Uncertainty Quantification for Safety of the Intended Functionality of Autonomous Driving: A Comprehensive Survey

With the development of deep learning technology in autonomous driving, to solve the Safety problems caused by uncertain factors, the Safety of the Intended Functionality (SOTIF) standard came into being. In the SOTIF standard, the perception link of autonomous driving is the most important and complex. The existing perception network framework still has the problem of insufficient accuracy, and there are many unknown Out-of-Distribution (OOD) scenes in the real road scene. However, previous research mainly focuses on the research of uncertainty methods and complex scene detection technology, the cause of danger, theoretical analysis and evaluation systems under the SOTIF standard, and its perception application under the SOTIF standard for autonomous driving, which are still lacking. This survey introduces the background of autonomous driving under the SOTIF standard. It analyzes the two major problems faced by the perception system: uncertainty factors and complex scene detection, including the causes and classification of the problems. Secondly, the perception detection evaluation system, which aligns with the SOTIF standard, was proposed and compared with the traditional evaluation criteria. Thirdly, the application of uncertainty techniques in the SOTIF perception tasks is analyzed, including uncertainty probability methods, OOD problem solving, spatial feature theory, Dirichlet networks, and uncertainty decomposition. Finally, we use the new evaluation system to discuss the advantages and disadvantages of various uncertainty methods under the SOTIF standard and summarize the current datasets that meet the requirements of the SOTIF standard. The paper also pointed out the shortcomings and directions for improvement of SOTIF-related technologies in autonomous driving scenarios. This research will promote the development and application of uncertainty quantification techniques under the SOTIF standard. To facilitate future research, we create a repository that includes links to relevant reviews and methodological code for learning at https://github.com/sotif-ma/ UQ-for-SOTIF.