基于城市作战的隐匿小目标识别技术  

Hidden Target Recognition Technology Based on Urban Combat

作　　者：张峰 范会兵 王科举 马洋 张晓曦 范金蕤 ZHANG Feng;FAN Huibing;WANG Keju;MA Yang;ZHANG Xiaoxi;FAN Jinrui(No.208 Research Institute of China Ordnance Industries,Beijing 102202,China;Unit 63936 of PLA,Beijing 102202,China)

机构地区：[1]中国兵器工业第二○八研究所,北京102202 [2]解放军63936部队,北京102202

出　　处：《火力与指挥控制》2023年第3期101-106,共6页Fire Control & Command Control

基　　金：国家自然科学基金(61801441);北京信息科学与技术国家研究中心跨媒体智能专项基金(BNR2019TD01022);“北京市高精尖”学科建设基金;中国传媒大学中央高校基本科研业务费专项基金资助项目(CUC2019B066,CUC18A002-2)。

摘　　要：随着人工智能技术的飞速发展,能够自动识别、锁定和打击目标的智能化武器系统逐渐出现,代替人类执行简单的决策命令,高精度目标识别算法是智能化武器精确打击的前提。目前城市作战越来越受到世界各军事强国的高度重视。城市战场态势瞬息万变,复杂的伪装技术、目标遮挡和恶劣环境条件,给智能目标识别带来严峻的挑战。以当前先进的目标识别模型YOLOv5为基础模型,提出了一种可以多尺度学习空间和通道信息的卷积注意力模块MS-CBAM,允许每个神经元根据输入信息自适应地调整其感受野大小。实验结果表明,在国际公开COCO数据集和自建数据集Long-distance PC Dataset上mAP分别提升了0.5%和2%。训练好的轻量级模型经过TensorRT加速部署在NVIDIA JETSON TX2,实时检测帧为20 ms,满足实时检测要求。该系统也可以作为智能武器系统的一个模块,对自主型武器和无人作战系统具有一定的借鉴意义。With the rapid development of artificial intelligence technology in recent years,the intelligent weapon systems that can automatically identify,lock and strike targets have gradually emerged,and can replace humans in executing the simple decision-making commands.High-precision target recognition algorithms are the prerequisites for intelligent weapons to accurately strike targets.With the acceleration of the global urban process,urban warfare has received more and more attention from each military power in the world.The situation of the urban battlefield is changing rapidly.The complex camouflage technology,target occlusion and harsh environmental conditions have brought severe challenges to intelligent target recognition.In response to the above problems,the current advanced target recognition model YOLOv5 is used as the basic model,and a convolution attention module that can learn spatial and channel information at multiple scales is proposed,allowing each neuron to adaptively adjust its receptive field size according to the input information.The experimental results show that mAP has increased by 0.5%and 2%on the internationally public COCO data set and the self-built Long-distance PC Dataset,respectively.The trained lightweight model is deployed on NVIDIA JETSON TX2 through TensorRT acceleration,and the real-time detection frame is 20 ms,which meets the real-time detection requirements.This system can also be used as a module of the intelligent weapon system,which has certain reference significance for autonomous weapons and unmanned combat systems.

关 键 词：城市作战 NVIDIA JETSON TX2 YOLOv5 智能武器系统 卷积注意力 

分 类 号：E836.7[军事—战术学]