一种改进YOLOv5的区域入侵检测算法  

An Improved Area Intrusion Detection Algorithm for YOLOv5

作　　者：陶晶 吴浩 TAO Jing;WU Hao(School of Automation and Information Engineering,Sichuan University of Light and Chemical Technology Zigong,Sichuan 643000,China;Key Laboratory of Artificial Intelligence,Zigong,Sichuan 643000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川自贡643000 [2]人工智能四川省重点实验室,四川自贡643000

出　　处：《宜宾学院学报》2024年第6期9-14,共6页Journal of Yibin University

基　　金：四川省科技厅项目(2021YFG0313,2022YFS0518,2022ZHCG0035);人工智能四川省重点实验室项目(2019RYY01);四川轻化工大学人才引进项目(2021RC12);四川轻化工大学研究生创新基金资助项目(Y2022116)。

摘　　要：针对现有基于传感器的入侵检测技术误报率高、存在安全隐患等问题,提出一种改进YOLOv5的区域入侵检测算法:以YOLOv5为基础,在Backbone中引入CBAM注意力机制,增强网络对特征的提取能力;在Neck中增加操作层继续对特征图进行上采样处理,并将操作后获取到的特征图与Backbone中第二层的特征图进行Concat融合,以此获取更大的特征图进行小目标检测;结合掩膜法与图像像素坐标系划分警戒区域,对进入警戒区域的可疑目标进行检测,以防止非法入侵的发生.实验结果表明:入侵检测算法mAP值为83.4%,分别较YOLOv5、YOLOX、SSD、Faster-RCNN提高1.8%、17.3%、28.2%、40.6%,检测速度达25.4 frame/s,仅次于YOLOv5,能够满足真实安防场景下对入侵目标的检测需求,且具备良好的泛化能力.Aiming at the problems of existing sensor-based intrusion detection techniques such as high false alarm rate and secu⁃rity risks,an improved area intrusion detection algorithm for YOLOv5 was proposed.Based on YOLOv5,the CBAM attention mechanism was introduced in Backbone to enhance the network’s feature extraction ability;the operation layer was added in Neck to continue the up-sampling processing of feature maps,and the feature maps obtained after the operation were Concat fused with the feature maps of the second layer in Backbone,so as to obtain larger feature maps for small target detection;the mask method was combined with the image pixel coordinate system to delineate the alert area and detect suspicious targets enter⁃ing the alert area in order to prevent illegal intrusion.The experimental results show that the intrusion detection algorithm has a mAP value of 83.4%,which is 1.8%,17.3%,28.2%,and 40.6%higher than YOLOv5,YOLOX,SSD and Faster-RCNN respec⁃tively,and reaches a detection speed of 25.4 frames/s,which is second only to YOLOv5,and is able to satisfy the real security sce⁃narios of intrusion target detection needs in real security scenarios,and has good generalisation capability.

关 键 词：入侵检测 YOLOv5 注意力机制 小目标检测 

分 类 号：TB391.41[一般工业技术—材料科学与工程]