作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (9): 254-260,267. doi: 10.19678/j.issn.1000-3428.0055701

• 图形图像处理 • 上一篇    下一篇

基于注意力机制的狭小空间人群拥挤度分析

张菁, 陈庆奎   

  1. 上海理工大学 光电信息与计算机工程学院, 上海 200093
  • 收稿日期:2019-08-09 修回日期:2019-09-23 发布日期:2019-10-14
  • 作者简介:张菁(1994-),女,硕士,主研方向为图像处理、人工智能、GPU并行计算;陈庆奎,教授、博士、博士生导师。
  • 基金资助:
    国家自然科学基金(61572325,60970012);高等学校博士学科点专项科研博导基金(20113120110008);上海重点科技攻关项目(14511107902,16DZ1203603);上海市工程中心建设项目(GCZX14014);上海智能家居大规模物联共性技术工程中心项目(GCZX14014);上海市一流学科建设项目(XTKX2012);沪江基金研究基地专项(C14001)。

Analysis of Crowd Congestion Degree in Narrow Space Based on Attention Mechanism

ZHANG Jing, CHEN Qingkui   

  1. School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
  • Received:2019-08-09 Revised:2019-09-23 Published:2019-10-14

摘要: 人群拥挤度的分析对维护公共安全极为重要,在空间狭窄的环境下,由于视角受到局限,人与人、人与物品的遮挡十分严重,并且人的尺度不一,密度不均匀,使得传统人群拥挤度监控方法较难直接统计出具体人数。为此,提出一种基于注意力机制的狭小空间人群拥挤度分析方法,旨在量化人群,通过卷积神经网络回归拥挤率分析当前空间内的人群拥挤程度。设计一个注意力模块作为网络的前端,通过生成对应尺度的注意力图区分背景和人群,保留精确的像素点位置信息,以减轻输入图像中各种噪声的影响。在此基础上,将注意图和原始图片通过对应像素点相乘,注入到微调的残差网络中训练得到人群拥挤率。实验结果表明,该方法能够预测出拥挤率,准确反映当前人群拥挤程度,实现人群的流量控制。

关键词: 人群拥挤度, 狭小空间, 注意力机制, 卷积神经网络, 残差网络

Abstract: The analysis of crowd congestion degree is very important to maintain public safety.Generally,in a narrow space the perspective is limited,and the human-human occlusion and human-item occlusion are serious.In addition,because of the different scales of people and uneven density,the traditional methods often fail to directly get the specific number of people in a narrow space.To address the problem,this paper proposes an analysis method of crowd congestion degree in a narrow space based on the attention mechanism in order to quantify the crowd.The method analyzes the congestion degree in the current space through the regression congestion rate of the Convolutional Neural Network(CNN).It designs an attention module as the front end of the network,distinguishes the background and the crowd by generating attention maps of corresponding scales,and retains accurate pixel position information to reduce the impact of various noises in the input image.The attention graph and the original image are multiplied by corresponding pixels and injected into the fine-tuned ResNet to train the crowd congestion rate.Experimental results show that the proposed method can predict the congestion rate,accurately reflect the current crowd congestion degree,and realize crowd flow control.

Key words: crowd congestion degree, narrow space, attention mechanism, Convolutional Neural Network(CNN), Residual Network(ResNet)

中图分类号: