Automatic Scenario Generation for Decision Algorithm Performance Evaluation of Autonomous Vehicle via Scenario Parameter Sweeping Method

摘要

Autonomous driving algorithms are always in operation during normal driving and require responses to many more driving situations than ADAS, which assists the driver only in specific risk scenarios. However, due to the absence of a quantitative algorithm evaluation method, most algorithms are evaluated in their own test scenarios. Even in similar cases in identical scenarios, different decisions may be made depending on small differences such as the speed or location of surrounding vehicles. Therefore, one representative scenario cannot cover all cases and it is difficult to quantitatively compare and evaluate algorithms. In this study, the parameters constituting the scenario are determined for lane change and intersection scenarios, typical scenarios in autonomous driving research. Then, all key parameters are swept within a certain range, and all cases that are slightly different are generated even in identical scenario. Simulations are performed automatically on the generated cases, and the vehicle risk is calculated on all cases based on the time to occupancy (TTO). Using this evaluation method, edge cases in which the autonomous algorithm may have weaknesses can be found.