基于多尺度注意力网络的密集人群计数  

MULTI-SCALE ATTENTION NETWORK FOR HIGHLY CONGESTED CROWD COUNTING

作　　者：文帅 蒋勇[1] 杨丹 马金刚 杨闻宇 Wen Shuai;Jiang Yong;Yang Dan;Ma Jingang;Yang Wenyu(College of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621000,Sichuan,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621000

出　　处：《计算机应用与软件》2025年第1期130-136,157,共8页Computer Applications and Software

摘　　要：针对拥挤场景下的尺度变化导致人群计数任务中精度较低的问题,提出一种基于多尺度注意力网络(MANet)的密集人群计数模型。通过构建多列模型以捕获多尺度特征,促进尺度信息融合;使用双注意力模块获取上下文依赖关系,增强多尺度特征图的信息;采用密集连接重用多尺度特征图,生成高质量的密度图,之后对密度图积分得到计数。此外,提出一种新的损失函数,直接使用点注释图进行训练,以减少由高斯滤波生成新的密度图而带来的额外的误差。在公开人群数据集ShanghaiTech Part A/B、UCF-CC-50、UCF-QNRF上的实验结果均达到了最优,表明该网络可以有效处理拥挤场景下的目标多尺度,并且生成高质量的密度图。Aimed at the problem of the poor performance in crowd counting tasks caused by scale various in highly congested scenes,a dense crowd counting model based on multi-scale attention network(MANet)is proposed.A multi-column convolutional neural network was constructed to capture multi scale features and to promote scale information fusion.A dual attention module was adopted to obtain contextual information and enhance the performance of multi scale feature.Dense connection was used to reuse multi scale feature maps,and generate high-quality density maps,and the density maps were integrated to count.A new loss function was proposed,which directly used the dot annotation map for training to reduce the additional error caused by the Gaussian filtering to smooth the dot annotation.The best results on the public datasets(ShanghaiTech Part A/B,UCF-CC-50,UCF-QNRF)demonstrate that our model can effectively handle multi-scale various in highly congested scenes and generate high-quality density maps.

关 键 词：密集人群计数 多尺度 卷积神经网络 注意力机制 密集连接 损失函数 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]