VDSA-YOLOv11: a vision-guided dynamic spatial adaptation network for UAV object detection in complex aerial scenes

Accurate multi-class object detection and localisation in UAV imagery remain challenging due to extreme scale variation, weak visual cues from tiny objects, dense object distributions, occlusion, and background clutter. This paper proposes VDSA-YOLOv11, a global–local adaptive detection framework built upon YOLOv11. The framework integrates OverLoCK dynamic convolutions for context-adaptive local feature extraction, a Vision Transformer for long-range dependency modelling, and a dual-stream attention-enhanced C3K2 module for local–global feature interaction. A BiFPN neck with DySample further supports weighted multi-scale fusion and adaptive upsampling. On the VisDrone2019 validation set, VDSA-YOLOv11 achieves 53.2% mAP50 and 41.4% mAP50:95, exceeding the YOLOv11 baseline by 16.6 and 19.8 percentage points, respectively. Module ablations indicate that ViT and OverLoCK yield the largest accuracy contributions among the evaluated components. Supplementary evaluations on UAVDT and DOTA-HBB show improvements under dataset-specific training and evaluation. The full model contains 123.0 million parameters and achieves 153 FPS on an NVIDIA A100 GPU at an input resolution of 640×640 and a batch size of one. This throughput is hardware-dependent and does not establish onboard real-time performance. All experiments use a single fixed random seed; consequently, the reported improvements reflect the evaluated conditions, while run-to-run variability remains unquantified.