DRIFT Model Cuts Occlusion Response Latency for Safer Self-Driving

A team of researchers led by Zian Wang has introduced DRIFT (Driving Risk Inference via Field Transmission), a novel approach to modeling risk in autonomous driving environments. Unlike traditional scalar safety metrics or static risk fields, DRIFT is a spatiotemporal risk field governed by an advection-diffusion-reaction partial differential equation—with an optional telegrapher term that captures wave-like propagation of risk. The model draws on three key sources: anisotropic Gaussian kernels to account for velocity-induced risk, latent hazard detection behind large vehicles that occlude sensors, and topology-coupled conflict pressure at merge zones. This allows the system to anticipate dangers that are not immediately visible, mimicking human driving intuition.

To evaluate DRIFT, the authors introduced four field-centric metrics: Lane-Change Risk Differential, Temporal Anticipation Index, Occlusion Sensitivity Index, and Occlusion Response Latency—complementary to existing Surrogate Safety Measures. In experiments on real-world traffic datasets, DRIFT notably reduced occlusion response latency and lowered the near-collision rate compared to selected baseline models in synthetic occlusion scenarios. Accepted at the IEEE International Conference on Intelligent Transportation Systems (ITSC) 2026, DRIFT represents a significant step toward human-like risk perception in autonomous vehicles, enabling safer navigation in complex, occluded environments.