Edge Feature-Enhanced Network for Collision Risk Assessment Using Traffic Scene Graphs

A traffic scene graph effectively models relationships among traffic entities and holds significant importance in enhancing the high-level scene-understanding capabilities of autonomous driving systems. Graph-learning-based methods are widely employed to model structural traffic scene graphs. However, for the collision risk assessment task, the majority of current state-of-the-art graph-learning models overlook essential visual appearance features and do not effectively incorporate the intrinsic graph edge features. In this article, we build a novel edge feature-enhanced collision risk assessment framework that can fully exploit edge features representing the relationships among traffic entities. Specifically, we introduce a node feature-enhanced module to fuse the visual appearance features and implement dynamic traffic scene graph extraction. Next, we design a novel hierarchical node feature update strategy that leverages cross-attention to sufficiently integrate the edge features. Moreover, the temporal transformer encoder is employed to capture the temporal dynamics in driving scenario video. Extensive experiments on the real-world IESG dataset and simulated 1043-carla-sg demonstrate the proposed approach’s superiority, achieving state-of-the-art performance in accuracy (Acc), area under the curve (AUC), and F1-score metrics.