摘要
机坪特种车辆作为航班保障服务的重要一环,其种类多,形状各异;目前已有的车辆检测算法在识别机坪特种车辆时检测精度不高,在遮挡时无法检测;针对于此问题,提出了一种基于改进YOLOv5s的机坪特种车辆检测算法;为了在机坪特种车辆检测中快速、准确地定位感兴趣区域,在主干网络中融合协同注意力机制;考虑到机坪监控场景下特种车辆尺度差别较大的情况,为了能够增强对不同尺度特种车辆的检测能力,提出了四尺度特征检测网络结构;为了提高检测网络多尺度特征融合能力,结合加权双向特征金字塔结构对网络的Neck部分进行改进;将改进后的算法在自建的机坪特种车辆数据集上进行训练、测试,实验结果表明,与YOLOv5s相比,改进后算法的精确度提升了1.6%,召回率提升了3.5%,平均精度mAP0.5和mAP0.5:0.95分别有2.3%和3.3%的提升。
Apron special vehicles are taken as an important part of flight guarantee service,which have the features of various types and different shapes.existing vehicle detection algorithms have the low detection accuracy of identifying special vehicles on the apron and cannot detect when obscured.Aiming at this problem,a special vehicle detection algorithm based on improved YOLOv5s is proposed.To locate the region of interest quickly and accurately in the detection of special vehicles on the apron,a coordinate attention mechanism is integrated into the backbone network.Considering that the scale of special vehicles varies greatly in the apron monitoring scene,a four-scale feature detection network structure is proposed to enhance the detection ability of special vehicles with different scales.To improve the multi-scale feature fusion capability of the detection network,the neck part of the network is improved by combining with the weighted bidirectional feature pyramid structure.The improved algorithm is trained and tested on the special vehicle dataset of the self-built apron.The experimental results show that compared with YOLOv5s,the precision of the proposed algorithm is improved by 1.6%,the recall by 3.5%,and the average precision mAP0.5 and mAP0.5:0.95 by 2.3%and 3.3%,respectively.
作者
诸葛晶昌
李想
ZHUGE Jingchang;LI Xiang(College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)
出处
《计算机测量与控制》
2023年第6期27-33,39,共8页
Computer Measurement &Control
基金
国家重点研发专项(2018YFB1601200)
中国民航大学中央高校基本科研业务费专项资金(3122019047)。
关键词
机坪特种车辆
协同注意力机制
四尺度特征检测
加权双向特征金字塔
特征融合
special vehicles on the apron
coordinate attention mechanism
four-scale feature detection
weighted bidirectional feature pyramid
feature fusion