Graph-based encoding of curve driving using spatial keypoints

Abstract

Current accident statistics show that the highest rate of fatal traffic accidents in Germany occurs on rural roads, particularly as a result of vehicles leaving the road. Advanced driver assistance systems (ADAS) and highly automated driving functions therefore have high potential to improve safety in this domain. A key challenge is lateral vehicle control, especially the selection of an appropriate trajectory when cornering in automated driving mode (SAE Level 3+). The aim of this work is to derive characteristic driving variants from real-world measurement data, which serve as a basis for the design of automated lateral vehicle control and contribute to achieving high customer acceptance at the same time. For this purpose, extensive data from real world field tests was collected, standardized, and segmented at defined nodes (curve entry, apex, curve exit). A subsequent cluster analysis identified typical driving styles. Based on this, various trajectory variants were systematically generated using graph theory methods. These variants differ in terms of vehicle class, curve radius, and preference for corner-cutting. In addition, environmental influences such as the presence of oncoming traffic were considered. The outcome is a catalog of reality-based trajectories that serves as the basis for future driving functions. This enables further investigations in which the influence of the variants on driving comfort and safety will be evaluated, both in the Dynamic Vehicle Road Simulator (DVRS) and in real-world driving tests with test vehicles.