您的位置:山东大学 -> 科技期刊社 -> 《山东大学学报(工学版)》

山东大学学报 (工学版) ›› 2018, Vol. 48 ›› Issue (6): 67-73.doi: 10.6040/j.issn.1672-3961.0.2018.192

• 机器学习与数据挖掘 • 上一篇    下一篇

基于随机擦除行人对齐网络的行人重识别方法

金翠(),王洪元*(),陈首兵   

  1. 常州大学信息科学与工程学院, 江苏 常州 213164
  • 收稿日期:2018-07-18 出版日期:2018-12-20 发布日期:2018-12-26
  • 通讯作者: 王洪元 E-mail:807214072@qq.com;hywang@cczu.edu.cn
  • 作者简介:金翠(1994—),女,江苏常州人,硕士研究生,主要研究方向为计算机视觉.E-mail:807214072@qq.com
  • 基金资助:
    国家自然科学基金资助项目(61572085)

Person re-identification based on random erasing pedestrian alignmentnetwork method

Cui JIN(),Hongyuan WANG*(),Shoubing CHEN   

  1. College of Information Science and Engineering, Changzhou University, Jiangsu 213164, Changzhou, China
  • Received:2018-07-18 Online:2018-12-20 Published:2018-12-26
  • Contact: Hongyuan WANG E-mail:807214072@qq.com;hywang@cczu.edu.cn
  • Supported by:
    国家自然科学基金资助项目(61572085)

摘要:

基于检测出的行人图像容易出现错位和深度网络容易出现过拟合现象的问题,使用行人对齐网络和随机擦除数据增强,对行人数据集进行预处理。使图像生成不同程度的遮挡,并通过仿射估计分支中的空间变换网络层将图像中的错位进行修正。裁剪背景大的部分,填补行人图像缺失的部分,从而降低网络过拟合的现象,提高网络泛化能力。Market1501、DuckMTMC-reID和CUHK03数据集上进行试验,结果表明在rank-1的值达到84%左右。将随机擦除行人对齐网络方法与其他方法进行比较,发现随机擦除行人对齐网络的行人重识别方法的试验结果要好。

关键词: 行人重识别, 行人对齐网络, 数据增强, 过拟合, 图像错位

Abstract:

The detected pedestrian images were prone to misalignment and the depth network was prone to over-fitting phenomenon. Pedestrian datasets were preprocessed using pedestrian alignment networks and random erasing data enhancements. It made the images generating different levels of occlusion, and corrected the misalignment in the images by the spatial transformation network layer in the affine estimation branch. It cropped the large part of the background and filled in the missing part of the pedestrian images, which reduced the phenomenon of network over-fitting and improved the generalization ability of the network. The tests were performed on the Market1501, DuckMTMC-reID and CUHK03 datasets, which showed the value of rank-1 reached approximately 84%. Compared the methods of randomly erasing pedestrian alignment network with other methods, it was found that the test results of pedestrian recognition method for randomly erasing pedestrian alignment network were better.

Key words: person re-identification, pedestrian alignment network, data enhancement, overfitting, image misalignment

中图分类号: 

  • TP391

图1

预处理效果图"

图2

经过对齐网络行人图像前后对比图"

图3

随机擦除数据增强行人对齐网络结构图"

表1

基础分支试验结果"

%
方法 Market1501 DukeMTMC-reID CUHK03(labeled) CUHK03(detected)
rank-1 rank-5 mAP rank-1 rank-5 mAP rank-1 rank-5 mAP rank-1 rank-5 mAP
A 80.17 91.69 59.14 65.20 78.88 44.99 31.14 52.00 29.80 30.50 51.07 29.04
B 81.26 92.35 60.02 69.93 80.67 52.49 32.22 54.03 31.45 31.56 52.36 30.70

表2

对齐分支试验结果"

%
方法 Market1501 DukeMTMC-reID CUHK03(labeled) CUHK03(detected)
rank-1 rank-5 mAP rank-1 rank-5 mAP rank-1 rank-5 mAP rank-1 rank-5 mAP
C 79.01 90.86 58.27 68.36 81.37 47.14 35.29 53.64 32.90 34.14 54.50 31.71
D 81.15 92.77 61.79 66.83 81.66 48.34 37.34 55.58 33.44 35.26 54.68 32.60

表3

随机擦除行人对齐网络与行人对齐网络试验结果"

%
方法 Market1501 DukeMTMC-reID CUHK03(labeled) CUHK03(detected)
rank-1 rank-5 mAP rank-1 rank-5 mAP rank-1 rank-5 mAP rank-1 rank-5 mAP
E 82.81 93.53 63.35 68.36 81.37 51.51 36.86 56.86 35.03 36.29 55.51 34.05
F 84.17 95.34 65.79 71.50 81.66 55.02 38.77 58.69 37.23 38.46 57.73 36.20

表4

Market1501数据集不同方法试验结果比较"

%
方法 rank-1 mAP
DADM 39.40 19.60
FisherNet 48.15 29.94
BoW+KISSME 44.42 20.76
ReRank 77.10 63.60
SVDnet 80.50 55.90
Transfer 83.70 65.50
GAN 83.97 66.10
Ours 84.17 65.79

表5

DukeMTMC-reID数据集不同方法试验结果比较"

%
方法 rank-1 mAP
LOMO+XQDA 30.75 17.04
BoW+KISSME 25.13 12.17
GAN 67.68 47.13
APR 70.69 51.88
ReRank 71.06 52.40
SVDnet 71.32 53.90
Ours 71.50 55.02

表6

CUHK03数据集不同方法试验结果比较"

%
方法 labeled detected
rank-1 mAP rank-1 mAP
LOMO+XQDA 14.80 13.60 12.80 11.50
Resnet50+XQDA 32.00 29.60 31.10 28.20
ReRank 63.50 68.27 64.12 69.26
SVDnet 67.86 72.54 68.63 73.31
Transfer 84.10 84.20
Ours 38.77 37.23 38.46 36.00
1 王金, 刘洁, 高常鑫, 等. 基于姿态对齐的行人重识别方法[J]. 控制理论与应用, 2017, 34 (6): 837- 842.
WANG Jin , LIU Jie , GAO Changxin , et al. Pedestrian re-identification method based on attitude alignment[J]. Control Theory & Applications, 2017, 34 (6): 837- 842.
2 杜久伦.多视图行人重识别算法研究与数据采集[D].济南:山东大学, 2017.
DU Jiulun. Research on multi-view pedestrian recognition algorithm and data acquisition[D]. Jinan: Shandong University, 2017.
3 黄新宇, 许娇龙, 郭纲, 等. 基于增强聚合通道特征的实时行人重识别[J]. 激光与光电子学进展, 2017, 54 (9): 119- 127.
HUANG Xinyu , XU Jiaolong , GUO Gang , et al. Real-time pedestrian re-identification based on enhanced aggregate channel characteristics[J]. Laser & Optoelectronics Progress, 2017, 54 (9): 119- 127.
4 张见威, 林文钊, 邱隆庆. 基于字典学习和Fisher判别稀疏表示的行人重识别方法[J]. 华南理工大学学报(自然科学版), 2017, 45 (7): 55- 62.
doi: 10.3969/j.issn.1000-565X.2017.07.008
ZHANG Jianwei , LIN Wenzhao , QIU Longqing . Pedestrian re-identification method based on dictionary learning and Fisher discriminant sparse representation[J]. Journal of South China University of Technology(Natural Science Edition), 2017, 45 (7): 55- 62.
doi: 10.3969/j.issn.1000-565X.2017.07.008
5 CHEN D , YUAN Z , WANG J , et al. Exemplar-guided similarity learning on polynomial kernel feature map for person re-identification[J]. International Journal of Computer Vision, 2017, 123 (3): 1- 23.
6 WANG X , ZHENG W S , LI X , et al. Cross-scenario transfer person reidentification[J]. IEEE Transactions on Circuits & Systems for Video Technology, 2016, 26 (8): 1447- 1460.
7 AHMED E, JONES M, MARKS T K. An improved deep learning architecture for person re-identification[C]//Proceedings of the 2015 IEEE Conf on Computer Vision and Pattern Recognition(CVPR). Boston, USA: IEEE Press, 2015: 3908-3916.
8 KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagenet classification with deep convolutional neural networks[C]//Proceedings of the International Conference on Neural Information Processing Systems. Lake Tahoe, Spain: [s.n.], 2012: 1097-1105.
9 ZHAO L, LI X, WANG J, et al. Deeply-learned part-aligned representations for person re-identification[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision(ICCV). Venice, Italy: IEEE Press, 2017: 3239-3248.
10 CHEN Y C , ZHU X , ZHENG W S , et al. Person re-identification by camera correlation aware feature augmentation[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2018, 40 (2): 392- 408.
11 JADERBERG M, SIMONYAN K, ZISSERMAN A, et al. Spatial transformer networks[C]//Proceedings of the 2015 IEEE Conf on Computer Vision and Pattern Recognition(CVPR). Boston, USA: IEEE Press, 2015: 2017-2025.
12 HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the 2015 IEEE Conf on Computer Vision and Pattern Recognition(CVPR). Boston, USA: IEEE Press, 2015: 770-778.
13 DING S , LIN L , WANG G , et al. Deep feature learning with relative distance comparison for person re-identification[J]. Pattern Recognition, 2015, 48 (10): 2993- 3003.
doi: 10.1016/j.patcog.2015.04.005
14 ZHENG L, SHEN L, TIAN L, et al. Scalable person re-identification: a benchmark[C]//Proceedings of the 2015 IEEE International Conference on Computer Vision(ICCV). Santiago, Chile: IEEE Press, 2015: 1116-1124.
15 LI W, ZHAO R, XIAO T, et al. DeepReID: deep filter pairing neural network for person re-identification[C]//Proceedings of the 2014 IEEE Computer Vision and Pattern Recognition(CVPR). Colombia, USA: IEEE Press, 2014: 152-159.
16 ZHENG Z, ZHENG L, YANG Y. Unlabeled samples generated by GAN improve the person re-identification baseline in vitro[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy: IEEE Press, 2017: 3774-3782.
[1] 陈大伟,闫昭*,刘昊岩. SVD系列算法在评分预测中的过拟合现象[J]. 山东大学学报(工学版), 2014, 44(3): 15-21.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
[1] 程代展,李志强. 非线性系统线性化综述(英文)[J]. 山东大学学报(工学版), 2009, 39(2): 26 -36 .
[2] 王勇, 谢玉东.

大流量管道煤气的控制技术研究

[J]. 山东大学学报(工学版), 2009, 39(2): 70 -74 .
[3] 刘新1 ,宋思利1 ,王新洪2 . 石墨配比对钨极氩弧熔敷层TiC增强相含量及分布形态的影响[J]. 山东大学学报(工学版), 2009, 39(2): 98 -100 .
[4] 胡天亮,李鹏,张承瑞,左毅 . 基于VHDL的正交编码脉冲电路解码计数器设计[J]. 山东大学学报(工学版), 2008, 38(3): 10 -13 .
[5] 田芳1,张颖欣2,张礼3,侯秀萍3,裘南畹3. 新型金属氧化物薄膜气敏元件基材料的开发[J]. 山东大学学报(工学版), 2009, 39(2): 104 -107 .
[6] 陈华鑫, 陈拴发, 王秉纲. 基质沥青老化行为与老化机理[J]. 山东大学学报(工学版), 2009, 39(2): 125 -130 .
[7] 赵延风1,2, 王正中1,2 ,芦琴1,祝晗英3 . 梯形明渠水跃共轭水深的直接计算方法[J]. 山东大学学报(工学版), 2009, 39(2): 131 -136 .
[8] 李士进,王声特,黄乐平. 基于正反向异质性的遥感图像变化检测[J]. 山东大学学报(工学版), 2018, 48(3): 1 -9 .
[9] 王汝贵,蔡敢为 . 两自由度可控平面连杆机构机电耦合系统的超谐波共振分析[J]. 山东大学学报(工学版), 2008, 38(3): 58 -63 .
[10] 赵科军 王新军 刘洋 仇一泓. 基于结构化覆盖网的连续 top-k 联接查询算法[J]. 山东大学学报(工学版), 2009, 39(5): 32 -37 .