面向自动驾驶的仿真场景自动生成方法综述

The traditional scenario enumeration method based on expert experience has failed to meet testing requirements owing to the increasing reliance of autonomous driving on virtual simulation scenarios for testing and verification. The automatic generation of simulation scenarios has substantial technical advantages in terms of scenario diversity, safety, interpretability, and generation efficiency. It plays a crucial role in improving the efficiency of autonomous driving tests, which have become a prevalent research topic. In recent years, researchers have intensively studied automatic scenario generation methods. In the present study, extensive research was conducted on the results obtained in the field of automatic scenario generation. Thus, the latest research progress in scenario definition, scenario deconstruction, scenario generation based on mechanism modeling, scenario generation driven by data, etc., is schematically presented in this paper. In addition, an analysis on some areas worthy of further study was performed, and prospective research directions are presented herein. In terms of scenario deconstruction, given that scenarios are abundant, extremely complex, and inexhaustible, substantial importance should be given to research on the deconstruction of heterogeneous complex scenarios with the coupling of "field-weather-traffic." Regarding mechanism modeling, to meet the requirements of testing scenario diversity and boundary generation, the focus should be on scenario combination generation, edge scenario optimization generation, and adaptive generation. Furthermore, data with rich content must be collected, laying the foundation for research. To fully exploit the test value of scenario data, attention should be paid to the research on scenario reconstruction, thereby accelerating the generation of test scenario databases and dangerous scenarios. Thus, future research should focus on the aspects mentioned above to establish a completely automatic simulation scenario generation system for autonomous driving. This will lay a theoretical foundation for performing large-scale simulation tests of high-level autonomous driving.