Traffic Incident Reasoning Through Multi-Grained Conceptual Graph Networks

The paper proposes a novel Conceptual Graph Reasoning Network (CGR) for analyzing traffic events in videos, which is crucial for autonomous vehicles, assisted driving, and intelligent transportation systems (ITS). CGR focuses on Visual Question Answering (VQA) at the event level by integrating spatial-temporal and linguistic representations using graph convolution networks. It includes a Diversity-aware Conceptual Reasoning (DCR) module that separates misleading correlations and dynamically adjusts the importance of different features. The CGR’s effectiveness is demonstrated using the SUTD-TrafficQA dataset, showing significant performance improvements through ablation studies.