Critical Scenario Identification for Testing of Autonomous Driving Systems
Abstract: Background: Autonomous systems have received considerable attention from academia and are adopted by various industrial domains, such as automotive, avionics, etc. As many of them are considered safety-critical, testing is indispensable to verify their reliability and safety. However, there is no common standard for testing autonomous systems efficiently and effectively. Thus new approaches for testing such systems must be developed.Aim: The objective of this thesis is two-fold. First, we want to present an overview of software testing of autonomous systems, i.e., relevant concepts, challenges, and techniques available in academic research and industry practice. Second, we aim to establish a new approach for testing autonomous driving systems and demonstrate its effectiveness by using real autonomous driving systems from industry.Research Methodology: We conducted the research in three steps using the design science paradigm. First, we explored the existing literature and industry practices to understand the state of the art for testing of autonomous systems. Second, we focused on a particular sub-domain - autonomous driving - and proposed a systematic approach for critical test scenario identification. Lastly, we validated our approach and employed it for testing real autonomous driving systems by collaborating with Volvo Cars.Results: We present the results as four papers in this thesis. First, we conceptualized a definition of autonomous systems and classified challenges and approaches, techniques, and practices for testing autonomous systems in general. Second, we designed a systematic approach for critical test scenario identification. We employed the approach for testing two real autonomous driving systems from the industry and have effectively identified critical test scenarios. Lastly, we established a model for predicting the distribution of vehicle-pedestrian interactions for realistic test scenario generation for autonomous driving systems. Conclusion: Critical scenario identification is a favorable approach to generate test scenarios and facilitate the testing of autonomous driving systems in an efficient way. Future improvement of the approach includes (1) evaluating the effectiveness of the generated critical scenarios for testing; (2) extending the sub-components in this approach; (3) combining different testing approaches, and (4) exploring the application of the approach to test different autonomous systems.
CLICK HERE TO DOWNLOAD THE WHOLE DISSERTATION. (in PDF format)