New Framework Enhances Autonomous Driving Risk Assessment.

Autonomous driving systems critically depend on their ability to accurately perceive and anticipate risks in complex, safety-critical environments. Current research has advanced collision prediction, but quantifying risk levels precisely from single camera inputs remains a significant challenge due to the intricate dynamics of multiple interacting agents and inherent real-world uncertainties. To address these issues, a new framework called NSF-HRPT has been developed. This approach integrates learning-based perception with structured reasoning for quantitative risk assessment. It features a Neural Semantic Field (NSF) trained on simulation data to model scene semantics, predict trajectories, and estimate probabilistic Time-to-Collision (TTC) distributions. During operation, the pre-trained NSF acts as a foundational prior for the Hierarchical Risk Perception Tree (HRPT), which facilitates efficient parallel computation and spatial reasoning about multi-agent risks. The framework also includes a Sim2Real enhancement strategy that leverages foundation models to improve real-world performance without requiring additional training. Evaluations show state-of-the-art results on synthetic benchmarks and competitive performance on real-world datasets for both TTC accuracy and risk localization.