Best et al. also proposed a high-fidelity simulation platform autonova-sim for autonomous driving data generation and driving strategy testing . autonova-sim is a set of advanced extensible modules. Similar to CARLA, it also supports specific configurations of vehicle sensor systems, changes in time and weather conditions, and non-vehicle participants such as cyclists and pedestrians.

In addition, several recent projects have attempted to establish simulation platforms to train end-to-end driving systems and provide rich virtual traffic scenarios for autonomous driving testing. For example, Apollo integrates large amounts of driving data from both real and virtual traffic. One limitation is that virtual traffic data is manually created with specific, well-defined obstacles and traffic signals, which does not provide the same level of realism and complexity as real traffic conditions .

Recently, Li et al. developed a simulation framework AADS that can enhance real images with simulated traffic flow to generate realistic images. Using data from lidar and cameras, the framework can synthesize simulated traffic flow into the background based on the actual trajectories of vehicles. Composite images can be modified to different viewpoints and fully annotated, and can be used in the development and testing of autonomous driving systems. This framework aims to overcome the high cost of manually developing virtual environments and the possible degradation of vehicle performance when using virtual images to train autonomous driving.

Another framework ADAPS developed by Li et al. obtains relevant autonomous driving data from accidents. The framework consists of two simulation platforms. The first simulation platform runs in 3D and is used to test learned strategies and simulate accidents; the second simulation platform runs in 2D and is used to analyze accidents that occurred in the first simulation platform and provide alternative safety trajectories by to resolve the accident. A large amount of annotated data is then generated based on the safety trajectories for training and updating control strategies. Compared with previous techniques such as DAGGER, ADAPS also represents a more efficient online learning mechanism that can greatly reduce the number of iterations required to generate robust control strategies.

6 Discussion

First, traffic simulation models should be able to model as many complex traffic behaviors as possible while maintaining computational efficiency. However, for existing microscopic traffic models, each vehicle behavior, such as acceleration/deceleration and lane changing, is modeled and controlled individually. In addition, microscopic traffic models focus more on vehicle motion in the forward direction, and lane changing behavior and vehicle lateral motion are generally ignored . In addition, according to the vehicle following law, the movement of the vehicle is mainly affected by the vehicle in front, so the simulation results obtained seldom involve the calculation of acceleration and deceleration of other vehicles in the field of view. To simulate more realistic traffic flows, a unified and scalable simulation framework needs to be developed for rich vehicle behaviors, including acceleration/deceleration, staying in lanes, lane changes, and interaction with non-vehicle traffic participants such as pedestrians and cyclist) interaction .

Third, for the evaluation of virtual traffic fidelity, dictionary-based metrics provide a feasible solution. However, the quality and composition of traffic data have a direct and substantial impact on the generated dictionary and thus the evaluation results. In addition, the framework extracts the acceleration, speed, relative speed and gap distance of each vehicle to the preceding vehicle to describe the instantaneous state of the vehicle. In order to better capture the traffic pattern for dictionary learning, the traffic flow should also be considered and extracted. More features, including vehicle kinematic constraints, road constraints and driver characteristics. For macroscopic traffic simulations, it is necessary to develop fidelity metrics that measure traffic flow in an aggregated manner, including flow density and speed .

Finally, for autonomous driving, addressing the interaction between autonomous vehicles and other road users remains a challenge. Existing simulators rarely consider the interaction between both parties . For example, two types of non-vehicle traffic participants are implemented in the Apollo simulation platform: pedestrians and cyclists. However, the behavior of these non-vehicle agents is predefined, so they cannot react to vehicles in real time. Although dynamic pedestrians are introduced in CARLA, the interaction between vehicles and pedestrians is handled in a simple, pre-specified way: pedestrians check to see if there are nearby vehicles before moving, and then continue without further checking. Mobile, this interaction mode is still too simple and needs further study .

Recommended event: SAE 2020 Automotive Electronics and Software Technology Forum (OEM free)