Batch human-like trajectory generation for multi-motion-state NPC-vehicles in autonomous driving virtual simulation testing

Non-player character vehicles (NPC-Vs) denote crucial components of autonomous driving systems (ADSs) and autonomous driving assistance algorithms (ADAAs) when conducting virtual simulation testing (VST). Human-like behaviors and trajectories of NPC-Vs could provide information on complex background traffic flow to the tested ADS and ADAA, thus ensuring rigorous tests on the reliability and stability of the ADS and ADAA. However, a VST based on data injection faces the problems of a small amount of data and difficulty in extracting critical scenarios and methods’ transplantation and reuse. To address these problems, this study takes intersection as a research scenario and proposes a probability-limited parameter combination method and a learning-based batch human-like trajectory generation model to generate different human-like trajectories according to different motion states of vehicles. First, an effective IM-sampling algorithm, which samples trajectory data and obtains equal-number-coordinate trajectories as trajectory generation model labels, is proposed. Second, dependency probabilities between different vehicle kinematic parameters (VKPs) are calculated to form a probability-limited generation tree, which generates different VKP combinations representing different vehicle motion states that are used as trajectory generation model inputs. Finally, a learning-based batch trajectory generation model is developed. After model training and testing, the generated VKP combinations are used for trajectory generation, and the generated trajectories are subjected to the human-like degree analysis considering multiple metrics. The experimental results show that the proposed model is capable of generating more complex human-like trajectories and behaviors than real trajectories in batch. The proposed model could be used to generate complex and human-like NPC-V trajectories for the autonomous driving VST and thus accelerate the autonomous driving VST.