基于局部-全局双分支网络的密集人群计数方法  

A Global-Local Dual Branch Network for Congested Crowd Counting

作　　者：邸慧军[1] 宋凌霄 余晓 王蔚然 DI Huijun;SONG Lingxiao;YU Xiao;WANG Weiran(School of Computer Science and Technology,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京理工大学计算机学院,北京100081

出　　处：《北京理工大学学报》2022年第11期1175-1183,共9页Transactions of Beijing Institute of Technology

摘　　要：基于卷积神经网络的人群计数方法促使人群计数精度取得了显著提高.然而,密集人群中的人头尺度变化与复杂环境干扰仍是影响网络计数精度的主要因素.本文提出了一种基于局部-全局双分支网络对密集人群计数.局部分支主要由尺度感知特征提取模块实现,以建模密集人群中人头的尺度变化.全局分支主要由位置感知注意力模块实现,以增强网络对人群与背景之间的判别力.提取到的局部特征与全局特征会送入特征融合分支处理,回归人群密度图.本文方法在3个常用的人群计数数据集与一个遥感目标计数数据集上进行了实验.定量与定性结果表明了本文方法的有效性.Convolutional neural network based crowd counting methods have promoted a significant improvement in the accuracy of crowd counting.However,for congested crowd,huge scale variations of crowd heads and complex scenes still hinder the accuracy of crowd counting.In order to overcome this problem a global-local dual branch network was proposed.The local branch was arranged with the proposed scale-aware feature extraction modules to model the scale changes of the heads in congested crowds.The global branch was arranged with a localization-aware attention module to enhance the network's ability to discriminate between the crowd and the background objects.Then the extracted local features and global features were sent to the feature fusion branch to produce a crowd density map.The proposed method was evaluated on three commonly-used crowd counting datasets and one remote sensing object counting dataset.The quantitative and qualitative results show the effectiveness of the proposed method.

关 键 词：人群密度估计 多尺度学习 视觉注意力 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]