AI Generates Realistic Conflict Trajectories for ADAS Evaluation

A new generative model called SevDiff has been introduced, designed to create realistic vehicle interaction trajectories conditioned on a specific conflict severity. This denoising diffusion probabilistic model (DDPM) addresses a significant challenge in ADAS (Advanced Driver-Assistance Systems) evaluation: the scarcity of real-world conflict events, especially rare but high-impact scenarios. Unlike previous generative approaches that condition on scene-level properties, SevDiff uniquely accepts a target Time-to-Collision (TTC) value as a scalar input and is engineered to produce trajectories that match this requested severity within a measurable error. Trained on a dataset of 468 interaction windows from an expressway, SevDiff demonstrated exceptional performance, achieving a 100% hit-rate within +/-0.5 seconds for TTC targets between 0.5 and 1.5 seconds, and maintaining high accuracy up to 2.5 seconds. The generated kinematic features were found to be physically plausible, with minimal out-of-range rates and no negative speed or gap values in the vast majority of samples. This capability allows for the systematic generation of diverse and controlled conflict scenarios, which is invaluable for robustly testing and validating ADAS systems against rare and critical events.