With the development of intelligent driving technology, human-machine cooperative driving is significant to improve driving safety in abnormal situations, such as distraction or incorrect operations of drivers. For human-machine cooperative driving, the capacity of pedestrian collision avoidance is fundamental and important. This paper proposes a novel learning-based human-machine cooperative driving scheme (L-HMC) with active collision avoidance capacity using deep reinforcement learning. Firstly, an improved deep Q-network (DQN) method is designed to learn the optimal driving policy for pedestrian collision avoidance. In the improved DQN method, two replay buffers with nonuniform samples are designed to shorten the learning process of the optimal driving policy. Then, a human-machine cooperative driving scheme is proposed to assist human drivers with the learned driving policy for pedestrian collision avoidance when the driving behavior of human drivers is dangerous to the pedestrian. The effectiveness of the human-machine cooperative driving scheme is verified on the simulation platform PreScan using a real vehicle dynamic model. The results demonstrate that the deep reinforcement learning-based method can learn an effective driving policy for pedestrian collision avoidance with a fast convergence rate. Meanwhile, the proposed human-machine cooperative driving scheme L-HMC can avoid potential pedestrian collisions through flexible policies in typical scenarios, therefore improving driving safety.

随着智能驾驶技术的发展，人机协作驾驶对于提高驾驶员分心或不正确操作等异常情况下的驾驶安全具有重要意义。对于人机协作驾驶来说，避免行人碰撞的能力是至关重要的。本文提出了一种新的基于学习的人机协同驾驶方案（L-HMC），该方案采用深度强化学习技术，具有主动防撞能力。首先，设计了一种改进的深度Q网络（DQN）方法，以学习避免行人碰撞的最佳驾驶策略。在改进的DQN方法中，设计了两个样本不一致的重播缓冲区，以缩短最佳驾驶策略的学习过程。然后，提出了一种人机协同驾驶方案，以在驾驶员的驾驶行为对行人构成危险时，为驾驶员提供学习的避免行人碰撞的驾驶策略。在仿真平台PreScan上使用真实的车辆动力学模型验证了人机协作驾驶方案的有效性。结果表明，基于深度强化学习的方法能够以快速收敛的速度学习有效的行人避碰驾驶策略。同时，提出的人机协同驾驶方案L-HMC在典型情况下可以通过灵活的策略避免潜在的行人碰撞，从而提高了驾驶安全性。