基于YOLOv5的无人机视角小目标检测算法

UAV Small Target Detection Algorithm Based on YOLOv5

针对无人机视角下的小目标检测精度较差、漏检较为严重的问题,提出一种基于改进YOLOv5的无人机图像检测算法。针对小目标尺度较小问题在骨干网络替换空间金字塔池化(Spatial Pyramid Pooling,SPP)为SPPCSPC-GS,增强密集区域关注能力,提取更多小目标有效特征;在颈部网络中引入CBAM注意力机制将头部C3模块替换为C3CBAM增强上下文信息,提高空间与通道特征表达能力;针对遮挡问题引入柔性非极大值抑制(Soft Non Maximum Suppression,Soft NMS)提升模型对遮挡和密集目标的检测能力;替换损失函数为EIOU加快收敛提升定位效果。改进后的模型在VisDrone数据集上平均检测精度为42.2%,相较于原始YOLOv5s算法提升10.7%,遮挡严重的小目标行人与人类别精度分别上升12%与13.3%。相较于其他先进算法,所提算法表现优秀,可以满足无人机视角图像检测任务要求。

Aiming at the problems of poor detection accuracy and serious missed detection of small targets from the perspective of UAV,a UAV image detection algorithm based on improved YOLOv5 is proposed.Aiming at the problem of small target scale,Spatial Pyramid Pooling(SPP)is replaced by SPPCSPC-GS in the backbone network to enhance the attention ability of dense areas and extract more effective features of small targets.The CBAM attention mechanism is introduced into the neck network to replace the head C3 module with C3CBAM to enhance the context information and improve the spatial and channel feature expression ability.Aiming at the occlusion problem,soft non maximum suppression(Soft NMS) is introduced to improve the detection ability of the model for occlusion and dense targets.The loss function is replaced with EIOU to accelerate convergence and improve positioning effect.The improved model has an average detection accuracy of 42.2% on the VisDrone dataset,which is 10.7% higher than the original YOLOv5s algorithm.The accuracy of small target pedestrians and people with severe occlusion increases by 12% and 13.3%,respectively.Compared with other advanced algorithms,the proposed algorithm performs well and can meet the requirements of UAV perspective image detection tasks.