基于YOLOV5与Jetson TX2的航拍场景目标检测  

Object Detection in Aerial Photography Scene Based on YOLOV5 and Jetson TX2

作　　者：张焕 刘清华 路钊 潘云凡 ZHANG Huan;LIU Qinghua;LU Zhao;PAN Yunfan(Aerospace Science and Industry Intelligent Operations Research and Information Security Research Institute(Wuhan)Co.,Ltd.,Wuhan 430040 China)

机构地区：[1]航天科工智能运筹与信息安全研究院(武汉)有限公司,湖北武汉430040

出　　处：《西华大学学报(自然科学版)》2025年第3期19-28,共10页Journal of Xihua University:Natural Science Edition

基　　金：国防科学技术预先研究基金项目(KO01071)。

摘　　要：基于卷积神经网络的目标检测技术得到快速发展与应用,但受限于检测速度,其在嵌入式平台大规模部署应用始终棘手,在保证模型精度基础上降低时间复杂度成为目标检测技术主要难题之一。为探索某领域基于微型处理器的目标自动检测方法,文章基于YOLOv5、DOTA数据集、Jetson TX2对航拍场景目标检测系统展开研究。首先,基于DOTA高分辨率航拍场景目标检测数据集在PC端完成YOLOv5模型训练,模型的准确率为54.76%,召回率为81.47%,mAP@0.5达到74.12%;其次,对船舶港口、机场、海港3种潜在目标场景进行目标检测分析,在高分辨率航拍场景下仍可以达到较好的检测效果,检测速度达到了181.8 FPS;最后,基于Jetson TX2与无人机设计目标检测系统,实现PC端向微处理器端算法移植,在Jetson TX2上完成模型检测,检测速度达到了16.13 FPS。The target detection technology based on convolutional neural network has been rapidly developed and applied.Limited by the detection speed,its large-scale deployment and application on embedded platforms are always difficult.Breaking through the model time complexity on the basis of ensuring model accuracy has become the main problem of target detection technology.In order to explore the automatic detection method of targets based on microprocessors in the military field,this paper studies the military target detection system in aerial photography scenes based on YOLOv5,DOTA data set,and Jetson TX2.Firstly,the YOLOv5 model training was completed on the PC side based on the DOTA high-resolution aerial scene target detection data set.The accuracy rate of the model was 54.76%,the recall rate was 81.47%,and the mAP@0.5 reached 74.12%;The target detection and analysis of three potential military target scenarios in the seaport can still achieve good detection results in high-resolution aerial photography scenarios and the inference speed reaches 181.8FPS.Finally,a military target detection system based on Jetson TX2 and UAV is designed to achieve The algorithm is transplanted from the PC side to the microprocessor side,and the model inference is completed on the Jetson TX2,and the inference speed reaches 16.13FPS.

关 键 词：目标检测 深度学习 YOLOV5 Jetson TX2 检测速度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]