Abstract: Highway cut-in scenarios can easily cause rear-end collisions and traffic flow disturbances, which is a key scenario for autonomous driving safety testing. Test parameters are usually directly sampled within a wide range, with initial spacing, speed difference, and lane changing time each bounded and combined, ignoring their kinematic coupling. The large number of samples that cannot be achieved or are far from the safe haven boundary dilutes the testing budget. A parameter optimization method for highway cut-in scenarios was proposed. Firstly, by combining horizontal accessibility with vertical relative motion, a physical support domain was constructed to eliminate parameter combinations that were not feasible or had unreasonable interactions. Subsequently, normalized safety margins based on vehicle response and braking capabilities were constructed, characterizing the proximity of parameter points to the safe haven boundary. Finally, the risk density within the critical region was estimated and the compact non-rectangular parameter subdomains was screened out through equivalent set filtering. Given the vehicle capability parameters and finite time domain kinematic discriminator, this subdomain occupied 4.35% of the global parameter space, and the collision sample rate increased from 31.09% of the full parameter space sampling to 86.76%. It was found that the average number of tests for a collision sample decreased from 3.22 to 1.15. Under 36 sets of grids and smoothing settings, feasible regions were obtained, and the collision rate was verified to be between 85.5% and 87.9%. Exploratory analysis of highD trajectories in the dataset shows that there is also similar enrichment in natural driving, that is, sorted by simulated risk density, the proportion of danger in the top 25% of samples is 13.64%, which is 3.41 times that of random testing. This proposed method is located upstream of scenario sampling, and the output can be directly used as parameter input for methods such as importance sampling.

Key words: traffic and transportation engineering; autonomous driving; cut-in scenario; parameter optimization; safety testing; accelerated testing

摘要： 高速公路切入易引发追尾与交通流扰动，是自动驾驶安全测试的重点场景。测试参数通常在较宽范围内直接采样，初始纵向间距、速度差与换道时间各自设界后组合，三者的运动学耦合被忽略，样本大量不可实现或远离避险边界，测试预算随之稀释。提出一种切入参数优选方法。首先，结合横向可达性与纵向相对运动构建物理支撑域，排除不可实现或交互不合理的参数组合；随后，基于车辆反应与制动能力构造能力归一化安全裕度，刻画参数点与避险边界的接近程度；最后，在临界区域内估计风险密度，并通过等值集筛选提取紧凑的非矩形参数子域。结果表明：在给定车辆能力参数和有限时域运动学判别器下，该子域占全局参数空间的4.35%，碰撞样本率由全参数空间采样的31.09%升至86.76%，一个碰撞样本所需的平均测试次数从3.22次降至1.15次；36组网格与平滑设置下均得到可行区域，验证碰撞率为85.5%~87.9%；数据集highD轨迹的探索性分析显示，自然驾驶中亦有类似富集，即按仿真风险密度排序，前25%样本中风险样本占比为13.64%，为随机测试的3.41倍。该方法处于场景采样的上游，输出可直接作为重要性采样等方法的参数输入。

关键词: 交通运输工程；自动驾驶；切入场景；参数优化；安全测试；加速测试