Current works on trajectory prediction and accident risk assessment for autonomous driving often rely on deep learning models but fail to utilize knowledge graphs (KGs) and a risk semantic framework for enhanced interpretability and accuracy. These models typically lack the structured relationships and dynamic risk assessment capabilities needed for real-time decision-making in complex traffic environments. Furthermore, while Graph Neural Networks (GNNs) are commonly used, they do not fully capture the semantic and relational richness that KGs could provide. In this work, we present a new traffic risk semantic framework, inspired by the nuScenes Knowledge Graph (nSKG) and enriched with available traffic video data. This framework incorporates a traffic risk feature into vehicle nodes, using existing vehicle dynamics such as speed, acceleration, and proximity to other vehicles, alongside road elements like lane geometry. To further enhance risk prediction, we engineer additional features including driver states (e.g., attention, fatigue), environmental conditions (e.g., weather), and human emotions during high-risk situations.