基于双向参考集矩阵度量学习的行人再识别

陈莹; 许潇月

doi:10.11999/JEIT190159

引用本文: 陈莹, 许潇月. 基于双向参考集矩阵度量学习的行人再识别[J]. 电子与信息学报, 2020, 42(2): 394-402. doi: 10.11999/JEIT190159 shu

Citation: Ying CHEN, Xiaoyue XU. Matrix Metric Learning for Person Re-identification Based on Bidirectional Reference Set[J]. Journal of Electronics and Information Technology, 2020, 42(2): 394-402. doi: 10.11999/JEIT190159 shu

基于双向参考集矩阵度量学习的行人再识别

陈莹 ^,
, 许潇月

江南大学轻工过程先进控制教育部重点实验室无锡 214000

作者简介: 陈莹: 女，1976年生，教授，博士生导师，研究方向为模式识别、信息融合;
许潇月: 女，1994年生，硕士生，研究方向为行人再识别

通讯作者: 陈莹,chenying@jiangnan.edu.cn

基金项目: 国家自然科学基金(61573168)

摘要: 针对行人再识别中由于外观差异不显著导致特征描述不准确的问题，该文提出一种基于双向参考集矩阵度量学习(BRM²L)的行人再识别算法。首先通过互近邻算法获得每个摄像头下的互近邻参考集，为保证参考集的鲁棒性，联合考虑各摄像头下的互近邻参考集获得双向参考集。通过双向参考集挖掘出困难样本进行特征描述，从而得到准确的外观差异描述。最后利用该特征描述进行更有效的矩阵度量学习。在多个公开数据集上的实验结果证明了该算法比现有算法具有更好的行人再识别性能。

关键词:

English

Matrix Metric Learning for Person Re-identification Based on Bidirectional Reference Set

Ying CHEN ^,
, Xiaoyue XU

Key Laboratory of Advanced Control Light Process, Jiangnan University, Wuxi 214000, China

Corresponding author: Ying CHEN,chenying@jiangnan.edu.cn

Received Date: 2019-03-18
Accepted Date: 2019-05-24
Available Online: 2020-02-01
Fund Project: The National Natural Science Foundation of China (61573168)

Abstract: To solve the problem of inaccurate feature representation caused by indistinctive appearance difference in person re-identification domain, a new Matrix Metric Learning algerithm based on Bidirectional Reference (BRM²L) set is proposed. Firstly, reciprocal-neighbor reference sets in different camera views are respectively constructed by the reciprocal-neighbor scheme. To ensure the robustness of reference sets, the reference sets in different camera views are jointly considered to generate the Bidirectional Reference Set (BRS). With hard samples which are mined by the BRS to represent feature descriptors, accurate appearance difference representations could be obtained. Finally, these representations are utilized to conduct more effective matrix metric learning. Experimental results on several public datasets demonstrate the superiority of the proposed method.

Key words:

1. [1]
  FARENZENA M, BAZZANI L, PERINA A, et al. Person re-identification by symmetry-driven accumulation of local features[C]. 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Francisco, USA, 2010: 2360–2367.
2. [2]
  李幼蛟, 卓力, 张菁, 等. 行人再识别技术综述[J]. 自动化学报, 2018, 44(9): 1554–1568. doi: 10.16383/j.aas.2018.c170505
  LI Youjiao, ZHUO Li, ZHANG Jing, et al. A survey of person re-identification[J]. Acta Automatica Sinica, 2018, 44(9): 1554–1568. doi: 10.16383/j.aas.2018.c170505
3. [3]
  ZHENG Lilei, DUFFNER S, IDRISSI K, et al. Pairwise identity verification via linear concentrative metric learning[J]. IEEE Transactions on Cybernetics, 2018, 48(1): 324–335. doi: 10.1109/TCYB.2016.2634011
4. [4]
  WANG Zheng, HU Ruimin, CHEN Chen, et al. Person reidentification via discrepancy matrix and matrix metric[J]. IEEE Transactions on Cybernetics, 2018, 48(10): 3006–3020. doi: 10.1109/TCYB.2017.2755044
5. [5]
  CHEN Xiaojing, AN Le, and BHANU B. Reference set based appearance model for tracking across non-overlapping cameras[C]. 2013 International Conference on Distributed Smart Cameras, Palm Springs, USA, 2013: 1–6.
6. [6]
  AN Le, KAFAI M, YANG Songfan, et al. Person reidentification with reference descriptor[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2016, 26(4): 776–787. doi: 10.1109/TCSVT.2015.2416561
7. [7]
  LIAO Shengcai and LI S Z. Efficient PSD constrained asymmetric metric learning for person re-identification[C]. 2015 IEEE International Conference on Computer Vision, Santiago, Chile, 2015: 3685–3693.
8. [8]
  ZHONG Zhun, ZHENG Liang, CAO Donglin, et al. Re-ranking person re-identification with k-reciprocal encoding[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 3652–3661.
9. [9]
  SUN Yipeng, TAO Xiaoming, LI Yang, et al. Robust two-dimensional principal component analysis via alternating optimization[C]. 2013 IEEE International Conference on Image Processing, Melbourne, Australia, 2013: 340–344.
10. [10]
  GRAY D, BRENNAN S, and TAO Hai. Evaluating appearance models for recognition, reacquisition, and tracking[C]. The 10th IEEE International Workshop on Performance Evaluation for Tracking and Surveillance, Rio de Janeiro, 2007: 1–7.
11. [11]
  ROTH P M, HIRZER M, KÖSTINGER M, et al. Mahalanobis Distance Learning For Person Re-identification[M]. London: Springer, 2014: 247–267.
12. [12]
  LI Wei, ZHAO Rui, and WANG Xiaogang. Human reidentification with transferred metric learning[C]. The 11th Asian Conference on Computer Vision, Daejeon, Korea, 2012: 31–44.
13. [13]
  WANG Xiaogang, DORETTO G, SEBASTIAN T, et al. Shape and appearance context modeling[C]. The 11th IEEE International Conference on Computer Vision, Rio de Janeiro, Brazil, 2007: 1–8.
14. [14]
  MATSUKAWA T, OKABE T, SUZUKI E, et al. Hierarchical gaussian descriptor for person re-identification[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 1363–1372.
15. [15]
  MATSUKAWA T and SUZUKI E. Person re-identification using CNN features learned from combination of attributes[C]. The 23rd International Conference on Pattern Recognition, Cancun, Mexico, 2016: 2428–2433.
16. [16]
  MIGNON A and JURIE F. PCCA: A new approach for distance learning from sparse pairwise constraints[C]. 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, USA, 2012: 2666–2672.
17. [17]
  MA Bingpeng, SU Yu, and JURIE F. Bicov: A novel image representation for person re-identification and face verification[C]. British Machive Vision Conference, Surrey, UK, 2012: 57. 1–57.
18. [18]
  KÖSTINGER M, HIRZER M, WOHLHART P, et al. Large scale metric learning from equivalence constraints[C]. 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, USA, 2012: 2288–2295.
19. [19]
  ZHAO Rui, OUYANG Wanli, and WANG Xiaogang. Unsupervised salience learning for person re-identification[C]. 2013 IEEE Conference on Computer Vision and Pattern Recognition, Portland, USA, 2013: 3586–3593.
20. [20]
  LI Zhen, CHANG Shiyu, LIANG Feng, et al. Learning locally-adaptive decision functions for person verification[C]. 2013 IEEE Conference on Computer Vision and Pattern Recognition, Portland, USA, 2013: 3610–3617.
21. [21]
  ZHAO Rui, OUYANG Wanli, and WANG Xiaogang. Learning mid-level filters for person re-identification[C]. 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, USA, 2014: 144–151.
22. [22]
  YANG Yang, YANG Jimei, YAN Junjie, et al. Salient color names for person re-identification[C]. The 13th European Conference on Computer Vision, Zurich, Switzerland, 2014: 536–551.
23. [23]
  LIAO Shengcai, HU Yang, ZHU Xiangyu, et al. Person re-identification by local maximal occurrence representation and metric learning[C]. 2015 IEEE Conference on Computer Vision and Pattern Recognition, Boston, USA, 2015: 2197–2206.
24. [24]
  YI Dong, LEI Zhen, LIAO Shengcai, et al. Deep metric learning for person re-identification[C]. 2014 International Conference on Pattern Recognition, Stockholm, Sweden, 2014: 34–39.
25. [25]
  CHEN Shizhe, GUO Chaochun, and LAI Jianhuang. Deep ranking for person re-identification via joint representation learning[J]. IEEE Transactions on Image Processing, 2016, 25(5): 2353–2367. doi: 10.1109/TIP.2016.2545929
26. [26]
  WANG Jin, WANG Zheng, GAO Changxin, et al. DeepList: Learning deep features with adaptive listwise constraint for person reidentification[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2017, 27(3): 513–524. doi: 10.1109/tcsvt.2016.2586851
27. [27]
  DE CARVALHO PRATES R F and SCHWARTZ W R. CBRA: Color-based ranking aggregation for person re-identification[C]. 2015 IEEE International Conference on Image Processing, Quebec City, Canada, 2015: 1975–1979.
28. [28]
  SHEN Yang, LIN Weiyao, YAN Junchi, et al. Person re-identification with correspondence structure learning[C]. 2015 IEEE International Conference on Computer Vision, Santiago, Chile, 2015: 3200–3208.
29. [29]
  CHEN Yingcong, ZHENG Weishi, and LAI Jianhuang. Mirror representation for modeling view-specific transform in person re-identification[C]. The 24th International Conference on Artificial Intelligence, Buenos Aires, Argentina, 2015: 3402–3408.
30. [30]
  YAO Wenbin, WENG Zhenyu, and ZHU Yuesheng. Diversity regularized metric learning for person re-identification[C]. 2016 IEEE International Conference on Image Processing, Phoenix, USA, 2016: 4264–4268.
31. [31]
  AHMED E, JONES M, and MARKS T K. An improved deep learning architecture for person re-identification[C]. 2015 IEEE Conference on Computer Vision and Pattern Recognition, Boston, USA, 2015: 3908–3916.
1. [1]
  陈鸿昶, 吴彦丞, 李邵梅, 高超. 基于行人属性分级识别的行人再识别. 电子与信息学报, 2019, 41(9): 2239-2246.
2. [2]
  杜宇宁, 艾海舟. 基于统计推断的行人再识别算法. 电子与信息学报, 2014, 36(7): 1612-1618.
3. [3]
  陈鸿昶, 陈雷, 李邵梅, 朱俊光. 基于显著度融合的自适应分块行人再识别. 电子与信息学报, 2017, 39(11): 2652-2660.
4. [4]
  孙锐, 方蔚, 黄启恒, 高隽. 基于支持样本间接式的行人再识别. 电子与信息学报, 2017, 39(12): 2953-2961.
5. [5]
  曾明勇, 吴泽民, 田畅, 付毅, 揭斐然. 基于外观统计特征融合的人体目标再识别. 电子与信息学报, 2014, 36(8): 1844-1851.
6. [6]
  周智恒, 刘楷怡, 黄俊楚, 陈增群. 一种基于等距度量学习策略的行人重识别改进算法. 电子与信息学报, 2019, 41(2): 477-483.
7. [7]
  张清华. Vague值(集)相似度量的研究. 电子与信息学报, 2007, 29(8): 1855-1859.
8. [8]
  董俊, 王锁萍, 熊范纶. 可变相似性度量的近邻传播聚类. 电子与信息学报, 2010, 32(3): 509-514.
9. [9]
  许良凤, 王家勇, 崔婧楠, 胡敏, 张柯柯, 滕文娣. 基于动态时间规整和主动外观模型的动态表情识别. 电子与信息学报, 2018, 40(2): 338-345.
10. [10]
  李乐, 章毓晋. 非负矩阵集分解. 电子与信息学报, 2009, 31(2): 255-260.
11. [11]
  朱建清, 曾焕强, 杜永兆, 雷震, 郑力新, 蔡灿辉. 基于新型三元卷积神经网络的行人再辨识算法. 电子与信息学报, 2018, 40(4): 1012-1016.
12. [12]
  徐晓滨, 文成林, 王迎昌. 基于模糊故障特征信息的随机集度量信息融合诊断方法. 电子与信息学报, 2009, 31(7): 1635-1640.
13. [13]
  于斌, 尹成友, 黄冶. 基于RBF神经网络的互耦逆矩阵快速计算方法. 电子与信息学报, 2008, 30(6): 1496-1499.
14. [14]
  司伟建, 吴迪, 陈涛, 吴娜. 一种阵列互耦矩阵与波达方向的级联估计方法. 电子与信息学报, 2014, 36(7): 1599-1604.
15. [15]
  孙锐, 张广海, 高隽. 基于深度分层特征表示的行人识别方法. 电子与信息学报, 2016, 38(6): 1528-1535.
16. [16]
  储岳中, 徐波, 高有涛, 邰伟鹏. 基于近邻传播聚类与核匹配追踪的遥感图像目标识别方法. 电子与信息学报, 2014, 36(12): 2923-2928.
17. [17]
  何怡刚, 佘培亮, 佐磊, 张超群. 超高频射频识别近场系统互耦效应中频率偏移研究. 电子与信息学报, 2019, 41(3): 602-610.
18. [18]
  陈兵, 查宇飞, 李运强, 张胜杰, 张园强. 平移变化性相似性学习的行人重识别算法. 电子与信息学报, 2018, 40(10): 2381-2387.
19. [19]
  黄汝激. 应用超图理论实现有向基本割集矩阵. 电子与信息学报, 1992, 14(1): 50-60.
20. [20]
  李玉博, 田立影. 基于正交矩阵构造非周期组间互补序列集. 电子与信息学报, 2018, 40(8): 2028-2032.

图 1 差异矩阵描述子

下载: 全尺寸图片幻灯片

图 2 算法框架图

下载: 全尺寸图片幻灯片

图 3 双向参考集与随机参考集结果对比图

下载: 全尺寸图片幻灯片

表 1 两个摄像头下参考集里的样本标签的重叠率$\sigma $(%)

行人	行人1	行人2	行人3	行人4
重叠率$\sigma $	20	50	10	25

下载: 导出CSV

表 2 在3个数据集上采用不同特征的匹配精度(%)

方法	VIPeR		CHUK01		PRID450S
方法	Rank-1	Rank-5	Rank-1	Rank-5	Rank-1	Rank-5
${{\rm{L}}_{\rm{2}}}$范数(GoG)	19.00	38.00	24.17	51.33	32.44	60.00
F范数(GoG)	20.17	41.83	34.50	69.83	52.22	80.22
${\rm{BR}}{{\rm{M}}^{\rm{2}}}{\rm{L}}$(GoG)	38.33	69.17	45.33	70.50	54.44	80.67
${{\rm{L}}_{\rm{2}}}$范数(FCTNN)	29.00	46.00	37.44	58.00	31.73	57.96
F范数(FCTNN)	30.00	49.83	46.56	72.11	44.40	72.84
${\rm{BR}}{{\rm{M}}^{\rm{2}}}{\rm{L}}$(FCTNN)	41.33	68.17	47.42	77.44	45.51	72.96

下载: 导出CSV

表 3 VIPeR数据集上的结果

方法	Rank-1	Rank-5	Rank-10	Rank-20
PCCA^[16]	19.3	48.9	64.9	80.3
KISSME^[18]	19.6	48.0	62.2	77.0
BiCov^[17]	20.6	43.2	56.1	68.0
eSDC^[19]	26.3	46.4	58.6	72.8
DeepMetric^[24]	28.2	59.3	73.4	86.4
Midfilter^[21]	29.1	52.5	65.9	79.9
LADF^[20]	30.0	64.0	80.0	92.0
FTCNN^[15]+XQDA	31.2	59.8	74.0	83.5
RD^[6]	33.3	65.1	78.3	88.5
GoG^[14]+XQDA	37.3	67.4	77.2	89.6
SCNCD^[22]	37.8	68.5	81.2	90.4
${\rm{D}}{{\rm{M}}^{\rm{3}}}$^[4]	38.3	67.2	77.0	89.3
DeepRanking^[25]	38.4	69.2	81.3	90.4
LOMO+XQDA^[23]	40.0	68.5	80.5	91.0
DeepList^[26]	40.5	69.1	80.1	91.2
${\rm{BR}}{{\rm{M}}^2}{\rm{L}}$(GoG)	38.33	69.17	81.50	89.50
${\rm{BR}}{{\rm{M}}^2}{\rm{L}}$(FTCNN)	41.33	68.17	82.00	90.33

下载: 导出CSV

表 4 PRID 450S数据集上的结果

方法	Rank-1	Rank-5	Rank-10	Rank-20
KISSME^[18]	33.0	59.8	71.0	79.0
CBRA^[27]	26.4	57.1	71.0	83.2
CSL^[28]	44.4	71.6	82.2	89.8
Mirror^[29]	55.4	79.3	87.8	93.9
DRML^[30]	56.4	–	82.2	90.2
DM³^[4]	56.7	83.1	88.4	94.7
${\rm{BR}}{{\rm{M}}^2}{\rm{L}}$(GoG)	54.44	80.67	89.78	95.56
${\rm{BR}}{{\rm{M}}^2}{\rm{L}}$(FTCNN)	59.20	84.53	94.53	99.78

下载: 导出CSV

表 5 CUHK01数据集上的结果

方法	Rank-1	Rank-5	Rank-10	Rank-20
SDALF^[1]	9.9	22.6	30.3	41.0
TML^[12]	20.0	43.5	56.0	69.3
MidFilter^[21]	34.3	55.1	65.0	74.9
ImprovedDeep^[31]	47.5	71.0	80.0	–
RD^[6]	31.1	–	68.5	79.1
${\rm{D}}{{\rm{M}}^{\rm{3}}}$^[4]	43.7	70.1	77.4	88.7
${\rm{BR}}{{\rm{M}}^2}{\rm{L}}$(GoG)	45.33	70.50	86.50	90.00
${\rm{BR}}{{\rm{M}}^2}{\rm{L}}$(FTCNN)	47.42	77.44	88.33	98.33