Dynamic-Occlusion-Aware Risk Identification for Autonomous Vehicles Using Hypergames
MetadataShow full item record
A particular challenge for both autonomous vehicles (AV) and human drivers is dealing with risk associated with dynamic occlusion, i.e., occlusion caused by other vehicles in traffic. In order to overcome this challenge, we use the theory of hypergames to develop a novel dynamic-occlusion risk measure (DOR). We use DOR to evaluate the safety of strategic planners, a type of AV behaviour planner that reasons over the assumptions other road users have of each other. We also present a method for augmenting naturalistic driving data to artificially generate occlusion situations. Combining our risk identification and occlusion generation methods, we are able to discover occlusion-caused collisions (OCC), which rarely occur in naturalistic driving data. Using our method we are able to increase the number of dynamic-occlusion situations in naturalistic data by a factor of 70, which allows us to increase the number of OCCs we can discover in naturalistic data by a factor of 40. We show that the generated OCCs are realistic and cover a diverse range of configurations. We then characterize the nature of OCCs at intersections by presenting an OCC taxonomy, which categorizes OCCs based on if they are left-turning or right-turning situations, and if they are reveal or tagging-on situations. Finally, in order to analyze the impact of collisions, we perform a severity analysis, where we find that the majority of OCCs result in high-impact collisions, demonstrating the need to evaluate AVs under occlusion situations before they can be released for commercial use.
Cite this version of the work
Maximilian Kahn (2021). Dynamic-Occlusion-Aware Risk Identification for Autonomous Vehicles Using Hypergames. UWSpace. http://hdl.handle.net/10012/17774