基于改进YOLOv7的复杂行人检测系统研究  

Research on a Complex Pedestrian Detection System Based on Improved YOLOv7

作　　者：孔凡国[1] 王鑫 仇展明 袁功兴 陈靖轩 尹福林 KONG Fan-guo;WANG Xin;QIU Zhan-ming;YUAN Gong-xing;CHEN Jing-xuan;YIN Fu-lin(School of Mechanical and Automation Engineering,Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学机械与自动化工程学院,广东江门529020

出　　处：《机械工程与自动化》2024年第6期58-61,共4页Mechanical Engineering & Automation

摘　　要：在智慧交通系统中,行人的安全和便利性同样是重要关注点。针对行人交通场景中由光照、遮挡、目标小以及背景复杂等因素导致的目标检测精度低、易出现漏检和误检问题等情况,提出了一种基于改进YOLOv7的行人目标检测算法。该算法在主干网络中嵌入动态稀疏注意力机制BiFormer,以增强特征提取能力;将三尺度检测头输出改为双检测头输出,以提升模型训练和推理速度;运用迁移学习初始化权重配值方法,以减少模型训练时间。实验结果表明,相比改进前的YOLOv7算法,所提出的改进算法在人员遮挡区和远距离小目标行人检测时具有更高的识别率,满足了复杂行人检测要求。In the context of pedestrian traffic scenarios within a smart transportation system,the article addresses the challenges posed by factors such as illumination,occlusion,small target sizes,and complex backgrounds,which lead to low precision in pedestrian object detection and the occurrence of missed detections and false positives.The article proposes an improved YOLOv7 based pedestrian object detection algorithm.The algorithm incorporates a dynamic sparse attention mechanism called BiFormer into the backbone network to enhance feature extraction capabilities.It modifies the output of the detection heads from a three-scale output to a dual-scale output,thereby improving model training and inference speeds.Additionally,the algorithm employs a transfer learning strategy to initialize weight assignments,reducing model training time.Experimental results demonstrate that,compared to the original YOLOv7 algorithm,the proposed improved algorithm exhibits higher recognition rates in detecting pedestrians in scenarios involving occlusion or distant small targets,meeting the requirements for complex pedestrian detection.

关 键 词：行人交通 YOLOv7 BiFormer 迁移学习 双检测头 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]