Fusing Driver Perceived and Physical Risk for Safety Critical Scenario Screening in Autonomous Driving

A research team has developed a novel AI method that significantly improves how autonomous vehicle developers identify dangerous driving scenarios in massive datasets. Current approaches rely heavily on manual risk annotation and expensive frame-by-frame analysis, creating bottlenecks in safety testing. The new method, detailed in arXiv paper 2603.20232, introduces a driver risk fusion approach that combines an improved Driver Risk Field with a dynamic cost model during training to generate superior risk supervision signals.

During inference, the system directly predicts scenario-level risk scores through fast forward passes, completely avoiding per-frame computation. This enables efficient large-scale ranking and retrieval of hazardous scenarios from naturalistic driving data. The technical innovations include a new risk height function, speed-adaptive look-ahead mechanism, and a dynamic cost model that integrates kinetic energy, oriented bounding box constraints, and Gaussian kernel diffusion smoothing for more accurate interaction modeling.

The team further designed a risk trajectory cross-attention decoder to jointly decode risk and trajectories, producing smoother and more discriminative risk estimates. When tested on the INTERACTION and FLUID datasets, the method demonstrated substantial improvements over existing approaches. On the FLUID dataset specifically, it achieved an Area Under the Curve (AUC) of 0.792 and an Average Precision (AP) of 0.825, outperforming the previous state-of-the-art PODAR method by 9.1% and 5.1% respectively.