基于车道检测性能的自动驾驶车辆平原地区高速公路适驾性评估  

Readiness Evaluation on Freeways in Plain Regions for Automated Vehicles Based on Lane Detection Performance

作　　者：叶欣辰 王雪松[1,2] 田炜 YE Xin-chen;WANG Xue-song;TIAN Wei(College of Transportation,Tongji University,Shanghai 201804,China;Key Laboratory of Road and Traffic Engineering of Ministry of Education,Tongji University,Shanghai 201804,China;School of Automotive Studies,Tongji University,Shanghai 201804,China)

机构地区：[1]同济大学交通学院,上海201804 [2]同济大学道路与交通工程教育部重点实验室,上海201804 [3]同济大学汽车学院,上海201804

出　　处：《中国公路学报》2024年第11期220-234,共15页China Journal of Highway and Transport

基　　金：国家重点研发计划项目(2024YFE0115400);上海市2023年度“科技创新行动计划”“一带一路”国际合作项目(23210750500);国家自然科学基金项目(51878498)。

摘　　要：为从车道检测性能层面评估面向自动驾驶车辆的平原地区高速公路适驾性,利用搭载同济大学车载全息信息采集系统的测试车辆,在京沪高速和沈海高速(上海段)开展基于激光雷达的车道检测性能实车测试。将自车与相邻车道环境车辆可能发生侧面碰撞的情况作为安全临界,计算车道宽度检测上限和下限阈值,提取车道检测失效事件。将失效类型作为标签,将道路几何设计、道路路段、交通标线、车辆运行和环境等五大类特征作为输入特征,采用XGBoost集成学习模型,构建失效类型与特征变量的关系。利用SHAP事后解释方法,量化全局特征重要度,解析单个特征和交互特征的变化对失效类型的影响。结果表明:行驶速度、路段类型、下游大型车距离、车道位置、指示和警告标线类型、曲率、纵坡变化率、标线维护情况和车道标线组合类型对失效有影响,重要度依次递减。具体为:(1)当道路主线的平均曲率小于0.4 km~(-1),纵坡变化率大于10%·km~(-1)(凸曲线)或30%·km~(-1)(凹曲线),自车位于出入口主线或右侧车道,渐变段标线衔接不连续,存在指示和警告标线,标线维护不佳,下游大型车距离0~55 m时,车道检测失效概率增加;(2)渐变段标线连续衔接,磨损标线和旧标线养护提升,重要车道施画实线等手段可减小失效概率。研究结果可在车道检测性能层面对平原地区高速公路开展适驾性评估与优化,为交管部门提供自动驾驶车辆设计运行范围的量化依据,为传感器厂商和车企提供激光雷达性能的重点优化方向。To evaluate the readiness of freeways in plain regions for automated vehicles(AVs)from the perspective of lane detection performance,field tests were conducted on the Beijing-Shanghai Freeway and Shenyang-Haikou Freeway in Shanghai using a test vehicle equipped with the Tongji University Road and Traffic Holographic Data Acquisition System.By considering scenarios in which the test vehicle and surrounding vehicles in adjacent lanes could potentially collide laterally as safety-critical conditions,the upper and lower thresholds for lane width detection were calculated,and lane-detection failure events were extracted.The lane-detection failure type was used as the label.Five feature types,namely road geometric design,road section,road marking,vehicle operation,and environment,were considered as input features.Using an XGBoost ensemble learning model,the relationship between the lane detection failure type and the features was established.As a post hoc interpretation technique,the SHapley Additive exPlanations(SHAP)was used to analyze the feature importance and the impacts of individual and interaction features on failure types.The results show that features including speed,segment type,leading-truck distance,lane location,special marking type,average curvature,rate of change of vertical curve,marking condition,and longitudinal line type affect failure probabilities,with feature importance decreasing in order.Specifically,the failure probability increases when the average curvature is smaller than 0.4 km^(-1),the change rates of crest and sag curves exceed 10%·km^(-1) and 30%·km^(-1),respectively,vehicles are on the mainline at entrance or exit or in the right-most lane,lane markings are connected with acceleration or deceleration tapers,special markings are present or poorly maintained,or the leading truck distance ranges from 0 to 55 m.Moreover,approaches such as continuous connection of markings at acceleration or deceleration taper extensions,better maintenance of worn-out and unerased markings,and the use of

关 键 词：交通工程 自动驾驶车辆 实车测试 高速公路适驾性 集成学习模型 车道检测 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]