Adoption of autonomous vehicles on urban streets requires safety validation and assurance of driverless vehicles against the diversity and richness of pedestrian behaviors. Recent developments emphasize testing the driving methods in the simulation where traffic participants such as pedestrians are created in such a way that challenges the driving methods by creating risky and unpredictable situations. However, the current pedestrian simulation models are simple and fail to capture the diversity and richness of real-world pedestrians and the scenarios involving pedestrians. Therefore, safety validation and reporting fail to explicitly state how well the driving systems tackle the different types of pedestrians in different situations.

A lack of knowledge about how pedestrian behavior behaves on the road is a significant barrier to safety validation and assurance. This work aims to systematically define how pedestrians behave on the road to shed light on the requirements of a simulation model that reasonably represents pedestrians. It identifies key pedestrian types and proposes a pedestrian behavior ontology that facilitates simulation models and communication among the stakeholders of autonomous vehicles.

Often, the scenarios that emerge from pedestrian behavior and traffic situations are too rare and do not emerge in scenario-based testing approaches in simulation. Concrete reconstruction of such scenarios cannot adapt to the changing behavior of self-driving methods. The work also aims to reproduce and adapt such rare scenarios in simulation.

In a nutshell, this work provides a clear direction on how to define and model real-world pedestrians and scenarios for simulation-based testing of self-driving vehicles in such a way that the test results are interpretable, reliable, and communicable. It proposes several novel methods following the direction.

Event Host: Golam Md Muktadir, Ph.D. Candidate, Computer Science & Engineering

Advisor: Jim Whitehead