To improve the agility and efficiency of the highway decision-making system and overcome the local optimal dilemma of the existing safety field, this paper builds an improved safety field to reflect the advantage of the reachable states and the learning process is further employed to make the decision long-term optimal. Firstly, the improved safety field is prepared by the kinematic model-based prediction of surrounding vehicles and the boundary is determined elaborately to ensure real-time performance. Then, the field is constructed by three individual fields. One is the kinematic field, which is built based the safe-distance model to measure the colliding risk of both moving or no-moving objects accurately. Another is the road field that reflects the lane-marker constraint. The last is the efficiency field, which is introduced creatively to improve efficiency. Furthermore, the learning algorithm is adopted to learn the long-term optimal state-action sequence in the safety field. Finally, the simulations are conducted in Prescan platform to validate the feasibility of the improved safety field in complex scenarios. The results show that the proposed decision algorithm can always drive autonomous vehicle to the state with a long-term optimal payoff and can improve the overall performance compared to the existing pure safety field and the interaction-aware method.

为了提高公路决策系统的敏捷性和效率，克服现有安全场的局部最优困境，本文构建了一个改进的安全场来体现可达状态的优势，并进一步采用学习过程使决策长期最优。首先，通过基于运动学模型对周围车辆的预测来准备改进的安全场，并精心确定边界以确保实时性。然后，该字段由三个单独的字段构建。一种是运动场，它是基于安全距离模型建立的，以准确测量移动或静止物体的碰撞风险。另一个是反映车道标记约束的道路字段。最后是效率字段，创造性地引入以提高效率。此外，采用学习算法学习安全领域的长期最优状态-动作序列。最后，在 Prescan 平台上进行仿真以验证改进的安全场在复杂场景中的可行性。结果表明，与现有的纯安全领域和交互感知方法相比，所提出的决策算法可以始终将自动驾驶汽车驱动到具有长期最优收益的状态，并且可以提高整体性能。