Year: 2026
Type: article
Abstract: Corner cases, characterized by a pronounced long-tail distribution, pose severe challenges to the safe operation of autonomous vehicles. However, the definition of corner cases remains unclear and the classification system requires further exploration, which has impeded progress in corner case generation and application. Moreover, natural driving databases, as critical carriers for understanding and revealing the essence of corner cases, are of great value for achieving high-confidence generation of corner cases in a data-driven manner. This paper reviews the existing definitions and classifications of corner cases. Based on a multi-dimensional analysis of their characteristics, it further summarizes frontier issues and recent research progress in data-driven scenario generation methods across various autonomous driving subsystems. Firstly, we systematically review the existing definitions and classification methods of corner cases. Then, focusing on corner case generation for autonomous driving perception system testing, we summarize the sources of perception data and data-driven corner case generation methods. Next, focusing on corner case generation for autonomous driving decision-making system testing, we review the major natural driving datasets and summarize corner case generation methods based on data-driven architectures. Finally, we discuss future trends in corner case generation within the data-driven framework and offer guidance to researchers.
Institutions: Jilin University, Chang'an University
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7214243136