Abstract: In Unmanned Aerial Vehicles (UAVs) monitoring,small UAVs face significant challenges in target detection in complex environments due to their small size,diverse flight attitudes,and tendency to produce motion blur.To address this,a high-precision detection model for small UVAs,UAV-Net,was proposed.This model integrated four core improvement mechanisms: First,Edge Aware Convolution (EAC) module was designed to enhance the model's ability to extract edge features;Second,the model's neck was optimized and a Multi-scale Feature Focusing (MFF) network was proposed to strengthen feature interaction and fusion; Third,Dynamic Task Decomposition and Alignment (DTDA) detection heads were utilized to optimize the collaboration mode for localization and classification tasks;Finally,the Generalized Intersection over Union (GIoU) loss function was adopted to improve the regression of bounding boxes.Experimental results on the public drone dataset DUT Anti-UAV show that the proposed UAV-Net model achieves improvements of 1.5,3.9,3.1,and 2.6 percentage points over the baseline model in terms of precision,recall,mAP50,and mAP50-95,respectively,thereby validating the effectiveness of UAV-Net as a high-performance anti-drone object detection model for real-world scenarios.Experimental results on the Det-Fly dataset and Visdrone2019 dataset also demonstrate the generalization and applicability of UAV-Net in both air-to-air anti-UAV target detection tasks and small target detection tasks.

Key words: Unmanned aerial vehicle target detection, Anti-UAV technology, Multi-scale features, Small target detection

摘要： 在无人机监测中，小型无人机因体积小、姿态多样且易产生运动模糊，在复杂环境中的目标检测面临巨大挑战。为此，提出一种改进的反无人机目标检测模型UAV-Net。该模型整合了4项核心改进机制：首先，自研设计边缘感知（EAC）模块增强模型对边缘特征的提取能力；其次，优化重构模型颈部提出多尺度特征聚焦（MFF）网络强化特征交互与融合；再次，利用动态任务分解与对齐(DTDA)检测头优化定位和分类任务协作模式；最后，采用GIoU（Generalized Intersection over Union）损失函数改进边界框回归。在公开无人机数据集DUT Anti-UAV上的实验结果表明，UAV-Net在精确率、召回率、mAP50及mAP50-95指标上较基线模型分别提升了1.5、3.9、3.1和3.6个百分点，从而验证了UAV-Net作为面向实际场景的高性能反无人机目标检测模型的有效性。在Det-Fly数据集和Visdrone2019数据集上的实验结果，则验证了UAV-Net在空对空反无人机目标检测任务及小目标检测任务中的泛化性和适用性。

关键词: 无人机目标检测, 反无人机技术, 多尺度特征, 小目标检测