WFG-RT-DETR: A Highly Robust Object Detection Model for Autonomous Vehicles in Adverse Weather Conditions

Object detection under adverse weather conditions remains a challenging problem for autonomous driving systems due to image degradation caused by rain, snow, haze, and complex illumination. Although RT-DETR achieves a good balance between detection accuracy and efficiency, its performance is limited in adverse weather scenarios due to weather-induced feature distortion and unreliable query selection. In this paper, we propose WFG-RT-DETR, a robust real-time object detection framework for adverse weather conditions. A Weather-Focused Guidance (WFG) module is introduced to enhance feature representation by combining frequency-domain noise suppression with spatial structural compensation. Furthermore, a Weather-Aware Query Selection (WAQS) module is proposed to improve query initialization by incorporating weather-aware noise estimation, reducing false-positive proposals caused by environmental interference. An Exponential Moving Average (EMA) strategy is also employed to stabilize model training. Extensive experiments on the DAWN and WEDGE datasets demonstrate the effectiveness of the proposed method. Compared with RT-DETR, WFG-RT-DETR improves the mAP@50 on the DAWN dataset from 64.62% to 68.89% while maintaining real-time inference capability. Cross-dataset evaluations on the WEDGE dataset further verify its robustness and generalization under diverse adverse weather conditions. The proposed framework provides an effective solution for reliable object detection in intelligent transportation and autonomous driving applications.