结合全局与局部的人群集体性卷积网络识别方法  

Crowd Collectiveness Recognition Convolutional Neural Network Combining Global and Local Features

作　　者：刘赏[1] 葛顶玉 耿明筱 LIU Shang;GE Dingyu;GENG Mingxiao(School of Science and Technology,Tianjin University of Finance and Economics,Tianjin 300202,China)

机构地区：[1]天津财经大学理工学院,天津300202

出　　处：《信息与控制》2023年第1期75-83,共9页Information and Control

基　　金：天津市自然科学基金(18JCYBJC85100);教育部人文社会科学研究一般项目(19YJA630046)。

摘　　要：人群运动集体性识别对公共场所人群管理具有重要意义。人群运动集体性不仅取决于运动个体,还受到人群局部运动状态的影响。针对以上分析,本文给出了结合局部特征和全局特征的人群集体性卷积网络识别方法。该方法首先基于光流向量构建人群集体性测度图作为卷积网络的输入;然后,在网络第一层卷积后加入通道注意力,获取人群运动的全局信息;并采用空洞卷积提取人群运动的局部信息。最后,本文在公共数据集上进行对比实验,以验证本文方法的有效性。实验结果表明:本文方法在进行人群场景集体性识别时,其加权平均召回率、加权平均准确率和加权平均精准率均优于其它模型。The recognition of collective crowd movement is crucial for crowd management in large public places.Crowd collectiveness depends on the motion of each individual and the motion state of the local crowd overall.Considering the abovementioned factors,in this study,we propose a novel crowd collectiveness recognition convolutional neural network combining global and local features.First,we generate collectiveness measurement images using optical flow vectors.Subsequently,the channel attention mechanism is applied to obtain the global crowd features,and the dilated convolution is used to obtain the local crowd features.Finally,the effectiveness of the proposed method is demonstrated by conducting comparison experiments on the public dataset.The experimental results demonstrate that the effectiveness of our recognition method is superior in the weighted average recall,weighted average accuracy,and weighted average precision.

关 键 词：人群运动 集体性 卷积神经网络 通道注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]