Mark2Cause Knowledge Dataset (M2CKD): Road Marking Causal Knowledge for Explainable Situation Awareness in Autonomous Driving

초록

In autonomous driving environments, deriving situational knowledge from the latent meanings and interrelationships among perceived objects is essential for achieving stable driving responses and reliable decision making. However, current perception systems are limited to recognizing only the surface-level attributes of dynamic and static objects in real time, making it difficult to infer the underlying semantics necessary for human-like predictive reasoning. To address this limitation, this paper proposes the Mark2Cause (M2C) framework and the M2C Knowledge Dataset (M2CKD), which enable explainable situation awareness by extracting the causal relationships between object semantics and situational outcomes. The proposed M2C framework models road markings central to driving context reasoning using a Subject-Predicate-Object (S-P-O) triplet structure, formalizing the minimal ontological unit for causal knowledge representation. The triplet structure is extended into a sequential M2C Knowledge inference process, comprising Object Meaning Detection → Rule Interpretation → Result Inference, which generalizes context reasoning patterns and response strategies through the construction of the M2C Knowledge Dataset (M2CKD). To validate the proposed approach, situation reasoning experiments were conducted by comparing models with and without M2CKD integration. Experimental results demonstrate that the M2C framework and M2CKD effectively integrate object rules, latent semantics, and causal relationships through the triplet-based structure, enabling explainable and adaptive situation awareness under dynamic driving contexts. The experiments show that the proposed M2C framework and M2CKD enhances the predictive and interpretable reasoning capability of autonomous driving systems.

키워드