基于注意力机制融合特征的车辆目标检测方法  

Vehicle object detection method based on attention mechanism integrated features

作　　者：过鑫炎 朱硕 孙佳豪 梁吉丰 汪宗洋 Guo Xinyan;Zhu Shuo;Sun Jiahao;Liang Jifeng;Wang Zongyang(Nanjing University of Information Science and Technology,Nanjing 210044,China;Wuxi University,Wuxi 214105,China;Wuxi Xiyuan Co.,Ltd.of Technology,Wuxi 214000,China)

机构地区：[1]南京信息工程大学,南京210044 [2]无锡学院,无锡214105 [3]无锡汐沅科技有限公司,无锡214000

出　　处：《电子测量技术》2024年第9期52-60,共9页Electronic Measurement Technology

摘　　要：为了解决道路监控下的车辆目标检测精度低的问题,本文提出一种改进YOLOv7的车辆检测方法。首先引入跨空间学习的高效多尺度注意机制EMA来提高对特征信息的关注;其次将颈部网络中的SPPCSPC模块替换为SPPFCSPC模块,裁剪CBS层,引入EMA注意力机制,以强化对小目标区域的关注,获取更准确的车辆特征;同时,将EMA注意力引入MP模块中,使网络融合更多重要的特征信息;最后,采用MPDIoU损失函数,加快模型收敛速度并提高检测精度。实验结果表明,改进后的YOLOv7检测精度为86.69%,相比原始YOLOv7网络提高了2.83%,可以有效地提升车辆目标检测精度,为道路视频监控等提供保证。To address the issue of low vehicle detection accuracy in road surveillance,this paper proposes an improved vehicle detection method based on YOLOv7.Firstly,we introduce the Efficient Multi-Scale Attention Mechanism(EMA)for cross-space learning to enhance attention to feature information.Secondly,we replace the SPPCSPC module in the neck network with the SPPFCSPC module,trim the CBS layer,and introduce the EMA attention mechanism to strengthen attention to small target areas,thereby obtaining more accurate vehicle features.Additionally,we incorporate the EMA attention into the MP module to fuse more important feature information.Finally,employing the MPDIoU loss function accelerates model convergence and enhances detection accuracy.Experimental results show that the improved YOLOv7 achieves a detection accuracy of 86.69%,which is a 2.83%improvement over the original YOLOv7 network.This enhancement effectively boosts the accuracy of vehicle object detection,providing assurance for applications such as road video surveillance.

关 键 词：车辆检测 YOLOv7 注意力机制 MPDIoU loss 

分 类 号：TN919.8[电子电信—通信与信息系统]