摘要
随着基于深度学习目标检测模型日渐成熟,将目标检测模型部署到航拍无人机上已成为重要研究方向,针对无人机机载推理设备算力和内存有限,提出一种结构重参化的YOLOV5(you only look once V5)航拍目标检测算法。首先替换YOLOV5模型的特征提取网络为结构可重参网络,然后借助开源数据集训练多分支YOLOV5模型,再对多分支网络重参化得到单路网络模型,最后把Pytorch模型转化为ONNX模型,完成在无人机嵌入式端推理部署。实验表明:重参化YOLOV5模型推理速度提高3倍左右,检出率增加0.03%,召回率增加0.02%,mAP0.5增加1.22。
As the deep learning-based target detection model becomes more and more mature,it has become an important research direction to deploy the target detection model to aerial unmanned aerial vehicle(UAV).Aiming at the limited computing power and memory of the onboard reasoning equipment of UAV,a structural reparameterized you only look once V5(YOLOV5)aerial target detection model was proposed.Firstly,the feature extraction network of the YOLOV5 model was replaced as the structurally reconfigurable network.Then,the multi-branch YOLOV5 model was trained with the help of the open source data set.Then,the multi-branch network was reparameterized to obtain the single-path network model.The experiment showed that the reasoning speed of the reparametric YOLOV5 model increased by about 3 times,the detection rate increased by 0.03%,the recall rate increased by 0.02%,and the mAP0.5 increased by 1.22.
作者
倪立
黄征
杨静
NI Li;HUANG Zheng;YANG Jing(United Digital(Hangzhou)Technology Company Limited,Hangzhou Zhejiang 310000,China;Hangzhou Land Survey,Design and Planning Institute Company Limited,Hangzhou Zhejiang 310000,China;Sinohydro Engineering Bureau 8 Company Limited,Changsha Hunan 410000,China)
出处
《北京测绘》
2023年第9期1232-1236,共5页
Beijing Surveying and Mapping
关键词
航拍无人机目标检测
网络模型重参化
YOLOV5模型
aerial photography unmanned aerial vehicle(UAV)target detection
network model reparameterization
you only look once V5(YOLOV5)
作者简介
倪立(1983—),男,浙江杭州人,工程师,从事工程测量、界限与不动产测绘及航空摄影测量工作。E-mail:179985395@qq.com
