Learning from the Unseen: Generative Data Augmentation for Geometric-Semantic Accident Anticipation

Researchers from multiple Chinese institutions have published a new approach to traffic accident anticipation for autonomous vehicles, a critical yet unsolved problem. Their dual-path framework tackles two core challenges: the lack of diverse, large-scale accident datasets and the difficulty of modeling complex interactions between road users. On the data side, they employ a controlled video synthesis pipeline that, guided by structured prompts, derives feature distributions from existing corpus and generates high-fidelity synthetic driving scenes that mirror real-world statistical patterns. This generative augmentation vastly expands training data without costly real-world collection.

For reasoning, the framework uses a graph neural network enriched with semantic cues. It dynamically models both spatial proximity and semantic relationships (e.g., vehicle type, pedestrian intent) among traffic participants, enabling earlier and more accurate anticipation of potential collisions. To validate their method, the team introduces a new benchmark dataset containing standardized, finely annotated video sequences across a broad spectrum of regions, weather conditions, and traffic densities. Experiments on existing datasets and their new benchmark demonstrate notable improvements in both prediction accuracy and anticipation lead time, highlighting the potential of generative data augmentation to enhance autonomous driving safety.