자율주행 위험 판단을 위한 증거 기반 설명형 추론 프레임워크(EGRR)

초록

This study proposes a risk reasoning framework designed to strengthen the connection between observed evidence and final conclusions in safety-critical environments such as autonomous driving. The proposed method transforms object information detected in road scenes into structured observational evidence through an Evidence Builder, and then uses this evidence within a Chain-of-Thought (CoT)-based intermediate reasoning process to sequentially generate decision-critical objects and a fact-grounded scene description before deriving the final risk level. In this way, the model does not merely predict a risk category, but also presents the observational basis on which the conclusion is reached. Experimental results show that explicit object-level evidence plays a key role in improving both risk description generation and risk level prediction, while the step-by-step reasoning structure contributes to enhancing the consistency and stability of final risk assessment.

키워드