Data-driven configurable scenario generation for testing autonomous driving systems in highway environments

Abstract

Human-like traffic flow provides a test environment suitable for evaluating the safety of autonomous driving systems. Currently, simulation testing based on data injection faces the problems of small data volumes and high acquisition costs. Although previous studies have been conducted on scenario generation, the following shortcomings remain: the inability to conduct continuous-scenario generation, lack of real-time simulations, and reliance on simulations oriented toward a single-function scenario. To address these shortcomings, this study proposed the concept of behavior incentives as a basis for configurable continuous-scenario generation. First, to better extract the behavioral characteristics of a vehicle, a sampling method was proposed to dimensionally homogenize vehicles’ sequence data. Second, using these processed data, the type of behavior incentive and its numerical format were determined, and a unified behavior incentive framework was developed and populated. Additionally, to complete the lane changing information in the behavior incentive, the vehicle motion and trajectory data were resampled, a velocity-trajectory generation neural network was proposed, and the lane changing trajectory for the behavior incentive framework was generated. After completing all behavior incentive frames, the proposed method was simulated in real time using the Simulation of Urban Mobility traffic-flow simulation software, and the key parameters and functions of the simulation were identified. The simulation results show that the proposed method can not only effectively generate continuous test scenarios, but can also facilitate the addition and modification of parameters to generate configurable test scenarios comprising different states, providing an excellent basis for testing autonomous driving systems.