To enhance the interpretability and coverage of high-risk scenarios in virtual
test scenarios for autonomous vehicles, we propose a method for generating
virtual test scenarios based on the VI-GAN (vehicle-interactive GAN) game neural
network. This method constructs a converging interaction game model by capturing
the interaction characteristics of vehicles converging on the ramp and those
driving in the main lane. The Nash equilibrium solution of the game strategy and
the convergence data are used to obtain the vehicle priority probability, and
the game model is embedded in the S-GAN neural network model to propose a game
trajectory generation model with the characteristics of a realistic interactive
gaming behavior. Meanwhile, in order to obtain high-risk convergence scenarios,
CT model is introduced to test the combination of real trajectories of
interacting vehicles in the observed area and used in VI-GAN algorithm to
generate more high-risk interaction trajectories with realistic game interaction
behaviors. By comparing VI-GAN with LSTM, S-LSTM, S-GAN, and other trajectory
generation algorithms, the results show that: (1) Compared with other algorithms
such as S-GAN, the model generates converging interaction trajectories in 3.2 s
and 4.8 s time domain, the ADE decreases by 25.30%/18.98%/7.02% on average, and
the FDE decreases by 17.54%/16.16%/7.87% on average. Higher accuracy of
interaction trajectories generated by VI-GAN algorithm. (2) The initial
trajectories were generated using combinatorial testing and combined with game
interaction scenarios in conflict situations. The number of generated
trajectories is 150 times that of the original trajectories. The game-generated
trajectories have more high-risk scenarios and higher scenario coverage compared
to the original trajectories. This is of practical significance for virtual
scenario-enhanced testing of self-driving cars. The virtual scene reinforcement
test is of practical significance.
