高精地图的知识图谱表达

Knowledge Graph Expression of High Definition Map

高精地图是自动驾驶的“传感器”,为自动驾驶提供必要的先验数据以及相应的超视距感知、校验定位、动态规划和决策控制。然而,高精地图数据供给与自动驾驶知识需求仍存在鸿沟,包括数据量大导致查询困难、数据关联弱导致语义理解和智能决策困难。知识图谱是将知识以图的结构表达出来,以描述实体及其关系,涉及实体抽取和关系抽取。为此,在高精地图数据基础上,引入知识图谱,提出高精地图知识图谱的构建方法,以架起地图数据供给与驾驶知识需求之间的桥梁,支撑高精地图数据到自动驾驶知识的转化。构建的知识图谱实例,一方面将高精地图海量数据采用图进行了二次表达,建立了类似于索引的结构;另一方面显式表达了面向自动驾驶需求的语义关系。实验结果表明,知识图谱能为高精地图的语义查询、知识推理和局部决策规划提供基础。所提出的方法能实现高精地图先验数据的语义结构化,推进高精地图由数据到信息到知识的跨越,为自动驾驶的落地贡献先验知识。

Objectives:High definition map(HDM)is the“sensor”for automated driving(AD),which integrates the real-time data collected by various sensors and the prior data collected in the early stage,and serves the application of AD,including providing the necessary prior data and the corresponding over-the-horizon perception,calibration positioning,dynamic planning and decision-making and control.However,there still exists a gap between the supply of HDM data and the demand for AD knowledge,including difficulties in data retrieval due to the large volume of data and challenges in semantic understanding and intelligent decision-making due to weak data correlation.Therefore,how to balance the data supply of HDM and the knowledge demand of AD is the main goal of this paper.Methods:Knowledge graph(KG)is a representation of knowledge in a graph structure to describe entities and their relationships,involving entity extraction and relationship extraction,so that it can make the AD with interpretable,understandable and inferential.This means that KG can serve as an alternative and explicit simulation of the human mind and map cognition that driverless vehicles are missing.Therefore,we introduce KG on the basis of HDM data,and propose a framework of HDM-KG-driving task,so as to support the transformation of HDM data to AD knowledge.The construction of HDM-KG adopts the top-down method,that is,the pattern layer is first followed by the data layer.As the conceptual hierarchy of KG,pattern layer defines the concept,attribute and relation of map ontology,and it can explicitly describe the indirect information and implicit correlation information of map domain from the perspective of traffic.The data layer is an instantiation of the pattern layer,a way to populate the pattern layer through instance matching.Results:To verify the validity of the proposed method,a virtual simulation dataset based on OpenStreetMap data is constructed and converted into two formats.The first is the OpenDRIVE format,which is used to build the KG of static data,including the static knowledge of roads,lanes,intersections,road signs and road markings.The other is the data format of CARLA AD-simulator,which is used for AD simulation and the construction of dynamic real-time KG,including the dynamic knowledge of self-driving cars and vehicles in front of them,pedestrians and other traffic participants.Through the application of the resulting HDM-KG in semantic query,knowledge reasoning and local decision planning,the results show that the HDM-KG explicitly expresses the semantic relationship,enhances the application of HDM in AD,and can further improve AD by connecting additional knowledge such as safety requirements,traffic rules or scene background.Conclusions:HDM-KG can structurally and explicitly express the traffic semantics contained in map objects and the semantic relations between objects,traffic rule constraints,traffic scenes and other knowledge,and realize the three-layer evolution of data-information-knowledge of data sources-HDM-KG,providing support for semantic relation query,logical reasoning and motion planning required by driving tasks.In addition,the introduction of KG eliminates the need for driving tasks to perceive and predict objects and semantic relationships from data collected by sensors,which are explicitly defined by HDM-KG,reducing computational latency and im-proving safety and reliability.The top-down method of constructing KG will cost more labor and time,and the subsequent research will carry out the bottom-up method to automatically fill the KG with multi-source data.