Abstract:

An integrated framework coupling decision-making with trajectory planning was proposed to enhance the emergency collision avoidance capability of vehicles in high-speed multi-obstacle scenarios and address the challenge of real-time responsiveness in decision-making and planning due to computational complexity. The high-dimensional game problem was simplified into a sequence of single-obstacle interaction processes by establishing a multi-vehicle non-cooperative game model to describe dynamic interactions and designing a sequential decision-making mechanism based on threat assessment. A graphics processing unit (GPU)-accelerated trajectory optimization algorithm was implemented using the open source machine learning framework PyTorch, generating safe and comfortable collision avoidance trajectories while satisfying vehicle dynamic constraints. The results show that the average decision-making computation time of the proposed method in typical high-speed scenarios is 20~50 ms, and trajectory planning takes 33.1~149.1 ms, outperforming traditional model predictive control (MPC) methods. The lateral velocity and acceleration of the planned trajectories are controlled within 4.0 m/s and 4.0 m/s², respectively, meeting safety and comfort requirements. When tracking the planned trajectories, the maximum lateral tracking error and speed error are 0.22 m and 0.59 m/s, respectively, fulfilling the requirements for high-speed emergency collision avoidance. In CARLA simulations, successful collision avoidance is achieved in all scenarios. The conclusion demonstrates that the proposed framework effectively balances decision-making optimality and real-time performance, providing a reliable solution for vehicle active safety in complex scenarios.

Key words: automotive active safety, complex scenarios, emergency collision avoidance decision-making, trajectory planning and tracking, open source machine learning framework PyTorch