多模态深度学习提升行人再识别性能

需积分: 9 7 浏览量更新于2024-09-04 收藏 9.09MB PDF 举报

该论文"Omnidirectional Feature Learning for Person Re-Identification"发表于2019年的IEEE Access，针对行人再识别（Person Re-Identification, Re-ID）这一领域提出了创新性的深度学习方法。作者们构建了一个包含五个分支的多分支深度模型，旨在从多个角度和层次提升特征学习的效率和性能。首先，模型的两个分支专注于水平和垂直方向的学习，这样能够捕捉到行人图像在空间布局中的关键信息，增强对个体身体结构的区分能力。这种设计有助于处理由于视角变化导致的行人图像变形问题，提高了特征的鲁棒性。其次，第三个分支关注不同特征通道之间的相关性，这表明模型不仅关注局部细节，还考虑了全局特征的整体协调性。通过分析和整合这些特征通道，模型能够更好地提取到与身份识别相关的整体特征模式。接着，剩下的两个分支分别负责识别任务和三元组损失优化。识别分支用于学习具有高度鉴别力的全局特征，这些特征对于确定一个人的身份至关重要。而三元组子网络则负责优化度量学习，确保相似的人被紧密聚类，不同的个体之间保持适当的距离，从而提高识别的准确性和召回率。论文的实验部分在大型基准数据集Market-1501、CUHK03和DukeMTMC-reID上进行了广泛比较，结果证明了所提深度框架在行人再识别任务中展现出显著的优势，相较于当时的先进方法，它在精度和效率上都有所提升。研究者还得到了国家自然科学基金等项目的资助，这表明该工作受到了学术界的认可和支持。这篇论文的核心贡献在于提出了一种全面的深度学习框架，它通过多维度的特征学习策略，有效地解决了行人再识别中的挑战，展示了在复杂场景下提高行人身份匹配性能的潜力。这对于推进计算机视觉领域的行人识别技术发展具有重要意义。

Received January 27, 2019, accepted February 21, 2019, date of publication February 26, 2019, date of current version March 18, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2901764

Omnidirectional Feature Learning for

Person Re-Identification

DI WU

, HONG-WEI YANG

, DE-SHUANG HUANG

, (Senior Member, IEEE),

CHANG-AN YUAN

, XIAO QIN

, YANG ZHAO

, XIN-YONG ZHAO

AND JIAN-HONG SUN

School of Electronics and Information Engineering, Institute of Machine Learning and Systems Biology, Tongji University, Shanghai 201804, China

Science Computing and Intelligent Information Processing of Guangxi Higher Education Key Laboratory,

Guangxi Teachers Education University, Nanning 530001, China

Beijing E-Hualu Information Technology Co., Ltd., Beijing 100043, China

Corresponding author: De-Shuang Huang (dshuang@tongji.edu.cn)

This work was supported in part by the National Science Foundation of China under Grant 61520106006, Grant 61732012, Grant

61861146002, Grant 61772370, Grant 61702371, Grant 61672203, Grant 61572447, Grant 61772357, and Grant 61672382, in part by the

China Postdoctoral Science Foundation under Grant 2017M611619, and in part by the BAGUI Scholar Program of Guangxi Province

of China.

ABSTRACT Person re-identiﬁcation (PReID) has received increasing attention due to it being an important

role in intelligent surveillance. Many state-of-the-art PReID methods are part-based deep models. Most

of these models focus on learning the part feature representation of a person’s body from the horizontal

direction. However, the feature representation of the body from the vertical direction is usually ignored.

In addition, the relationships between these part features and different feature channels are not considered.

In this paper, we introduce a multi-branch deep model for PReID. Speciﬁcally, the model consists of ﬁve

branches. Among the ﬁve branches, two branches learn the part features with spatial information from

horizontal and vertical orientations; one branch aims to learn the interdependencies between different feature

channels generated by the last convolution layer of the backbone network; the remaining two branches

are identiﬁcation and triplet sub-networks in which the discriminative global feature and a corresponding

measurement can be learned simultaneously. All ﬁve branches can improve the quality of representation

learning. We conduct extensive comparison experiments on three benchmarks, including Market-1501,

CUHK03, and DukeMTMC-reID. The proposed deep framework outperforms other competitive

state-of-the-art methods. The code is available at https://github.com/caojunying/person-reidentiﬁcation.

INDEX TERMS Person re-identiﬁcation, deep learning, part feature, triplet model, identiﬁcation model.

I. INTRODUCTION

As a fundamental task of intelligent surveillance, person

re-identiﬁcation (PReID) aims to re-identify a speciﬁc person

from multiple camera views. It has been of considerable inter-

est to the computer vision community in recent years. Great

progress has been made in PReID, however, the visual appear-

ance of a person may undergo signiﬁcant variations when

facing unpredictable changes in illumination, background

clutter as well as person pose, which creates a challenging

issue.

In current studies, PReID is resolved from the follow-

ing two angles: 1) Extracting discriminative descriptors to

The associate editor coordinating the review of this manuscript and

approving it for publication was Hugo Proenca.

represent different identities. 2) Learning an effective dis-

tance metric to make the relative distance between the inter-

classes larger than intra-class.

Beneﬁting from the considerable development of deep

learning in the computer vision community, a large number

of deep architecture-based methods have been introduced

for PReID. Different from traditional hand-crafted meth-

ods, these deep learning-based methods integrate feature

and distance metric learning in an end-to-end way. It is

worth noting that the most recent state-of-the-art results

have been achieved by deep learning-based models. Many

of them attempt to learn global pedestrian features. When

the pedestrian global features are generated by the deep

model, the Euclidean metric is applied to measure the dis-

tance between the two pedestrians. However, global feature

28402

2169-3536  2019 IEEE. Translations and content mining are permitted for academic research only.

Personal use is also permitted, but republication/redistribution requires IEEE permission.

See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

VOLUME 7, 2019

下载后可阅读完整内容，剩余9页未读，立即下载

姣孙孙孙

粉丝: 1
资源: 4

多模态深度学习提升行人再识别性能

Omni-Scale Feature Learning for Person Re-Identification

ISO-IEC 14443-3-2016

py_cfar-master.zip_Python__Python_

Large-Scale Direct SLAM for Omnidirectional Cameras.pdf

Dual_Polarized_Omnidirectional_Array_Element_for_MIMO_Systems.pdf

EIKI爱其LC-W5 用户手册.pdf

EIKI爱其LC-X7用户手册.pdf

EIKI爱其LC-WUL100 用户手册.pdf

EIKI爱其EIP-HDT30 用户手册.pdf

EIKI爱其LC-XL100 用户手册.pdf

最新资源