Weather severity models domain shift: each sensor modality is trained on mostly clear-weather data, so its detection confidence for the pedestrian ahead degrades along a different curve as conditions worsen. Camera confidence collapses fastest in fog and rain (visible light scatters); LiDAR loses returns in heavy snow (flakes scatter the beam); radar barely notices weather but starts from a lower baseline (coarse resolution).
conf_i(s) = base_i Β· dist_falloff Β· (1 β k_i(weather) Β· s^1.3) + noise
Ο_i = 0.05 + 0.35Β·(1 β conf_i)
w_i = 1 / Ο_iΒ² (inverse-variance weight)
fused = Ξ£ w_iΒ·conf_i / Ξ£ w_i fused Ο = β(1/Ξ£ w_i)
- Uncertainty-weighted fusion lets confident sensors dominate the estimate, so if only camera is blinded by fog the fused confidence barely drops β this is what "sensor fusion" buys you.
- Naive average treats every sensor equally even when one is unreliable, so a single blinded sensor drags the whole estimate down β switch modes to see the gap widen as severity rises.
- Cross-sensor disagreement (standard deviation across enabled modalities) is a practical proxy for out-of-distribution input: sensors that normally agree start to diverge once conditions leave the training distribution.
- OOD alert fires when fused confidence drops below your threshold or disagreement spikes β in a real stack this is exactly the signal that hands control to a conservative fallback controller instead of trusting a shaky detection.
Toggle sensors off to see graceful degradation: two independent, differently-degrading modalities recover far more accuracy under domain shift than any single sensor alone.