基于改进YOLOv7-tiny的坦克车辆检测方法  

Tank vehicle detection method based on improved YOLOv7-tiny

作　　者：郑陆石 胡晓锋[2] 于伟国 赵东志 张鸿涛 ZHENG Lushi;HU Xiaofeng;YUWeiguo;ZHAO Dongzhi;ZHANG Hongtao(l.School of Mechatronics Engineering,North university of China,Taiyuan 030051,China;Norinco Group Norendar Internationa Ltd.,Shijiazhuang 050011,China;Hua An Industry Group Co.,Ltd.,Qiqihar 161006,China)

机构地区：[1]中北大学机电工程学院,太原030051 [2]北方工程设计研究院有限公司,石家庄050011 [3]北方华安工业集团有限公司,黑龙江齐齐哈尔161006

出　　处：《兵器装备工程学报》2023年第12期285-292,共8页Journal of Ordnance Equipment Engineering

基　　金：山西省技术基础科研项目(JSZL2019408B001)。

摘　　要：针对不同种类无人机航拍高度相差较大、图像分辨率不佳引起的坦克车辆检测算法效果不佳、速度慢等问题,提出一种基于改进YOLOv7-tiny的无人机视角坦克车辆检测算法。首先构建包含568幅图像、2132个目标的坦克车辆数据集。其次对YOLOv7-tiny网络进行3个方面改进:提出了AC-ELAN网络结构并加入3D注意力机制,提高对目标信息的提取能力;引入SPPCSPC结构进一步扩大模型的感受野,同时能够有效减少训练学习时间;将损失函数计算方法替换为WIoU,聚焦于普通质量锚框,加速了模型收敛。最后实验结果表明,改进算法在自建数据集上表现优异,比传统的YOLOv7-tiny平均精度提升5.0%,在GPU设备上检测速度达到71帧/s,能够在无人机计算平台实现实时检测。Aiming at problems such as the ineffective and slow speed of the tank vehicle detection algorithm caused by the large difference in the aerial photography height of UAV,we proposed a tank vehicle detection algorithm from UAV perspective based improved YOLOv7-tiny.In terms of data sets,a tank vehicle dataset containing 568 images and 2132 targets was constructed.In terms of algorithms,an AC-ELAN structure was proposed to enhance image feature recognition and a 3D attention mechanism was incorporated to improve the ability to extract target information;The SPPCSPC structure was introduced to further expand the receptive field,at the same time,it can also effectively reduce the training and learning time;The loss function calculation method was replaced by WIoU,which focuses on the common quality anchor box,and this method accelerates the model convergence.The experimental results show that the improved algorithm in this paper performs well on the self-built dataset,compared with the traditional YOLOv7-tiny,the average precision is increased by 5.0%.The detection speed on the GPU device reaches 71 FPS,the experimental results illustrate that our algorithms can achieve real-time detection on UAV computing platforms.

关 键 词：目标检测 YOLOv7-tiny网络 非对称卷积 3D注意力机制 WIoU损失 

分 类 号：TJ85[兵器科学与技术—武器系统与运用工程] TP7[自动化与计算机技术—检测技术与自动化装置] TP391.4[自动化与计算机技术—控制科学与工程]