FAD-DETR: A Frequency-Adaptive and Haze-Aware Detection Transformer for Autonomous Driving in Adverse Weather

Robust perception in adverse meteorological conditions remains a critical challenge for autonomous driving systems. Generic object detectors often suffer from feature degradation and domain shifts in foggy environments due to the loss of high-frequency details and contrast attenuation. To address these limitations, we propose the Frequency-Adaptive and Haze-Aware Detection Transformer (FAD-DETR). Building upon the Real-Time Detection Transformer (RT-DETR), an atmospheric scattering suppression module recalibrates the feature distribution in the latent space based on physical scattering priors to mitigate haze interference. Subsequently, an edge-guided enhancement module utilizes Laplacian operators to enhance structural boundaries blurred by fog. Finally, a frequency-adaptive context aggregation module is applied to reduce spectral aliasing during cross-scale feature fusion. Evaluations on the real-world RTTS benchmark demonstrate that FAD-DETR achieves a favorable performance, improving the mAP@50 by 4.1% compared to the RT-DETR-L baseline. Furthermore, zero-shot experiments on the synthetic Foggy Cityscapes dataset indicate the generalization capability of the model without additional fine-tuning. This framework provides a practical solution to visual perception tasks in complex weather.