A Knowledge Graph-Based Approach for Situation Comprehension in Driving Scenarios

Making an informed and right decision poses huge challenges for drivers in day-to-day traffic situations. This task vastly depends on many subjective and objective factors, including the current driver state, her destination, personal preferences and abilities as well as surrounding environment. In this paper, we present CoSI (Context and Situation Intelligence), a Knowledge Graph (KG)-based approach for fusing and organizing heterogeneous types and sources of information. The KG serves as a coherence layer representing information in the form of entities and their inter-relationships augmented with additional semantic axioms. Harnessing the power of axiomatic rules and reasoning capabilities enables inferring additional knowledge from what is already encoded. Thus, dedicated components exploit and consume the semantically enriched information to perform tasks such as situation classification, difficulty assessment, and trajectory prediction. Further, we generated a synthetic dataset to simulate real driving scenarios with a large range of driving styles and vehicle configurations. We use KG embedding techniques based on a Graph Neural Network (GNN) architecture for a classification task of driving situations and achieve over 95% accuracy whereas vector-based approaches achieve only 75% accuracy for the same task. The results suggest that the KG-based information representation combined with GNN are well suited for situation understanding tasks as required in driver assistance and automated driving systems.