Real-world Test Drive Vehicle Data Management System for Validation of Automated Driving Systems
(2019)
For the validation of autonomous driving systems, a scenario-based assessment approach seems to be widely accepted. However, to verify the functionality of driving functions using a scenario-based approach, all scenarios that may be relevant for the validation have to be identified. Real-world test drives are mandatory to find relevant and critical scenarios. However, the identification of scenarios and the management of the captured data requires computational assistance to validate driving functions with reasonable effort. Therefore, this work proposes a highly-modularised multi-layer Vehicle Data Management System to manage and support analysing large-scale test campaigns for the scenario-based validation of automated driving functions. The system is capable of aggregating the vehicle sensor data to time-series of scenes by utilising temporal discretisation. Those scenes will be enriched with information from various external sources, providing the foundation for efficient scenario mining. The practical usefulness of the proposed system is demonstrated using a real-world test drive sequence, by finding lane-change scenarios and evaluating an onboard system.
For the introduction of new automated driving functions, the systems need to be verified extensively. A scenario-driven approach has become an accepted method for this task. But to verify the functionality of an automated vehicle in the simulation in a certain scenario such as a lane change, relevant characteristics of scenarios need to be identified. This, however, requires to extract these scenarios from real-world drivings accurately. For that purpose, this work proposes a novel framework based on a set of unsupervised learning methods to identify lane-changes on motorways. To represent various types of lane changes, the maneuver is split up into primitive driving actions with an Hidden Markov Model and Divisive Hierarchical Clustering. Based on this, lane change maneuvers are identified using Dynamic-Time-Warping. The presented framework is evaluated with a real-world test drive and compared to other baseline methods. With a f1 score of 98.01\% in lane-change identification, the presented approach shows promising results.