Target-Guided Query Allocation and Boundary-Aware Multi-Scale Feature Fusion for Object Detection in UAV Imagery

Object detection in unmanned aerial vehicle (UAV)-based visual sensing is important for aerial monitoring and intelligent perception. However, UAV imagery often contains small and densely distributed objects, complex backgrounds, and substantial scale variations. Repeated downsampling weakens the boundary and localization cues of small objects, while query initialization based mainly on classification responses may overlook objects with weak semantic features. To address these challenges, we propose TQ-DETR, a target-guided end-to-end detector based on RT-DETRv2. A Target-Oriented Branch (TOB) learns class-agnostic objectness priors to provide reliable foreground spatial guidance independent of category confidence. Based on this prior, Dynamic Query Allocation (DQ) supplements semantic queries with objectness-guided candidates under a fixed query budget, improving candidate coverage for small objects with weak classification responses. In addition, Objectness-Guided Multiscale Boundary Fusion (OG-MBF) selectively extracts and injects target-relevant shallow boundary information while suppressing high-frequency background interference, thereby recovering fine-grained localization cues without adding an extra Transformer feature level. Experiments on VisDrone2019 show that, over three independent runs, TQ-DETR achieved an AP of 30.21 ± 0.51% and an AP50 of 50.28 ± 0.65%, outperforming RT-DETRv2-R18 by 1.79% and 2.75%, respectively, demonstrating improved detection of small and densely distributed UAV targets.