Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2021, Vol. 47 ›› Issue (1): 239-245,254. doi: 10.19678/j.issn.1000-3428.0057000

• Graphics and Image Processing • Previous Articles     Next Articles

Pedestrian Re-Identification Method Based on Multi-Task Pyramid Overlapping Matching

XU Longzhuang, PENG Li, ZHU Fengzeng   

  1. Engineering Research Center of Internet of Things Technology and Applications, Ministry of Education, College of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Received:2019-12-23 Revised:2020-02-07 Published:2020-02-09

多任务金字塔重叠匹配的行人重识别方法

徐龙壮, 彭力, 朱凤增   

  1. 江南大学 物联网工程学院 物联网技术应用教育部工程研究中心, 江苏 无锡 214122
  • 作者简介:徐龙壮(1994-),男,硕士研究生,主研方向为行人再识别、图像处理;彭力,教授、博士、博士生导师;朱凤增,博士研究生。
  • 基金资助:
    国家自然科学基金(61873112);国家重点研发计划(2018YFD0400902);教育部中国移动科研基金(MCM20170204)。

Abstract: In order to address the low accuracy of existing local-feature-based pedestrian re-identification methods in the cases of pedestrian dislocation and posture changes,this paper proposes a re-identification method based on multi-task pyramid overlapping matching features.In the training stage,the improved ResNes50 is used as the backbone network to extract the feature map,which is segmented and re-combined to form a pyramid overlapping matching network.The global eigenvectors are obtained and multiple local eigenvectors including multi-scale features were got by global average pooling.The Softmax loss function,Triplet loss function and Center loss function are jointly used to learn the global and local eigenvectors,and the Feature Normalization(FN) layer is used to reduce the influence of conflicted learning targets of loss functions.In the reasoning stage,several local eigenvectors are fused into a new eigenvector for similarity matching to obtain better matching results.Experimental results on Market1501,DukeMTMC-reID and CUHK03 datasets show that compared with the mainstream re-identification methods such as PSE and MultiScale,the proposed method has higher re-identification accuracy,and its extracted features have better robustness and recognizability.

Key words: deep learning, pedestrian re-identification, feature fusion, pyramid overlapping matching, multi-task joint learning

摘要: 针对基于局部特征的行人重识别方法在行人错位和姿态变化时识别精度较低的问题,提出一种采用多任务金字塔重叠匹配特征的重识别方法。在训练阶段,使用改进的ResNes50作为主干网络提取特征图,将其切分组合形成金字塔重叠匹配网络,获得全局特征向量并经全局平均池化得到包含多尺度特征的多个局部特征向量,联合使用Softmax损失函数、三元组损失函数和中心损失函数学习全局和局部特征向量,并利用特征归一化层减少损失函数学习目标冲突的影响。在推理阶段,将多个局部特征向量融合为一个新特征向量进行相似性匹配,以获取更好的匹配结果。在Market1501、DukeMTMC-reID和CUHK03数据集上的实验结果表明,与PSE、MultiScale等主流重识别方法相比,该方法重识别精度更高,提取的特征具有较好的鲁棒性和识别度。

关键词: 深度学习, 行人重识别, 特征融合, 金字塔重叠匹配, 多任务联合学习

CLC Number: