UAV Small Target Detection Algorithm Integrating Dynamic Features

To address the issues of missed and false detections caused by tiny target scales, complex background environments, and extremely uneven target distribution in dense scenes in UAV aerial images, we propose a lightweight small target detection algorithm based on an improved YOLOv11. A light-weight dynamic feature fusion module (LDFM) is designed, which effectively enhances the feature saliency of tiny targets in complex backgrounds through parallel large-kernel depthwise convolutions and geometric feature extraction branches combined with an adaptive gating mechanism. Secondly, a learnable density-adaptive task alignment (LDA-TAL) strategy is proposed, which dynamically adjusts the number of positive samples by estimating grid density, significantly alleviating the problem of insufficient sample allocation in dense occluded scenarios. Finally, a quality-balanced IoU loss (QB-IoU) is introduced, which reweights regression loss according to the matching quality of predicted boxes, improving the localization accuracy of low-quality difficult samples. Experimental results on the VisDrone2019 dataset show that the algorithm achieves an mAP@0.5 of 46.1%, an increase of 11.3% compared with the baseline model, and, while maintaining a lightweight advantage with only 2.87\times 10^6 parameters, the detection accuracy is significantly better than that of existing mainstream detection algorithms.