面向自动驾驶的驾驶环境视觉感知复杂度量化评估方法

Method for the quantitative evaluation of the visual perception complexity of driving environments for autonomous vehicles

为优选自动驾驶汽车开放测试、示范道路并支撑其驾驶环境的优化,提出面向自动驾驶的驾驶环境视觉感知复杂度量化评估方法。以百度街景地图作为驾驶环境数据源,运用脚本文件以及截图工具PicPick搭建自动化街景图像数据提取平台,并在不同区域、不同道路等级下采集上海市50条典型道路的驾驶环境数据;从行人、交通标志、交通标线、红绿灯、车辆5方面出发,构建驾驶环境要素感知平台,并开展感知精度的量化评估;在单要素感知准确率的基础上,采用熵权法确定多维感知要素权重,计算各道路综合感知准确率,并应用轮廓系数法与K-means++聚类算法进行视觉感知复杂度分级。结果表明,上海市典型道路的驾驶环境视觉感知复杂度分为三级,大部分道路的视觉感知复杂度属于2级;对比不同等级道路发现,支路的视觉感知复杂度总体上低于主干路。

To support the optimal selection of public testing road sites and provide instructions for improving driving environments for autonomous vehicles,a quantitative evaluation method for driving environment visual perception complexity was proposed.Based on the driving environment dataset from Baidu street view map,an automated extraction platform of street view image data was established using a script file and the screenshot tool PicPick.Driving environment data of 50 typical Shanghai roads of different areas and different road grades were collected.Then,an element perception platform of driving environment was established from five aspects—pedestrians,traffic signs,road markings,traffic lights,and vehicles—to quantitatively evaluate perception accuracy.Based on single-element perception accuracy,the weights for multidimensional perception elements were determined using the entropy method,and integrated perception accuracy at the road level was calculated.The silhouette coefficient method and K-means++clustering algorithm were used to propose the classification of visual perception complexity.Results showed that the visual perception complexity of the road network in Shanghai can be divided into three levels.The majority of driving environment visual perception complexities of roads belong to level 2.In addition,comparisons among roads of different grades indicated that the visual perception complexity of local roads is generally lower than that of arterial roads.