VD-DETR: a small UAVs detection method based on visible light and DQ-DETR

Abstract Unauthorized unmanned aerial vehicles (UAVs) incursions into restricted airspace have emerged as a growing security threat. However, detecting these UAVs in visible-light imagery remains difficult, as their small size and inconspicuous appearance are easily obscured by cluttered backgrounds and variable environments. Current methods frequently face limitations in detection accuracy and robustness across complex scenarios. To overcome these challenges, we propose a small UAVs detection method based on visible-light and DQ-DETR (VD-DETR). To mitigate the loss of semantic cues across scales, we propose stacked feature aggregation module that preserves fine structural details of small UAVs that are otherwise easily suppressed by channel compression. To improve localization under varying illumination, we propose optimized coordinate attention that strengthens long-range dependencies through decomposed pooling and dynamic convolution, enhancing boundary perception when the outlines of small UAVs become blurred. To handle crowded airspace and distractor interference, we propose anchor-centered noise mechanism to reduce overlap among adjacent small UAVs and a distance-aware loss that enhances query-instance alignment in dense small object scenarios by jointly considering confidence and spatial consistency. Evaluations on the Ac-UAVs and Det-Fly datasets demonstrate that VD-DETR achieves AP scores of 86.8% and 67.6%, surpassing the baseline DQ-DETR by 1.4% and 0.7%. Comparative experiments against state-of-the-art detectors, including Deformable-DETR, DINO, and RT-DETR, further highlight the superiority of our method in accuracy and robustness. These results confirm the potential of VD-DETR as a solution for UAVs surveillance in low-altitude airspace and public safety applications.