Detection and Analysis of Trajectory-Related Corner Cases at a Signalized Urban Intersection

Kurzfassung

Automated driving has been an emerging technology in the automotive industry for several years, and it is expected to revolutionize how we commute. One of the major challenges in automated driving is ensuring that the system can handle all possible driving scenarios, including rare and critical ones, also called corner cases. For the validation of automated driving functions, it is necessary to test the corner cases in simulation environments. Simulation allows for testing complex and rare scenarios that may be difficult or dangerous in the real world. However, the effectiveness of simulation-based testing depends on the availability of realistic test data that accurately reflect real-world scenarios. This work aims to detect, cluster, and parametrize rare and unforeseen traffic scenarios based on real-world traffic data from an urban intersection. By doing so, we provide realistic test cases for the validation of automated vehicles. Corner cases can be challenging to detect and classify due to their infrequency and complexity. Therefore, we conduct long-term traffic observations and apply anomaly detection methods and surrogate measures of safety to classify corner cases as rare and critical compared to normal behavior. Afterward, we use the characterization of the 6-layer model from the PEGASUS project to describe the whole traffic environment, including its traffic participants, and cluster the corner cases in a structured manner. Once the scenarios have been detected and clustered, we parametrize them to create realistic test cases for the validation of automated vehicles. This involves describing the scenario using a digital map and relevant parameters such as the traffic participants' kinematic behavior, interaction level, criticality, the time the scenario occurs, and environmental factors such as weather conditions and traffic light status. For this study, the beforementioned methodology is applied to a large dataset from the AIM Research Intersection in Braunschweig, Germany. The dataset contains trajectory, traffic light, and weather data from a period of multiple weeks and a digital map of the intersection. Our results show that the dataset contains corner cases, including red light violations, emergency vehicles forcing other vehicles to stop inside the intersection, large trucks utilizing multiple lanes for turning maneuvers, unusual traffic events, and critical braking maneuvers, some of them under rare environmental conditions. The comparison of the detected corner cases with the normal behavior of traffic participants in similar scenarios shows the unusualness of the corner cases. Furthermore, the scenarios are clustered, parameter distributions are generated, and realistic test cases are derived. In summary, this work provides data from trajectory-related corner cases by extracting relevant scenarios from a large dataset of trajectories, traffic light states, and weather conditions from an urban intersection. Following the 6-layer model and considering many parameters to describe the traffic scenarios, we show that the infrastructural setup enables detecting multiple corner cases like traffic rule violations, unusual traffic events, and critical events leading, for example, to strong braking maneuvers. These corner cases are clustered and parameterized to provide real-world traffic scenarios for the testing of automated vehicles. For now, our methodology for detecting corner cases mainly depends on expert knowledge, i.e., we define the desired corner case and the parameters required for detection. Further work may focus on a more general approach to corner case detection that does not require defining the desired corner case.