基于高速公路自然驾驶数据的自动驾驶车辆跟驰智能性评价模型

An Evaluation Model ofAutonomous Vehicle Car-Following Intelligence Based on Expressway Natural Driving Data

为解决当前自动驾驶车辆跟驰智能性评价中存在的以主观评价为主、缺少微观驾驶行为数据支撑的问题,以高速公路自然驾驶数据为基础,从自动驾驶车辆与人工驾驶车辆驾驶行为一致性的角度出发,构建自动驾驶车辆跟驰智能性评价模型。首先,通过无人机视频拍摄和图像处理,获取了国内18个省份部分高速公路上的高精度车辆轨迹,利用K-means聚类方法提取了15446组稳定跟驰数据。然后,采用描述性统计方法对速度、加速度、跟车间距及跟车时距等指标进行分析。通过Gamma分布拟合不同速度下的跟车间距,以不同速度下跟车间距众数为中心,将跟车间距按照样本量的70%、20%、10%划分为与人工驾驶车辆驾驶行为一致性较好、一般、较差等3种情况,以此为基础建立自动驾驶车辆跟驰智能性评价模型。最后,通过自动驾驶车辆跟驰试验,证明所建模型适用于自动驾驶车辆跟驰智能性评价,相比既有研究,该模型的特点是能基于全过程、微观跟驰行为数据对自动驾驶车辆做出综合的量化评价。这表明基于自然驾驶数据与驾驶行为一致性构建的模型能客观、量化评价自动驾驶车辆跟驰行为,可用于自动驾驶车辆跟驰行为研究与技术参数设计。

In order to solve the problems of subjective evaluation and lack of micro driving behavior data support in the current evaluation of autonomous vehicles for car-following intelligence,this paper constructed a car-following intelligence evaluation model for autonomous vehicles based on natural driving data on expressways,and considering the similarity in driving behavior between autonomous and manual vehicles.Firstly,the high-precision vehicle trajectories on some expressways in 18 provinces of China were obtained through drone photography and image processing,and K-means clustering method was used to extract 15446 sets of stable car-following data.Then,descriptive statistical method was used to analyze the indicators of speed,acceleration,following distance,and following time distance.Afterwards,the following distance at different speeds was fitted using a Gamma distribution.Taking the mode of following distance at different speeds as the center,based on 70%,20%,and 10%of the sample size,the following distance was divided into three categories:good consistency,moderate consistency,and poor consistency with manual driving behavior.Finally,a car-following evaluation model for autonomous vehicles was established,and empirical analysis was conducted through car-following experiments on autonomous vehicles.The results showed that the model was suitable for evaluating the following performance intelligence of autonomous vehicles;compared to previous research findings,the model′s feature is providing a comprehensive quantitative evaluation of autonomous vehicles based on the entire process and microscopic car-following behavior data.It indicates that the model constructed based on the consistency between natural driving data and driving behavior can objectively and quantitatively evaluate the following behavior of autonomous vehicles,and can be used for research on following behavior of autonomous vehicles and the development of technical parameters.