Semantic Classification of Urban Traffic Scenarios for the Validation of Automated Driving Systems

The safety argumentation of an automated vehicle is an essential condition before its use on public roads. For this reason, a thorough Verification and Validation (V&V) process is a fundamental aspect of the development and commercial release for every automated vehicle.
In recent years, a number of scientific publications have reasoned that a distance- based V&V approach, aimed at providing a stochastic safety argument by achieving a desired failure-rate over a defined testing distance, will not be feasible to implement for automated vehicles. The main reason for this is the fact that the distance required to be driven with development vehicles exceeds economical and practical limits by far. To overcome this challenge, scenario-based V&V approaches are currently a subject of many research activities. These methodologies aim to evaluate the safety of an automated vehicle by testing it in a variety of different traffic scenarios. This allows to decompose the safety V&V into smaller units in the form of scenarios instead of having to achieve one large statistical argument for the safety of the system. However, urban traffic scenarios themselves form a complex, high dimensional state-space, which makes it challenging to argue for the completeness of scenario-based V&V approaches.
This work aims to address this challenge through a semantic classification of urban traffic scenarios. By extracting them in a structured manner from large volumes of recorded driving data, it is possible to analyze them statistically and to make empirical, data-driven contributions to the V&V process of automated vehicles. To this end, a catalog of driving maneuvers is introduced to describe the behavior of vehicles in urban traffic on a semantic level. Next, algorithms for the automated classification of these maneuvers are implemented and evaluated with respect to their detection accuracy. Based on this automated maneuver classification, an empirical analysis of urban traffic scenario diversity is conducted. Here, a special focus is put on saturation effects during data collection as well as the observed exposure of various semantic scenario elements.
The results of the investigations provide some of the first quantitative insights into the Long Tail-problem of automated driving V&V, which is often mentioned in current literature in this field. From the empirical findings it is further concluded that a semantic scenario classification has the potential to contribute substantially to a data-driven, scenario-based safety argumentation for automated vehicles.