Generation of Signalized Intersection Test Scenarios Based on Traffic Participant Model

Abstract Safety testing is a crucial phase in the development of autonomous vehicles, with scenarios serving as the fundamental medium for these tests. The construction of scenarios that accurately reflect the real-world traffic behavior is a key challenge for autonomous vehicle testing technology, given the need for authenticity and effectiveness in testing. Addressing this, this study introduces a method for generating test scenarios specifically for intersections. This method involves decoupling the scenario into interactive and non-interactive layers. For the interactive elements, namely the traffic participants, this study proposes a novel approach to construct a human-like decision-making model adapted to intersections, with the OF-T-GAIL algorithm at its core. This algorithm, through a novel perception-decision representation and an omnidirectional coordinate system transformation, effectively reduces the cumulative error in the trajectory generation. Furthermore, this model is extended to multiple agents, allowing the bulk generation of intersection test scenario segments. These scenarios and the traffic participant (TP) models are validated using the SinD dataset, demonstrating the applicability and effectiveness of the approach in creating testing environments for autonomous driving systems.