• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 人工智能与数据挖掘 • 上一篇    下一篇

基于深层特征融合的行人重识别方法

熊炜1,2,熊子婕1,杨荻椿1,童磊1,刘敏1,曾春艳1
  

  1. (1.湖北工业大学电气与电子工程学院,湖北 武汉 430068;
    2.美国南卡罗来纳大学计算机科学与工程系,南卡 哥伦比亚 29201)
  • 收稿日期:2019-07-13 修回日期:2019-08-29 出版日期:2020-02-25 发布日期:2020-02-25
  • 基金资助:

    国家自然科学基金(61571182,61601177);湖北省自然科学基金(2019CFB530);国家留学基金(201808420418)

Pedestrian re-identification based on deep feature fusion

XIONG Wei1,2,XIONG Zi-jie1,YANG Di-chun1,TONG Lei1,LIU Min1,ZENG Chun-yan1   

  1. (1.School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;
    2.Department of Computer Science and Engineering, University of South Carolina,Columbia,SC 29201,USA)
  • Received:2019-07-13 Revised:2019-08-29 Online:2020-02-25 Published:2020-02-25

摘要:

针对现有基于深度学习的行人重识别方法对于行人姿态变化、部分遮挡等引起的行人判别特征信息缺失的问题,提出了一种深层特征融合的行人重识别方法。首先,利用卷积层和池化层多次提取网络深层特征,从空间维度提升网络性能,使用融合后的深层特征作为行人图像的全局特征属性;其次,为提高模型的泛化能力,在深层融合特征后加入一个批量归一化层,同时采用标签平滑损失函数和三元组损失函数对模型进行联合训练。实验结果表明,所提的深层特征融合方法具有很好的表达能力。在Market1501、DukeMTMC-reID、CUHK03和MSMT17 4个数据集上对所提方法进行了验证,其中在Market1501数据集上,Rank-1值达到了95.0%,mAP达到了85.6%。
 
 

关键词: 行人重识别, 深层特征融合, Se-resnet50, 批量归一化, 标签平滑损失, 三元组损失

Abstract:

Aiming at the problem that the existing deep learning based pedestrian re-identification methods lack pedestrian discriminative feature information caused by pedestrian posture change and partial occlusion, a pedestrian re-identification network model based on deep feature fusion is proposed. Firstly, the convolution layer and the pooling layer are used to extract the deep features of the network many times to improve the network performance from the spatial dimension. Secondly, in order to improve the generalization ability of the model, a batch normalization layer is added after the deep fusion features. Finally, the label smoothing loss function and the triple loss function are used to train the model. The experimental results show that the proposed deep feature fusion method has good expression ability. The proposed method is validated on four datasets: Market1501, Duke MTMC-reID, CUHK03 and MSMT17. On Market1501, Rank-1 and mAP reach 95.0% and 85.6%.

 

 

 

Key words: pedestrian re-identification, deep feature fusion, Se-resnet50, batch normalization, label smoothing loss, triplet loss