一种改进的无人机小目标检测算法USTD-YOLOv8sOA
An Improved USTD-YOLOv8s Algorithm for UAV Small Target Detection
针对无人机小目标检测(UAV small target detection,USTD)中目标尺寸小、分布不均、易遮挡以及无人机平台硬件存储限制等特点,基于YOLOv8s提出一种适于无人机检测的改进模型USTD-YOLOv8s.将骨干网络引入RFAConv、SCDown、C2FCIB和SimSPPF_PSA模块进行重构,加强对小目标特征提取能力;增加一个小目标检测头(SODH),并进行残差网络连接,专用于从具有更高分辨率的特征图中检测小目标;将颈部中的标准卷积替换为幻影卷积(GhostConv),减少SODH增加的计算成本;用Wise-IoUv3替换CIoU,降低低质量样本对训练影响.将改进模型在VisDrone2019数据集上进行实验验证,实验结果表明:mAP50、mAP50-95值可达45.8%、27.6%,相比YOLOv8s分别提升6.7%、4.2%;模型大小下降3.48 MB,下降16.2%,USTD-YOLOv8s在保证轻量化的同时,有效提升了无人机小目标检测的精度与实用性.
Aiming at the characteristics of small target detection,uneven distribution,easy occlusion,and hardware storage limitations of unmanned aerial vehicle platforms,an improved model USTD-YOLOv8s suitable for unmanned aerial vehicle detection is proposed based on YOLOv8s.Introducing backbone networks into RFAConv,SCDown,C2FCIB,and SimSPPF_PSA modules for reconstruction to enhance feature extraction capabilities for small targets.Add a prediction head small object detection head(SODH)and perform residual network connections specifically for detecting small targets from feature maps with higher resolution.Replace standard Conv in Neck with GhostConv to reduce the computational cost of SODH.Replace CIoU with Wise-IoUv3 version to reduce the impact of low-quality samples on training.The improved model is implemented and validated on the VisDrone dataset,and the results show that the MAP50,MAP50-95 values reached 45.8%and 27.6%,respectively,which are 6.7%and 4.2%higher than YOLOv8s.The model size decreased 3.48 MB,a decrease of 16.2%.While remaining lightweight,USTD-YOLOv8s significantly enhances the precision and applicability of UAV small target detection.
金梅;马维鑫
燕山大学 电气工程学院,河北 秦皇岛 066004燕山大学 电气工程学院,河北 秦皇岛 066004
通用工业技术
计算机视觉无人机检测SODHRFAConvSimSPPF_PSAWise-IoUGhostConv
computer visionunmanned aerial vehicle detectionSODHRFAConvSimSPPF_PSAWise-IoUGhostConv
《计量学报》 2026 (8)
1132-1139,8
国家重点研发计划(2020YFB1711001)
评论