基于无人机图像地面建筑物目标检测算法研究  

Research on the Algorithm for Detecting Objects on the Ground Based on Images by UAV

作　　者：张艳珠 包慧哲 于长海 刘彬 ZHANG Yanzhu;BAO Huizhe;YU Changhai;LIU Bin(Shenyang Ligong University,Shenyang 110159,China;Navy Military 3rd Representative Office in Tianjin,Tianjin 300462,China)

机构地区：[1]沈阳理工大学自动化与电气工程学院,沈阳110159 [2]海装驻天津地区第三军事代表室,天津300462

出　　处：《沈阳理工大学学报》2023年第4期1-6,14,共7页Journal of Shenyang Ligong University

基　　金：国家重点实验室基金项目(2021JCJQLB055006);辽宁省教育厅高等学校基本科研项目(LJKZ0245)。

摘　　要：利用无人机图像基于深度学习对地面建筑物目标进行检测,特别是针对偏僻地区房屋目标的检测,相对人工方式提高了检测效率。应用无人机航拍技术获取视频数据,建立专用的建筑物数据集,提出一种YOLOv5算法的改进模型,融合注意力机制并改进损失函数,添加针对小目标的检测层。实验结果表明,建筑物目标识别的准确率为94.2%,平均精确度均值为96.4%,检测速度为每秒88.9帧。Using images captured by unmanned aerial vehicle to detect building targets on the ground,especially houses targets in remote areas,has a higher detection efficiency than manual methods.In this paper,UAV aerial photography technology is applied to acquire video data,and special building data set is established.An improved version of deep learning-based YOLOv5 algorithm is proposed.Attention mechanism is integrated,loss function is improved,and small target detection layer is added.The experimental results show that the accuracy of building target recognition is 94.2%,the average accuracy is 96.4%,and the detection speed is 88.9 frames per second.

关 键 词：深度学习 目标检测 YOLOv5 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]