FogBEV: Degradation-Aware Bird's-Eye-View Fusion for Autonomous Driving in Dense Fog

Abstract

Dense fog is a persistent failure mode for autonomous-driving perception because optical contrast is reduced by atmospheric scattering while range sensors may suffer from sparsification, backscatter and unstable returns. This paper presents FogBEV, a degradation-aware bird's-eye-view fusion framework that combines gated/NIR image evidence with LiDAR evidence in a shared BEV representation. Unlike fixed camera-LiDAR fusion, FogBEV estimates per-modality reliability from feature contrast, point-density and false-return cues, then uses these reliabilities to adjust fusion weights before BEV detection. The design is motivated by the strong gated vision and LiDAR integration framework, but moves the adaptation from image/point features into a BEV-centric formulation that is directly compatible with modern planning stacks. To avoid fabricated results, this draft reports only a lightweight synthetic BEV fog benchmark generated by the accompanying code. The benchmark simulates three meteorological optical range levels, 200 m, 80 m and 40 m, and evaluates BEV occupancy average precision over three random seeds. FogBEV improves dense-fog BEV AP from 0.471 for fixed fusion and 0.561 for magnitude-based fusion to 0.670, while shifting the mean LiDAR weight below 0.01 in dense fog. The results support the central claim that degradation-aware weighting can prevent corrupted range evidence from suppressing visible gated structures. The paper also clearly states the remaining need for full validation on real adverse-weather datasets such as SeeingThroughFog/DENSE before any safety deployment claim is made.