Graph Neural Reasoning for Semantic Understanding of Traffic Signs in Intelligent Transportation Systems

In the domain of intelligent transportation systems, interpreting informative traffic signs requires more than recognizing isolated objects. Our study is dedicated to Traffic Sign Semantic Understanding (TSSU), with the goal of examining diverse traffic signs and producing structured semantic descriptions. To accomplish this, we propose an integrated hybrid framework encompassing element identification, content inference, and semantic description generation. The first module detects text, symbols, and arrowheads. The second module uses an Edge-weighted Graph Attention Mechanism (EGAT) to jointly refine node and edge representations, infer component relations, and incorporate local and global layout information. The final module organizes the inferred content into relation groups and generates key–value descriptions using attributes and a dynamic prediction transformer. Experiments on the 5000-image CASIA–Tencent CTSU dataset attain a component F1 of 0.9676, a detection-conditioned relation F1 of 0.9198, and Information Matching F1 of 0.8016. Generation-input ablations show sensitivity to relational information, semantic attributes, and layout input. The framework links visual components and their relations to explicit, structured traffic information.