Identifying and Explaining Safety-critical Scenarios for Autonomous Vehicles via Key Features

• DOI: 10.1145/3640335
• Publication Date: 2024-01-11T12:26:34Z
• Authors: Neelofar Neelofar; Aldeida Aleti
• Keywords: SAFER; Test suite; Scenario testing; Test strategy; Critical area

Ensuring the safety of autonomous vehicles (AVs) is utmost importance, and testing them in simulated environments a safer option than conducting in-field operational tests. However, generating an exhaustive test suite to identify critical scenarios computationally expensive, as representation each complex contains various dynamic static features, such AV under test, road participants (vehicles, pedestrians, obstacles), environmental factors (weather light), road’s structural features (lanes, turns, speed, etc.). In this article, we present systematic technique that uses Instance Space Analysis (ISA) significant affect their ability reveal unsafe behaviour AVs. ISA identifies best differentiate safety-critical from normal driving visualises impact these on scenario outcomes (safe/unsafe) two dimensions. This visualisation helps untested regions instance space provides indicator quality terms percentage feature covered by testing. To predictive identified train five Machine Learning classifiers classify safe or unsafe. The high precision, recall, F1 scores indicate our proposed approach effective predicting outcome without executing it can be used for generation, selection, prioritisation.