TrafﬁcGen Learning to Generate Diverse and Realistic Trafﬁc Scenarios Lan Fengx Quanyi Liy Zhenghao Peng Shuhan Tanz Bolei Zhou xETH ZurichyThe University of Edinburgh

2025-04-24 0 0 3.75MB 10 页 10玖币

侵权投诉

TrafﬁcGen: Learning to Generate Diverse and Realistic Trafﬁc Scenarios

Lan Feng§*, Quanyi Li†*, Zhenghao Peng♠*, Shuhan Tan‡, Bolei Zhou♠

§ETH Zurich, †The University of Edinburgh,

♠University of California, Los Angeles, ‡The University of Texas at Austin

Abstract— Diverse and realistic trafﬁc scenarios are crucial

for evaluating the AI safety of autonomous driving systems in

simulation. This work introduces a data-driven method called

TrafﬁcGen for trafﬁc scenario generation. It learns from the

fragmented human driving data collected in the real world and

then generates realistic trafﬁc scenarios. TrafﬁcGen is an au-

toregressive neural generative model with an encoder-decoder

architecture. In each autoregressive iteration, it ﬁrst encodes the

current trafﬁc context with the attention mechanism and then

decodes a vehicle’s initial state followed by generating its long

trajectory. We evaluate the trained model in terms of vehicle

placement and trajectories, and the experimental result shows

our method has substantial improvements over baselines for

generating trafﬁc scenarios. After training, TrafﬁcGen can also

augment existing trafﬁc scenarios, by adding new vehicles and

extending the fragmented trajectories. We further demonstrate

that importing the generated scenarios into a simulator as an

interactive training environment improves the performance and

safety of a driving agent learned from reinforcement learning.

Model and data are available at https://metadriverse.

github.io/trafficgen.

I. INTRODUCTION

Autonomous driving (AD) is transforming our daily life

with promised beneﬁts like safe transportation and efﬁcient

mobility. One of the biggest hurdles for deploying AD

in the real world is to ensure the vehicles controlled by

algorithms operate safely and reliably in all kinds of trafﬁc

scenarios. Before the real-world deployment of AD, the

simulation environment becomes an ideal testbed to evaluate

the reliability and safety of AD systems. However, most of

the existing simulators like CARLA [5] and SMARTS [32]

have hand-crafted trafﬁc-generating rules and maps, while

the trafﬁc scenarios available for testing are also far from

enough to emulate the complexity of the real world. As a

result, it is difﬁcult to evaluate how the AD systems make

safe-critical decisions and react to other trafﬁc participants in

complex trafﬁc scenarios. Thus, creating diverse and realistic

trafﬁc scenarios in simulation becomes crucial for thoroughly

evaluating the AI safety of AD systems.

Existing methods tackle the challenge by replaying vehicle

trajectories collected from the real world [15, 16, 30].

Though the trajectories replayed from real data preserve

the ﬁdelity of the real world, there are two issues. First, it

requires a time-consuming data collection process, particu-

larly on a large scale. Second, the trajectories of vehicles

in the commonly used driving datasets such as Waymo

Open Dataset [20] and Argoverse [2] are fragmented and

incomplete. Most of the trajectories only span a short period

Lan Feng, Quanyi Li and Zhenghao Peng contribute equally to this work.

of time due to the fact that they are collected by a data

collection vehicle in moving, which is often occluded by

other trafﬁc participants. For instance, only 30% of the

trajectories collected in Waymo Motion Dataset [20] last

more than 10 seconds, and only 12% cover the whole

scenario. Thus, the replayed trafﬁc scenarios are incomplete

and insufﬁcient for a thorough evaluation of AD systems.

One solution for covering more trafﬁc scenes is to design

test cases manually [10] or use heuristic methods like proce-

dural generation (PG) [8, 16] to create a huge number of test

cases. However, the generated scenarios cannot well reﬂect

the complexity of real-world trafﬁc and road structure. In

addition, it requires a substantial amount of human effort

and domain knowledge to design rules for placing vehicles,

determining their initial states, and setting trigger conditions

for assuring the interaction between the ego vehicle and other

trafﬁc participants [7].

Our goal is to enable the automatic generation of realistic,

complete, and diverse trafﬁc scenarios that learn from real-

world data. We develop a data-driven trafﬁc scenario gen-

erator TrafﬁcGen that can synthesize both the initial states

of trafﬁc vehicles and their long and complete trajectories.

TrafﬁcGen learns from the fragmented and noisy trajectories

collected in the real-world driving dataset such as Waymo

Open Dataset [20]. TrafﬁcGen follows an encoder-decoder

neural architecture. The encoder transforms the HD map and

states of vehicles into a scenario representation. The decoder

generates initial state distributions for placing vehicles and

long-term multi-modal probabilistic trajectories for realistic

simulation.

TrafﬁcGen can generate diverse and realistic trafﬁc scenar-

ios given HD maps after training, which greatly enlarges the

set of trafﬁc scenarios available for AD testing. As shown

in Fig.1, TrafﬁcGen can also be used to edit and augment

the existing scenarios, by (1) generating new trafﬁc on the

same HD map. (2) adding new vehicles and trajectories,

(3) inpainting a trajectory segment into a longer one. The

generated scenarios are further imported into the simulator

to improve the driving agent trained from reinforcement

learning (RL). The experimental results show that the safety

of driving agents can be substantially improved when being

trained on the generated scenarios with higher complexity

and trafﬁc density.

II. RELATED WORK

The majority of the trafﬁc scenarios in the existing driving

simulators are either pre-recorded in the real world [1, 15]

arXiv:2210.06609v2 [cs.RO] 5 Mar 2023

Fig. 1: Different applications of TrafﬁcGen. The colored vehicles and blue trajectories are generated by TrafﬁcGen. The

black vehicles and yellow trajectories are from the real data replay.

or synthesized by hand-crafted rules [4, 6, 12, 21], which

lack sufﬁcient diversity and realism.

To generate realistic driving scenarios, researchers attempt

to apply data-driven methods to learn from a large-scale

real-world dataset. Liang et al. [19] learn a probabilistic

distribution over trafﬁc ﬂows from where new scenes can

be sampled. Tan et al. [27] propose a neural autoregressive

model SceneGen, which inserts actors (vehicles) of various

classes into the scene and synthesizes their sizes, orienta-

tions, and velocities. These methods model high-dimensional

features of scenarios, improving the ﬁdelity and scalability.

However, they either require predeﬁned heuristics and priors,

e.g., scene grammars that encapsulate assumptions about

the trafﬁc rules [19] or rely heavily on post-processing

to guarantee the validity of the generated scene [27]. For

example, SceneGen needs to ﬁlter out improperly placed

vehicles beyond road boundaries and can only generate trafﬁc

snapshots from an empty map. Compare to SceneGen, our

method does not need post-processing and can augment

existing trafﬁc scenarios. Moreover, all these methods can

only generate snapshots of scenarios, which are static and

can not be used for interactive simulation, while our approach

allows dynamic trafﬁc ﬂows via trajectory generation.

Apart from generating static trafﬁc snapshots, another

line of works focuses on simulating realistic driving behav-

iors [11, 22, 25] and trajectories, such as TrafﬁcSim [26]

and SimNet [1]. TrafﬁcSim only generates trajectories on

real-world trafﬁc scenarios. SimNet ﬁrst generates trafﬁc

snapshots by generating bird-eye-view images of the map

and then generates vehicle trajectories. Their method lacks

the ability to edit the scenario. Differently, the proposed

TrafﬁcGen can edit the trafﬁc scenario and generate the long

trajectories of vehicles simultaneously.

Thus compared to the previous works, our method is an

integrated framework for generating both trafﬁc snapshots

and long vehicle trajectories, making it possible to produce

countless new trafﬁc scenarios as well as augment existing

ones.

III. METHOD

Problem Formulation. A trafﬁc scenario is denoted as τ=

(m,s1:T), which lasts Ttime steps and contains the High-

Deﬁnition (HD) road map mand the state series of trafﬁc

vehicles s1:T= [s1, ..., sT]. Each element st={s1

t, ...sN

is a set of states of Ntrafﬁc vehicles at time step t. Given

an existing scenario τ= (m,s1:T), the goal of TrafﬁcGen

is to learn to generate new trafﬁc scenarios τ0= (m,s0

1:T0)

that have similar distribution with τand different states s0

and longer time steps T0. After training, TrafﬁcGen takes

mas input and generates τ0as a totally different scenario.

Furthermore, with additional trafﬁc fragment [s1, ...st]as

input, it can augment or inpaint the existing scenario. The

neural architecture of TrafﬁcGen is illustrated in Fig. 2,

where a novel encoder-decoder architecture is designed to

generate trafﬁc scenarios. We introduce each component in

detail as follows.

A. Encoder for Trafﬁc Context

Vector-based Trafﬁc Representation.

We develop a new representation to encode the map and

the information about all vehicles on the map. In driv-

ing datasets like Waymo Open dataset [20] and Argoverse

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TrafcGen:LearningtoGenerateDiverseandRealisticTrafcScenariosLanFengx*,QuanyiLiy*,ZhenghaoPeng*,ShuhanTanz,BoleiZhouxETHZurich,yTheUniversityofEdinburgh,UniversityofCalifornia,LosAngeles,zTheUniversityofTexasatAustinAbstractDiverseandrealistictrafcscenariosarecrucialforevaluatingtheAIsafetyofa...

展开>> 收起<<

TrafﬁcGen Learning to Generate Diverse and Realistic Trafﬁc Scenarios Lan Fengx Quanyi Liy Zhenghao Peng Shuhan Tanz Bolei Zhou xETH ZurichyThe University of Edinburgh.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

TrafﬁcGen Learning to Generate Diverse and Realistic Trafﬁc Scenarios Lan Fengx Quanyi Liy Zhenghao Peng Shuhan Tanz Bolei Zhou xETH ZurichyThe University of Edinburgh

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: