One Ontology to Rule Them All: Corner Case Scenarios for Autonomous Driving

Abstract

The core obstacle towards a large-scale deployment of autonomous vehicles currently lies in the long tail of rare events. These are extremely challenging since they do not occur often in the utilized training data for deep neural networks. To tackle this problem, we propose the generation of additional synthetic training data, covering a wide variety of corner case scenarios. As ontologies can represent human expert knowledge while enabling computational processing, we use them to describe scenarios. Our proposed master ontology is capable to model scenarios from all common corner case categories found in the literature. From this one master ontology, arbitrary scenario-describing ontologies can be derived. In an automated fashion, these can be converted into the OpenSCENARIO format and subsequently executed in simulation. This way, also challenging test and evaluation scenarios can be generated.

From Traffic Sensor Data To Semantic Traffic Descriptions: The Test Area Autonomous Driving Baden-Württemberg Dataset (TAF-BW Dataset)

“…Another topic that is discussed in the context of scenariobased testing are corner cases. Bogdoll et al [32] propose the use of ontologies to help create and model different corner case scenarios for autonomous vehicles. The proposed master ontology is able to model any scenario found in the research literature related to corner cases, which can be converted into the OpenSCENARIO format and tested in simulation.…”

“…Most ontologies are in the range where the rough geometry of a vehicle, but mostly their relations and trajectories or actions can be described. In general, a trend can be seen that newer ontologies and V&V methods, for example from Feilhauer [29], Medrano-Berumen [28] or Bogdoll [32], have a higher level of detail, which could be explained by more comprehensive simulation tools and more available computing power.…”

“…positions (p), velocities (v) or accelerations (a) are relevant and how they behave in general in the scenario (actions/intentions). Moreover, we agree that actions or at least the intentions of the traffic participants should be Bagschik [13] p Klück [10] pv Wotawa [9] pva Driver Type Chen [14] pv Medrano-Berumen [28] pv Feilhauer [29] pva Schuldt [45] v Elgharbawy [5] p Ulbrich [2] p Hülsen [35] p Büchel [36] pv Regele [37] pva Bogdoll [32] pva Huang [39] pv Zhao [40] pv Fang [38] pv Kohlhaas [41] pv Herrmann [31] p Westhofen [33] pva explicitly specified. A general rule does not always lead to the expected behaviour.…”