DI-RTDETR: Decoupled Interaction Encoding and Condition-Aware Distillation for Robust UAV Object Detection

Object detection in UAV-captured road scenes is highly susceptible to severe adverse imaging conditions. Existing detectors often suffer from limited robustness and poor cross-scenario generalization when deployed under such condition shifts. This paper addresses multi-condition UAV road-scene object detection by proposing a unified architecture–training framework built upon RT-DETR. On the architectural side, we design a lightweight Decoupled Interaction Encoding Head (DIEH) that explicitly separates semantic representation and spatial localization into two parallel branches, enhancing small-object localization through two-stage cross-branch interaction and intra-branch refinement. On the training side, we introduce a scene-specific teacher distillation strategy, where condition-specialized teachers are trained on clear and foggy subsets, and adaptively selected to guide a unified student model. Extensive experiments on HazyDet2025 and VisDrone2019 demonstrate consistent improvements in detection accuracy, small-object performance, and cross-scenario robustness, while maintaining real-time efficiency.