Abstract: To solve the problems that small targets are difficult to detect, dense targets and complex environment lead to the increase of missed detection probability in UAV perspective, an improved YOLOv7-tiny UAV target detection algorithm is proposed. Firstly, a parallel network is added on the basis of the backbone network to enhance the capability of extracting feature map information. Secondly, the sampling scale of small targets is increased and the FPN structure is improved, so that the feature map output of the backbone network can be used for subsequent up-sampling and down-sampling, and the network accuracy is improved. Then, coordinate attention (CA) is added to optimize the output feature map of backbone network and reduce the loss of feature information. Finally, WIoU loss function is used to calculate location loss, which enhances the detection ability of small targets. Experimental results show that compared with the original algorithm, improved YOLOv7-tiny algorithm accuracy and recall rate increased by 2.8 and 2.7 percentage points respectively, mAP@0.5 and mAP@0.5:0.95 increased by 3.8 and 3.2 percentage points respectively, effectively improve the detection accuracy of the algorithm.

Key words: UAV, YOLOv7-tiny, target detection, coordinate attention (CA), loss function

摘要： 针对无人机视角下小目标难以检测、目标密集和环境复杂导致漏检概率增加的问题，提出一种改进YOLOv7-tiny的无人机目标检测算法。在原主干网络的基础上增加一个并行网络，加强主干网络对特征图信息的提取能力；增加细小目标采样尺度并改进FPN结构，使主干网络输出的特征图可以用于后续上采样和下采样当中，提高网络精度；加入CA注意力机制，优化主干网络输出特征图，减少特征信息损失；使用WIoU损失函数计算定位损失，增强网络对小目标的检测能力。实验结果表明，相较于原算法，改进YOLOv7-tiny算法的准确率和召回率分别提升了2.8和2.7个百分点，mAP@0.5和mAP@0.5:0.95分别提升了3.8和3.2个百分点，有效提高了算法的检测精度。

关键词: 无人机, YOLOv7-tiny, 目标检测, CA注意力机制, 损失函数