|
Abstract:
|
To test for the safety of an automated driving system (ADS) is one of the most important tasks for ADS evaluation and certification. However, the performance of the dynamic driving task by an automated driving system (ADS) involves the integration of multiple subsystems and the domains in which ADSs will operate are complex. The high-frequency, high dimension driving data constitutes the testable space but is challenge to be quantitatively evaluated. In this study we introduced a novel deep-learning-based approach to elite the testable space through feature engineering and clustering. The variational autoencoder was used to conduct feature engineering from thousands of the real-life line change events from the Second Strategic Highway Research Plan Naturalistic Driving Study. A subsequent clustering analysis identified dozens of distinct groups of lane change cases. The results provide crucial information for guide testable case generation and ADS testing.
|
