无信号灯十字路口自动驾驶汽车直行关键场景构建

自动驾驶汽车在开发过程中需要进行全面的测试以确保其安全性，关键场景是有条件和高度自动驾驶汽车进行安全测试验证的重要基础。首先，针对测试场景参数离散化全组合后测试用例数量巨大的问题，分别采用3路定强度和变强度组合测试策略及场景初筛规则，以自动驾驶主车直行通过无信号灯十字路口及周围2辆干扰车的测试场景为例，使测试用例数量从全组合的3.9×10⁷组分别减少到6 525组和26 496组。其次，采用碰撞时间、后侵占时间和最大减速度作为场景关键性识别和评价指标。利用基于动态安全区域的避障策略和基于模型预测控制的自动驾驶主车运动规划与控制模型，以及搭建的基于MATLAB/Simulink和CarSim软件的自动驾驶汽车联合仿真平台，通过仿真并与安全指标阈值比较，获得数量大幅减少的关键测试用例。最后，采用基于加权欧氏距离的K-medoids聚类方法对变强度组合策略获得的2 234组关键测试用例进行聚类，获得9组典型关键测试用例。研究结果表明：碰撞时间、后侵占时间和最大减速度指标阈值可以用于识别十字路口场景关键性；相对3路定强度参数组合策略，变强度参数组合策略提供了更多的关键测试用例，对其随机抽样获得少量随机关键测试用例；随机和典型关键测试用例可以应用于封闭试验场验证无信号灯十字路口自动驾驶汽车直行运动规划的安全性。

Construction of Critical Scenarios for Automated Vehicle Moving Straight at Intersection Without Traffic Lights

Automated vehicles need to be subjected to extensive tests to ensure their safety during the development process. The critical scenarios are an important basis for the safety testing and verification of conditional and highly automated vehicles. First, given the problem of the many test cases after the full combination of parameter discretization of the test scenarios, the three-way constant-intensity and variable-intensity combinatorial test strategies and scenario preliminary screening rules were adopted. The test scenario in which an automated ego-vehicle moves straight through the intersection with no signal lights and the surrounding two interfering vehicles was taken as an example. The number of test cases were reduced from 3.9×10⁷ groups of the full combination to 6 525 and 26 496 groups. Second, the collision time, post-encroachment time, and maximum deceleration were used as the identification and evaluation indicators of the critical scenarios. Using the obstacle avoidance strategy based on dynamic safety zones and the automated ego-vehicle motion planning and control models based on model predictive control, a joint simulation platform for the automated ego-vehicle based on MATLAB/Simulink and CarSim software was built. The critical test cases with a significantly reduced number were obtained through simulations and comparison with the safety metric thresholds. Finally, the K-medoids clustering method based on weighted Euclidean distance was used to cluster 2 234 sets of critical test cases obtained by the combinatorial approach of variable intensity, and nine sets of typical critical test cases were obtained. The results show that the index thresholds of the collision time, post-encroachment time, and maximum deceleration can be used to identify the criticality of the intersection scenarios. Moreover, the variable-intensity parameter combinatorial strategy provides more critical test cases compared with the three-way constant-intensity parameter combinatorial strategy, and a small number of random critical test cases are obtained through random sampling. The random and typical critical test cases can be used in a closed-test field to verify the motion planning safety for the automated ego-vehicles moving straight at intersections without traffic lights.