为解决无人机“黑飞”造成的安全隐患，针对现有的基于深度学习的无人机目标检测算法模型参数量大、训练耗时长的问题，提出了一种基于YOLOv5算法的轻量级无人机实时目标检测（ED-YOLOv5s）算法。首先结合轻量化模型（efficient model, EMO）对YOLOv5骨干网络的特征提取部分进行重构；其次引入归一化高斯Wasserstein距离（normalized Gaussian Wasserstein distance, NWD）和比例因子计算候选框之间的相似度来部分替代IoU（intersectiom over union）；然后引入无参注意力机制SimAM以优化权重分布，提升检测精确度，以达到对YOLOv5网络的Backbone、Head进行优化的效果；最终，得到模型大小为4.57 M，浮点运算量为26.1 GFLOPs的ED-YOLOv5s轻量级无人机检测算法。实验数据表明，改进后的算法提高了检测精度，实现了模型轻量化，所提算法在DUT Anti-UAV数据集上AP@50值达到96.7%，在RTX2060显卡上检测速度达到37 frame/s。

In order to solve the security risks caused by UAV "black flight", aiming at the problems of large model parameters and long training time of the existing UAV target detection algorithm based on deep learning, a lightweight real-time object detection algorithm for UAV based on YOLOv5 (ED-YOLOv5s) was proposed. Firstly, the feature extraction part of YOLOv5 backbone network was reconstructed by combining the lightweight model efficient model (EMO). Secondly, normalized gaussian wasserstein distance (NWD) and scale factor were introduced to calculate the similarity between candidate boxes to partially replace IoU. Then, SimAM was introduced to optimize the weight distribution and improved the detection accuracy, so as to achieved the effect of optimizing the Backbone and Head of YOLOv5 network. Finally, the model size is 4.57 M. ED-YOLOv5s lightweight UAV detection algorithm with floating point computation of 26.1 (GFLOPs). The experimental data shows that the improved algorithm improves the detection accuracy and realizes the model lightweight. The proposed algorithm achieves 96.7% AP@50 value on the DUT Anti-UAV dataset, and the detection speed reaches 37 frame/s on the RTX2060 GPU.