引用本文:郭强,刘全利,王伟.Fast SqueezeNet算法及在地铁人群密度估计上的应用[J].控制理论与应用,2019,36(7):1036~1046.[点击复制]
GUO Qiang,LIU Quan-li,WANG Wei.Fast SqueezeNet Algorithm with Applicationin Crowd Density Estimation[J].Control Theory and Technology,2019,36(7):1036~1046.[点击复制]
Fast SqueezeNet算法及在地铁人群密度估计上的应用
Fast SqueezeNet Algorithm with Applicationin Crowd Density Estimation
摘要点击 2701  全文点击 901  投稿时间：2018-05-22  修订日期：2018-09-21
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2018.80376
  2019,36(7):1036-1046
中文关键词  人群密度估计,SqueezeNet,稀疏化方法,地铁
英文关键词  Crowd  Density, SqueezeNet, Sparse  Techniques, Metro
基金项目  国家自然科学基金(61773085)资助
作者单位E-mail
郭强 大连理工大学控制科学与工程学院 329013203@mail.dlut.edu.cn 
刘全利* 大连理工大学控制科学与工程学院 liuql@dlut.edu.cn 
王伟 大连理工大学控制科学与工程学院  
中文摘要
      针对地铁视频监控一直缺乏一种有效的人群密度估计算法的问题,本文提出了一种人群密度估计算法-Fast SqueezeNet,用于解决在地铁嵌入式计算平台有限的硬件资源限制下,实现对地铁车厢内人群的密度估计。该算法基于轻型卷积神经网络SqueezeNet,结合权值稀疏化和结构稀疏化方法,具有如下3点优势：1)以原始图片作为输入,并在处理的过程中自动提取纹理特征用于拥挤人群密度的估计；2)SqueezeNet经过权值稀疏化后,具有更少的模型参数,可以灵活的布置在ARM等具有有限硬件资源的嵌入式平台上；3)结构稀疏化方法使得SqueezeNet具有更快的在线预测速度,提高其在地铁车厢上的图片处理效率。基于Fast SqueezeNet算法设计了一个三分类的人群密度分类器,在三个人群密度数据集：PETS-2009,Mall,ShangHai metro上,与采用深度卷积神经网络和单纯的权值稀疏化SqueezeNet网络的分类器进行对比实验研究,结果表明：在预测准确率、参数规模和运行时间三个维度上,基于Fast SqueezeNet的分类器均表现出了较好的性能,有效地克服了地铁车厢拥挤人群中存在的高密度、高耦合、透视变形等图像识别难题对估计结果的影响,同时满足了地铁嵌入式平台有限的硬件资源对计算模型在功耗和体积等方面的应用要求。
英文摘要
      In the absence of an effective crowd density estimation in metro video surveillance,the paper proposes an crowd density estimation algorithm-Fast SqueezeNet, which is used to solve crowd density estimation in a subway car, subject to limited hardware resource of subway embedded platform. It is based on Smaller Convolutional Neural Network-SqueezeNet, combining with weights sparsity and structure sparsity, so the proposed method has following three advantages: Firstly, it receives whole crowd image as input and learns texture features for crowd density estimation automatically. Secondly, the SqueezeNet with weights pruning has fewer parameters and can be more flexibility applied on ARM or other embedded platforms, which have limited hardware resource. Last but not least, the structure sparsity enhances online predicting speed of SqueezeNet and heightens image processing efficiency in a subway car. Finally, a three-classifier based on Fast-SqueezeNet for estimating crowd density is proposed, and the three datasets: PETS_2009, Mall and ShangHai metro(SH_METRO) are used to validate this approach. Compared to the state of the art CNNs and only weights sparsity SqueezeNet, the experiment results indicate that the classifier based on the proposed method has better performance at prediction accuracy, number of parameters and running time, thus it can effectively solve the difficulties of image recognizing of crowd people in undergroundScarriages such as high density, severe occlusion and perspective distortion that affect crowd density estimation, moreover, it meets the application demand of subway embedded platform which has limited hardware resource for power consumption and model size.