MetaScenario: A Framework for Driving Scenario Data Description, Storage and Indexing

作者：Cheng Chang, Dongpu Cao, Long Chen, Kui Su, Kuifeng Su, Yuelong Su, Fei–Yue Wang, Jue Wang, Ping Wang, Junqing Wei, Gansha Wu, Xiangbin Wu, Huile Xu, Nanning Zheng, Li Li · 发表于：IEEE Transactions on Intelligent Vehicles · 年份：2022 · DOI：10.1109/tiv.2022.3215503 · 被引用次数：46 · 研究领域：Data Management and Algorithms、Traffic Prediction and Management Techniques、Time Series Analysis and Forecasting

Autonomous driving related researches require the analysis and usage of massive amounts of driving scenario data. Compared to raw data collected by sensors, scenario data provide a preliminary abstraction of driving tasks and processes, explicitly integrate information about the road environment and the dynamic and static attributes of traffic participants, making it easier to conduct task understanding and decision making. However, many existing driving scenario datasets have the following two problems. First, it is not clear which data fields need to be recorded for driving scenarios. The data storage formats and organization standards are inconsistent. Second, the datasets cannot establish driving scenario indexing effectively. Existing datasets are sparsely annotated and difficult to index, which is detrimental to data sampling and extraction for machine learning process, thus hindering efficient fusion and reuse. In this paper, we propose MetaScenario, a framework for driving scenario data. We describe driving scenarios and design the centralized and unified data framework for the storage, processing, and indexing of scenario data based on relational database. The concept of atom scenario is proposed and characterized using semantic graphs. We also annotate and classify behaviors and interactions of traffic participants in atom scenarios by extracting the spatiotemporal evolution of semantic information. The annotation facilitates the indexing and extraction of data. The...