Trustworthy open-weather 3D detection via Radar-Guided Dynamic Target Feature Alignment
Year: 2026
Type: article
Abstract: Trustworthy 3D detection in open weather requires a detector to preserve object semantics and localization accuracy when the weather category, intensity, or their combination is absent from training. LiDAR alone is vulnerable to spatially nonuniform attenuation, backscatter, and point sparsification, whereas 4D radar provides complementary range and Doppler evidence. This paper proposes Radar-Guided Dynamic Target Feature Alignment (R-DTFA) to support trustworthy open-weather detection. A short radar-conditioned diffusion process refines weather-displaced LiDAR bird’s-eye-view features while a structural residual limits unintended geometric modification. A frozen normal-weather detector constrains class and box semantics, a frozen weather discriminator reduces weather-dependent distribution shift, and gated fusion controls the contribution of each modality. Trustworthiness is evaluated through leakage-controlled, condition-resolved detection performance under held-out weather evidence rather than through closed-weather aggregate accuracy. On K-Radar, R-DTFA improves mean BEV and 3D AP over L4DR by 3.83 and 7.73 points under unseen weather types and by 1.28 and 2.32 points under unseen severity. Under joint type–severity shift, mean unseen 3D AP exceeds V2X-R and L4DR by 1.76 and 2.18 points, respectively. Loss ablations show that semantic and distribution guidance both improve unseen 3D AP over structural alignment alone. These results provide empirical evidence that radar-guided alignment can improve 3D detection across the evaluated open-weather shifts, while not constituting calibrated uncertainty or a formal safety guarantee.
Institution: Nanjing University of Aeronautics and Astronautics
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,880
OpenAlex ID: W7219017083