某大口径自行加榴炮末敏弹识别模型轻量化研究  

Research on a Lightweight Recognition Model for Protecting Large Caliber Self-propelled Howitzers from Sensor-Fuzed Munitions

作　　者：冯颖龙 唐旭 许耀峰 王振明[1] 祁万龙 刘爱峰[1] 刘建成 FENG Yinglong;TANG Xu;XU Yaofeng;WANG Zhenming;QI Wanlong;LIU Aifeng;LIU Jiancheng(Northwest Institute of Mechanical&Engineering,Xianyang 712099,Shaanxi,China)

机构地区：[1]西北机电工程研究所,陕西咸阳712099

出　　处：《火炮发射与控制学报》2023年第3期55-60,85,共7页Journal of Gun Launch & Control

摘　　要：末敏弹对地面装甲目标的杀伤威力和命中率极高。有效反末敏弹对降低武器装备战场损失非常重要,识别末敏弹是地面装甲反末敏弹的降本增效的一个重要途径。针对某大口径自行加榴炮面对末敏弹的攻击反应速度慢、传统目标检测算法模型较大、应用于嵌入式硬件平台实时性不高的问题,提出了基于MobileNetV3-small的轻量化神经网络模型,通过标注图片的方法构建伞形目标数据集,优化训练方法,将模型部署于搭载高通骁龙835的嵌入式硬件。实验表明,相较于Yolo-v4-tiny,该模型推理速度提高了7.3帧/s,内存开销降低了716.8 MB,具有较好的实时性,可以在战场中辅助某大口径自行加榴炮及时发现末敏弹的攻击并做出有效反应,提高其战场生存能力。Sensor-fuzed munitions(SFM)have extremely high killing power and hit rate on armored ground targets.Effective countermeasures against SFM are of critical importance to reduce weaponry losses,where identification of SFM is considered an effective way for ground armor to counter them.The problem of large caliber self-propelled howitzers in the face of SFM attacks is slow reaction time because traditional target detection algorithms are large and have low real-time performance when applied to embedded hardware platforms.Therefore,a lightweight neural network model based on MobileNet V3-small was proposed.An umbrella target data set was first built by labeling images,then the model was optimized and deployed on embedded hardware equipped with Qualcomm Snapdragon 835.Experiments showed that the proposed model inference was 7.3 frames/second faster,with 716.8 MB less memory usage and better real-time performance compared to Yolo-v4-tiny.It can be used to assist large caliber self-propelled howitzers on the battlefield to detect SFM attacks and react effectively,which significantly improves their battlefield survivability.

关 键 词：末敏弹识别 轻量化神经网络 嵌入式硬件平台 战场生存能力 

分 类 号：TJ013[兵器科学与技术—兵器发射理论与技术]