基于YOLOv4-tiny的无人机目标检测算法  

UAV Target Detection Algorithm Based on YOLOv4-tiny

作　　者：王新博 李杰[1] 王岩[1] 姜涛 WANG Xinbo;LI Jie;WANG Yan;JIANG Tao(College of Electronic Information Engineering,Changchun University,Changchun 130022,China;Integrated Intelligent Energy Management Department of Jilin Jidian New Energy Co.LTD)

机构地区：[1]长春大学电子信息工程学院,吉林长春130022 [2]吉林吉电新能源有限公司综合智慧能源管理部

出　　处：《通化师范学院学报》2024年第12期53-59,共7页Journal of Tonghua Normal University

基　　金：吉林省科技厅项目(YDZJ202301ZYTS180)。

摘　　要：无人机作为一种新兴的信息与物质传输工具,具有极高研究价值.受限于无人机硬件条件,嵌入无人机的目标检测算法需要轻量化的模型.在无人机目标检测中,往往存在目标尺度变化大,图像存在相移和目标遮挡的问题,导致检测精度降低.针对无人机目标检测精度低的问题,该文提出了一种基于YOLOv4-tiny的改进算法.该改进算法基于YOLOv4-tiny算法模型,融合了递归特征金字塔以加强特征语义表达,设计了可融合深层特征与浅层特征的特征转换和特征融合模块以增强算法性能,提升算法精确度.经Visdrone数据集训练、测试,mAP值达到了0.146,算法精确度优于其他同级轻量化算法.Unmanned Aerial Vehicle(UAV)is of high research value as an emerging tool for information and material transmission.Limited by its hardware conditions,the target detection algorithm embedded in UAV requires lightweight models.The problems of large variation in target scale,image aberrations and target occlusion when UAV detecting targets,leading to failure in detection.To address the issue of low accuracy in drone target detection,an improved algorithm based on YOLOv4-tiny is proposed in this paper.It is based on the YOLOv4-tiny algorithm model,fusing the recursive feature pyramid to enhance the semantic expression of the features,and designing the feature conversion and feature fusion module,which can fuse the deep and shallow feature,to enhance algorithm performance and improve algorithm accuracy.After training and testing on Visdrone dataset,the mAP value reaches 0.146,which is better than other lightweight algorithms in the same class.

关 键 词：YOLOv4-tiny 递归特征金字塔 目标检测 特征融合 

分 类 号：TP306[自动化与计算机技术—计算机系统结构]