Aerial vehicle detection algorithm based on multi-scale fusion and small-object enhancement

To address the challenges of low detection accuracy, high miss rates, and elevated false-positive rates in vehicle detection from UAV perspectives, this paper proposes a vehicle detection model termed YOLO-VC (YOLO-Vehicle Capture), built upon YOLOv11n. First, a Partial Multi-Scale Feature Aggregation (PMSFA) module is introduced in the backbone network to replace the original C3k2 module, leveraging a differentiated fusion strategy across scales to enhance multi-scale feature extraction capability. Second, PSConv (Pyramid-Structured Convolution) is incorporated into the neck network, whose asymmetric convolutional kernels facilitate the capture of subtle features amid complex background clutter. Third, a small-object feature enhancement pyramid is constructed by fusing P2-level and P3-level features, which are subsequently refined by a Dual-Branch Interactive Enhancement (DBIE) module to boost small-object detection accuracy. Finally, the PIoU V2 (Powerful-IoU Version 2) loss function is adopted to strengthen the focus on medium-quality anchor boxes, thereby improving detection precision in dense scenes. Experimental results on the VisDrone2019 dataset demonstrate that the proposed algorithm achieves a 6.2 percentage-point improvement in mAP@50 over the baseline model, while maintaining a compact model size of only 3.72 MB.