基于自然驾驶数据的跟车场景潜在风险评估

为弥补传统风险评价指标对相对速度较小的跟车场景危险水平评价能力的不足，减少跟车场景中追尾事故的发生，提出了跟车场景潜在风险的概念。将假定前车以较大制动减速度减速条件下的风险称为潜在风险，并构建了代表驾驶人在潜在危险跟车场景下进行避撞操作需满足的最大反应时间的参数时间裕度（TM）。由于追尾危险工况中常见采取的避撞操作可分为制动和制动转向两大类，分别对典型制动避撞过程和制动转向避撞过程进行了分析，从而推导出分别针对2种跟车潜在危险场景的TM计算方式。通过自动筛选与人工筛选结合，获得了中国自然驾驶数据库（China-FOT）中具有中国驾驶人特点的制动避撞危险工况87个和转向制动避撞危险工况40个进行分级，并基于图像处理方法提取了前车制动开始时刻的TM值，从而得到跟车场景潜在风险两级危险域的划分。结果表明：制动避撞场景下，本车车速过低和过高时，TM值的变化均不明显；而在制动转向避撞场景中，只有速度较高时阈值才保持不变。通过对正常驾驶视频的分析，引入对比组共计正常跟车制动工况163例和正常跟车转向变道工况151例的车头时距（THW）值，其统计分析结果与支持向量机分类结果均难以清晰描述跟车场景危险水平与本车速度之间的关系。为研究跟车场景潜在风险评价在自动驾驶中的应用，对于制动避撞场景，在设定TM值不变和相对速度不变的条件下，分别对基于TM的最小相对距离和距离碰撞时间（TTC）值进行了仿真，仿真结果显示，相比于TTC而言，TM的评价稳定性受相对速度影响小，在自动驾驶跟车策略开发和避免其发生责任追尾事故中有重要应用价值。

Assessment of the Potential Risk in Car-following Scenario Based on Naturalistic Driving Data

The concept of potential risk in a car-following scenario with low relative velocity is proposed to avoid rear-end collision and to counter the disadvantages of traditional risk assessment indicators. The potential risk refers to the existing risk of a front car braking with a high braking deceleration. The time margin (TM) is created as an assessment parameter, which represents the driver maximum response time for taking action to avoid a collision in a potentially risky car-following scenario. The common collision avoidance operations taken in critical rear-end situations can be classified into two types: braking, and steering while braking. These two typical processes were analyzed to determine the calculation methods for TM in both types of car-following scenarios with potential risk. In a combination of automated and manual screening, 87 cases using braking and 40 cases using braking and steering were obtained from the China Field Operation Test (China-FOT) database, which includes characteristics of Chinese drivers. After grading these cases by risk level, the TM values at the first moment when the front car brakes were extracted via video post-processing. As a result, two levels of potentially risky car-following scenarios were distinguished on the basis of the TM. The results indicate that by choosing only braking to avoid collision, the variation of TM is not obvious when the velocity of the ego-car is either too low or too high. However, for drivers that chose both braking and steering operations, the threshold is only stable with a high ego-car velocity. Based on the analysis of normal driving records, the time headway (THW) values of 163 normal car-following braking cases and 151 normal car-following lane-changing cases were introduced as a comparison group. However, the statistical and classification results obtained by support vector machine could not be used to describe the relationships between velocity and risk level in those car-following scenarios. Under the conditions that the TM value is constant (group 1) and the relative speed is constant (group 2), the minimum relative distance calculated based on TM and time-to-collision (TTC) were simulated, respectively. To investigate the use of TM in autonomous driving, especially for braking cases, simulation results show that the evaluation stability of TM is better than TTC, which is only slightly affected by the relative speed. This result has critical application value to the strategy of avoiding rear-end collisions of autonomous vehicles in car-following scenarios.