Autonomous driving systems must provide safe, predictable, and consistent behaviors across diverse scenarios to enhance user experience. The geometric driver risk field (GDRF) method is proposed for human-like, interpretable, and fast risk estimations. The method models the driver’s subjectively perceived risk with fields formed by sets of geometric shapes. Unlike current safety measures adopted in trajectory planning algorithms, the proposed method aligns with driver risk perception properties. It provides risk estimations based on probable future vehicle states and the risk consequences at different locations. Compared with existing potential field methods for risk estimations, the mathematical form of the method has a high calculation efficiency while keeping the ability to model the details of the risk information. Influences of different obstacle types, multiple regions with different risk levels, and the relative motions between vehicles are modeled in a unified manner. The graph-search and optimization trajectory planning algorithms with a hierarchy framework are also designed to attain desired risk objectives in the planned trajectories. The framework mitigates the possible suboptimality problems in previous hierarchy trajectory planning algorithms. The proposed GDRF method and planning algorithms are examined in driving scenarios with different traffic states and road geometries. Results showed that the proposed risk estimation method outperforms previous risk-oriented potential field methods by more human-like and safer trajectories with evidence from existing studies and safety measures like time headway (THW), time to collision (TTC), and time to lane crossing (TLC). It also provides more consistent behaviors while maintaining a short computation time. The integration of the proposed risk estimation method and trajectory planning algorithms holds great potential for improving the safety and user experience of autonomous driving systems in real-world scenarios.

自动驾驶系统必须在不同场景下提供安全、可预测且一致的行为，以增强用户体验。提出了几何驾驶员风险场（GDRF）方法，用于类人、可解释且快速的风险估计。该方法利用由几何形状组形成的场对驾驶员主观感知的风险进行建模。与当前轨迹规划算法中采用的安全措施不同，所提出的方法符合驾驶员风险感知特性。它根据未来可能的车辆状态和不同地点的风险后果提供风险估计。与现有的风险估计势场方法相比，该方法的数学形式在保持对风险信息细节建模的能力的同时，具有较高的计算效率。以统一的方式对不同障碍物类型、不同风险级别的多个区域以及车辆之间的相对运动的影响进行建模。具有层次结构的图搜索和优化轨迹规划算法也旨在在规划轨迹中实现所需的风险目标。该框架缓解了先前层次轨迹规划算法中可能存在的次优问题。所提出的 GDRF 方法和规划算法在不同交通状态和道路几何形状的驾驶场景中进行了检验。结果表明，所提出的风险估计方法优于以前的风险导向势场方法，具有更人性化和更安全的轨迹，并有现有研究和安全措施（例如车头时间（THW）、碰撞时间（TTC）和车道时间）的证据交叉（TLC）。它还提供更一致的行为，同时保持较短的计算时间。所提出的风险估计方法和轨迹规划算法的集成对于提高现实场景中自动驾驶系统的安全性和用户体验具有巨大的潜力。