基于树结构Parzen估计器的自动驾驶仿真测试关键场景生成方法

打开文本图片集

主题词：自动驾驶场景生成TPEMOTPECARLA中图分类号：U467.3；TP391.9 文献标志码：A DOI： 10.19620/j.cnki.1000-3703.20240092

Automatic Generation Method of Autonomous Driving Simulation Test ScenariosBasedon Tree-StructuredParzen Estimator

QinQin'，Yang Zhisheng1，LiDaoxin1，Shen Zhiwei²，Cao Xiaolin³ （1.Shanghai Polytechnic University，Schoolof Intellgent Manufacturingand ControlEngineering，Shanghai 201209; 2.School of Electrical Engineering and Telecommunications，Universityof New South Wales，Sydney 2052; 3.College of Automotive Engineering，Jilin University，Changchun 130015）

【Abstract】In viewof theexponential increase in thenumberof keyscene scenariosgenerated in high-dimensional space，andthediicultyoftraditionalartificialconstructionorrandomsearchmethods tobalancecoverageandeficiencythis paper proposes a search method basedon single-objective Tree structure Parzen Estimator（TPE）and Multi-Objective Tree structureParzen Estimator （MOTPE）.Asoftware-in-the-loopautomatedsimulationtesting framework isbuiltbyusing the CARLA simulator.Taking weather elementsasanexample，thecriticalscenario generationefectsofthediffrent search algorithmsarecompared.Theexperimentalresultsindicate thattheTPE-basedsearch methodandthe MOTPE-based method increase thenumberofkeyscenariosgeneratedby3.11timesand2.O6times，respectively，comparedtotherandomsearch method.TheMOTPEmethodis1.53times beter than TPEintermsofscenarioquality.When combinedwithscenario automaedgenerationandtesting frameworks，thesemethodsefectivelyadress theissueof explodingscenario numbers， allowing for the discovery of scenarios with high testing value.

Key words：Autonomous driving，Scenario generation，Tree-structured Parzen Estimator （TPE），Multi-Objective Tree-structured Parzen Estimator （MOTPE），CARLA

【引用格式】秦琴，杨志胜，李道鑫，等.基于树结构Parzen估计器的自动驾驶仿真测试关键场景生成方法[J].汽车技术，2025（5）：39-46.QINQ，YANGZS，LIDX，etal.Automatic Generation Methodof AutonomousDriving SimulationTestScenarios BasedonTree-Structured ParzenEstimator[J].Automobile Technology，2025（5）： 39-46.

1前言

随着车辆系统复杂性的提升，交通环境变化、驾驶任务多样性等因素成为自动驾驶车辆测试评价面临的全新挑战。（剩余11609字）