A Combined Maneuver-Trajectory Approach to Cluster Real-World Driving Data

The advancement of highly automated driving presents significant opportunities for mobility and safety. However, the widespread integration of this technology faces a key challenge - the Validation and Verification (VV) of driving functions within new Operational Design Domains (ODDs). Running extensive testing campaigns for every new ODD is not feasible. Scenario-based testing aims to describe ODDs using scenarios that can represent different driving situations. The aim of this approach is to aggregate reoccurring scenarios into clusters and describe the ODDs through these clusters. This work proposes using traffic observation data to generate scenarios and cluster them using a combination of trajectory-based clustering and maneuver identification. Analysis using manually crafted reference clusters shows that the combined approach produces a higher clustering quality (v-measure 0.946 to 0.921 or 0.811) than a trajectory-based or maneuver-based approach alone. Furthermore, the combined approach can be improved more effectively by the introduction of a two-stage clustering process.