Scene to Scenario: Data-driven Pipeline for Extracting and Re-Simulation of Test Scenarios for Highly Automated Driving Functions

The development of highly automated driving has made remarkable progress in recent years. Numerous technical challenges have been solved, allowing the first partly autonomous vehicles to drive on public roads. However, the safeguarding process of Highly Automated Vehicles (HAVs) remains a significant challenge. The verification and validation of a corresponding system play a crucial role for manufacturers, as they ensure that their own systems are free of faults. Furthermore, verification carried out by testing institutions is relevant in order to allow new systems on public roads without posing a safety risk to traffic participants.
The highly complex nature of HAVs makes validation utilizing statistical tests considerably more difficult. In this context, scenario-based testing in a simulation opens up promising perspectives for the validation.

This dissertation presents and discusses a concept for the creation and evaluation of test scenarios for Highly Automated Driving Functions (HADFs).
An essential aspect in the creation of a test scenario is the mutual consideration of the behavior of all traffic participants within the simulation.
... mehr
The trajectory and solution strategies may vary depending on the HADF being tested. Consequently, the other simulated agents must react consistently.
To remain as comprehensive as possible, the dissertation does not consider any specific HADF and its behavior. Rather, this thesis deals in particular with the initial configuration, i.e. the traffic scene, which serves as the starting point for conducting a subsequent test.

A description model, referred to as the Semantic Scene Graph, is presented. This model allows the description of individual traffic scenes based on the constellations of traffic participants and relationships between them. The model facilitates a direct comparison between traffic scenes and serves as a basis for following modules.
In the next step, the abstract description model is used to check scenes for similarity using a graph-based self-supervised machine learning method. The result is a structured scene space that can be organized into clusters of similar traffic scenes.

After the scenes have been grouped into similar categories, they have to be evaluated based on their relevance for testing. A number of different criticality metrics serve as an objective measure to evaluate the criticality of a situation.
In order to make a meaningful statement about the relevance of an initial scene (seed-scene), many different possible futures are simulated that can develop from a seed-scene.
In addition to rule-based methods with varying parameterizations and combinations, two machine-learned behavior models developed in this thesis are used to simulate a broad spectrum of potential futures for a seed-scene.
The resulting child-scenarios are then evaluated using the criticality metrics.
A subsequent analysis of the criticality distribution provides insight into how far and in what aspects a seed-scene can be relevant for the generation of a test scenario. The criticality profile, in combination with the information pertaining to the associated cluster, can be utilized to identify traffic scenes that are characterized by important attributes.

In summary, this dissertation presents an approach for identifying important scenes from real-world data that can be used for testing HADFs. Within this pipeline, traffic scenes are evaluated based on their criticality and organized into clusters of similar scenes within the scene space.