Page 246 - 《软件学报》2026年第4期

田浩翔 等: 基于路网建模的自动驾驶关键场景生成与自适应演化方法 1687

个安全违背. LEADE 的安全违背暴露频率更高, 这表明 LEADE 能够更高效地生成更多具有安全风险的场景. 在
具体场景生成的时间成本上, LEADE 每生成并运行一个具体场景需要 23 s, BehAVExplor 需要 41 s, 而 MOSAT
为 24 s, AV-Fuzzer 为 55 s. LEADE 生成并执行具体场景的时间成本几乎与 MOSAT 相同, 低于 AV-Fuzzer 和
BehAVExplor. 对比结果表明, LEADE 能够更有效地生成更多样化的安全违背场景.
4 总 结

本文提出了一种基于路网建模的自动驾驶系统关键场景生成与自适应演化方法——LEADE. 通过解析用户
测试需求, 构建抽象场景; 基于路网建模, 生成具体场景. 运用改进的自适应遗传算法, 包括细粒度评估、自适应选
择、自适应变异, LEADE 能够有效地生成多样化的安全关键场景, 从而克服现有方法在生成效率和场景多样性上
的局限. 实验结果表明, LEADE 在工业级全栈自动驾驶系统平台上的应用表现优异, 不仅高效地发现了自动驾驶
系统中的多种安全违背, 还揭示了当前技术未能覆盖的新类型的安全关键场景. 在生成和执行具体场景时, 场景中
的从车和行人必须严格遵守交通法规. 虽然这一约束可能导致 LEADE 遗漏一些由于参与者异常行为导致的安全
违背场景, 但考虑到这些异常行为在现实交通中属于极端情况, 缺少实际意义; 此外, 目前自动驾驶系统存在的安
全问题较多, 因此当前测试的首要目标是发现由自动驾驶车辆引起的安全违背, 尚未进入测试自动驾驶系统面对
极端情况的应急能力的评估阶段. 故而保证测试场景中从车和行人的行为轨迹遵守交通法规, 能够确保发现的安
全违背问题, 是由自动驾驶系统导致的. 未来, LEADE 可以进一步扩展以适应更复杂的场景要求, 例如动态交通环
境和多模态传感器输入, 从而更全面地支持自动驾驶系统的安全评估与优化.

References
[1] Bertoncello M, Wee D. Ten ways autonomous driving could redefine the automotive world. 2015. https://www.mckinsey.com/industries/
automotive-and-assembly/our-insights/ten-ways-autonomous-driving-could-redefine-the-automotive-world#/
[2] Barr ET, Harman M, McMinn P, Shahbaz M, Yoo S. The oracle problem in software testing: A survey. IEEE Trans. on Software
Engineering, 2015, 41(5): 507–525. [doi: 10.1109/TSE.2014.2372785]
[3] Zofka MR, Kuhnt F, Kohlhaas R, Rist C, Schamm T, Zöllner JM. Data-driven simulation and parametrization of traffic scenarios for the
development of advanced driver assistance systems. In: Proc. of the 18th Int’l Conf. on Information Fusion (FUSION). Washington:
IEEE, 2015. 1422–1428.
[4] Singh S, Saini BS. Autonomous cars: Recent developments, challenges, and possible solutions. IOP Conf. Series: Materials Science and
Engineering, 2021, 1022(1): 012028. [doi: 10.1088/1757-899X/1022/1/012028]
[5] Chen L, Wu PH, Chitta K, Jaeger B, Geiger A, Li HY. End-to-end autonomous driving: Challenges and frontiers. arXiv:2306.16927,
2024.
[6] Pre-crash scenario typology for crash avoidance research. 2022. https://www.nhtsa.gov/sites/nhtsa.gov/files/pre-crash_scenario_typology-
final_pdf_version_5-2-07.pdf
[7] Feng S, Yan XT, Sun HW, Feng YH, Liu HX. Intelligent driving intelligence test for autonomous vehicles with naturalistic and
adversarial environment. Nature Communications, 2021, 12: 748. [doi: 10.1038/s41467-021-21007-8]
[8] Gambi A, Huynh T, Fraser G. Generating effective test cases for self-driving cars from police reports. In: Proc. of the 27th ACM Joint
Meeting on European Software Engineering Conf. and Symp. on the Foundations of Software Engineering. Tallinn: ACM, 2019.
257–267. [doi: 10.1145/3338906.3338942]
[9] Bashetty SK, Amor HB, Fainekos G. DeepCrashtest: Turning dashcam videos into virtual crash tests for automated driving systems. In:
Proc. of the 2020 IEEE Int’l Conf. on Robotics and Automation (ICRA). Paris: IEEE, 2020. 11353–11360. [doi: 10.1109/ICRA40945.
2020.9197053]
[10] Althoff M, Lutz S. Automatic generation of safety-critical test scenarios for collision avoidance of road vehicles. In: Proc. of the 2018
IEEE Intelligent Vehicles Symp. (IV). Changshu: IEEE, 2018. 1326–1333. [doi: 10.1109/IVS.2018.8500374]
[11] Zhang XD, Cai Y. Building critical testing scenarios for autonomous driving from real accidents. In: Proc. of the 32nd ACM SIGSOFT
Int’l Symp. on Software Testing and Analysis. Seattle: ACM, 2023. 462–474. [doi: 10.1145/3597926.3598070]
[12] Harman M. Automated test data generation using search based software engineering. In: Proc. of the 2nd Int’l Workshop on Automation
of Software Test (AST 2007). Minneapolis: IEEE, 2007. 2. [doi: 10.1109/AST.2007.4]