基于行人属性先验分布的行人再识别

吴彦丞; 陈鸿昶; 李邵梅; 高超

doi:10.16383/j.aas.c170691

基于行人属性先验分布的行人再识别

doi: 10.16383/j.aas.c170691

吴彦丞^1, ,,
陈鸿昶^1,,
李邵梅^1,,
高超^1,

国家数字交换系统工程技术研究中心郑州 450002

基金项目:

国家自然科学基金 61601513

国家自然科学基金 61521003

详细信息

作者简介:
陈鸿昶  国家数字交换系统工程技术研究中心教授.主要研究方向为电信网信息关防, 信息安全.E-mail:chc@ndsc.com.cn

李邵梅  国家数字交换系统工程技术研究中心副研究员.主要研究方向为计算机视觉.E-mail:lishaomei_may@126.com

高超  国家数字交换系统工程技术研究中心助理研究员.主要研究方向为计算机视觉.E-mail:chaosndsc@163.com

通讯作者:
吴彦丞国家数字交换系统工程技术研究中心硕士研究生.2016年获得清华大学自动化系学士学位.主要研究方向为机器学习, 计算机视觉.本文通信作者.E-mail:wuyc1994@163.com

计量
- 文章访问数: 2991
- HTML全文浏览量: 757
- PDF下载量: 643
- 被引次数: 0
出版历程
- 收稿日期: 2017-12-07
- 录用日期: 2018-06-09
- 刊出日期: 2019-05-20

Person Re-Identification Using Attribute Priori Distribution

National Digital Switching System Engineering & Technological R & D Center of China, Zhengzhou 450002

Funds:

National Natural Science Foundation of China 61601513

National Natural Science Foundation of China 61521003

More Information

Author Bio:
Professor at China National Digital Switching System Engineering and Technological Research and Development Center. His research interest covers telecom network information gateway, and information security

Associate professor at China National Digital Switching System Engineering and Technological Research and Development Center. Her main research interest is computer vision

Assistant professor at China National Digital Switching System Engineering and Technological Research and Development Center. His main research interest is computer vision

Corresponding author: WU Yan-Cheng Master student at China National Digital Switching System Engineering and Technological Research and Development Center. He received his bachelor degree from Tsinghua University in 2016. His research interest covers machine learning and computer vision. Corresponding auther of this paper

摘要

摘要: 为了提高基于深度学习和属性学习的行人再识别的识别精度，提出一种联合识别行人属性和行人ID的神经网络模型.相对于已有的同类方法，该模型有三大优点：1）为了提高网络在微调后的判别能力，在网络中增加了一层保证模型迁移能力的全连接层；2）基于各属性样本的数量，在损失函数中对各属性的损失进行了归一化处理，避免数据集中属性类之间的数量不均衡对识别效果的影响；3）利用数据中各属性分布的先验知识，通过数量占比来调整各属性在损失层中的权重，避免数据集中各属性正负样本的数量不均衡对识别的影响.实验结果表明，本文提出的算法具有较高的识别率，其中在Market1501数据集上，首位匹配率达到了86.90%，在DukeMTMC数据集上，首位匹配率达到了72.83%，在PETA数据集上，首位匹配率达到了75.68%，且对光照变化、行人姿态变化、视角变化和遮挡都具有很好的鲁棒性.
- 行人再识别 /
- 数据先验分布 /
- 权重调整 /
- 深度学习 /
- 卷积神经网络
Abstract: In order to improve the recognition accuracy of pedestrian recognition based on deep learning and attribute learning, a new neural network model is proposed to identify pedestrian attributes and IDs. Compared with the existing methods, this model has three advantages, one is to increase an all connected layer in the network to guarantee the model's migration which improves the network's discriminant ability after fine-tuning; Second, based on the number of samples of each attribute, we normalized processing the loss of each attribute in the loss function, to avoid the number of disequilibrium effect between classes among the data set which attributes to identify results; third, use the data of each attribute distribution in the prior knowledge, through a number of ratios to adjust each attribute's weight to avoid the influence of positive and negative samples' disequilibrium. Experimental results show that the proposed algorithm has high recognition rate, the rank-1 accuracy reached 86.90% on Market1501 dataset, 72.83% on DukeMTMC dataset and 75.68% on PETA dataset, it also shows that the proposed algorithm has good robustness on illumination changes, pedestrians posture changes, perspective changes and occlusions.
- Person re-identification /
- data priori distribution /
- weight adjustment /
- deep learning /
- convolutional neural network (CNN)
注释:

1) 本文责任编委王亮

HTML全文

图 1 网络结构示意图

Fig. 1 Schematic diagram of network structure

下载: 全尺寸图片幻灯片

图 2 数据集行人图片举例

Fig. 2 Example of dataset pedestrian picture

下载: 全尺寸图片幻灯片

图 3 网络参数及结果对比

Fig. 3 Comparison of network parameters and results

下载: 全尺寸图片幻灯片

图 4 行人属性识别结果举例

Fig. 4 Example of the result of pedestrian attributes

下载: 全尺寸图片幻灯片

图 5 行人再识别结果举例

Fig. 5 Example of the re-id result

下载: 全尺寸图片幻灯片

表 1 Market 1501数据集中的属性类别

Table 1 The attribute category of Market 1501 dataset

属性类($G$)	属性	数量($K$)
Gender	male, female	2
Age	child, teenager	4
Hair length	long, short	2
Lower clothing length	long, short	2
Lower clothing type	pants, dress	2
Wearing hat	yes, no	2
Carrying bag	yes, no	2
Carrying backpack	yes, no	2
Carrying handbag	yes, no	2
Upper clothing color	black, white, red$\cdots$	8
Lower clothing color	black, white, pink$\cdots$	9

下载: 导出CSV

表 2 Market 1501数据集中行人属性训练样本数量及占比

Table 2 Statistics of Market 1501 dataset

属性	数量	占比	属性	数量	占比
upblack	113	0.15	male	431	0.57
upwhite	228	0.30	female	320	0.43
upred	78	0.10	short hair	506	0.67
uppurple	30	0.04	long hair	245	0.33
upyellow	36	0.05	long sleeve	39	0.05
upgray	86	0.11	short sleeve	712	0.95
upblue	46	0.06	long lower body	110	0.15
upgreen	56	0.07	short lower body	641	0.85
handbag no	665	0.89	dress	294	0.39
handbag yes	86	0.11	pants	457	0.61
young	14	0.02	downgray	123	0.16
teenager	569	0.76	downblack	293	0.39
adult	160	0.21	downwhite	58	0.08
old	8	0.01	downpink	29	0.04
bag no	566	0.75	downpurple	2	0.00
bag yes	185	0.25	downyellow	10	0.01
backpack no	552	0.74	downblue	123	0.16
backpack yes	199	0.26	downgreen	14	0.02
hat no	731	0.97	downbrown	69	0.09
hat yes	20	0.03

下载: 导出CSV

表 3 Market 1501数据集各属性识别准确率(%)

Table 3 Accuracy rate of each attribute recognition of Market 1501 dataset (%)

属性	gender	age	hair	L.slv	L.low	S.cloth	B.pack	H.bag	bag	C.up	C.low	mean
APR	86.45	87.08	83.65	93.66	93.32	91.46	82.79	88.98	75.07	73.4	69.91	85.33
本文算法	86.73	88.14	84.12	93.5	94.54	91.86	85.99	90.67	82.36	77.83	73.82	86.32

下载: 导出CSV

表 4 DukeMTMC数据集各属性识别准确率(%)

Table 4 Accuracy rate of each attribute recognition of DukeMTMC dataset (%)

属性	gender	hat	boots	L.up	B.pack	H.bag	bag	C.shoes	C.up	C.low	mean
APR	82.61	86.94	86.15	88.04	77.28	93.75	82.51	90.19	72.29	41.48	80.12
本文算法	82.73	89.02	87.17	89.33	81.33	95.81	86.74	93.12	73.04	43.21	82.15

下载: 导出CSV

表 5 PETA数据集各属性识别准确率(%)

Table 5 Accuracy rate of each attribute recognition of PETA dataset (%)

属性	gender	age	carry	style	hat	hair	shoes	K.up	K.low	bag	glasses	mean
APR	89.51	86.37	78.28	84.69	92.12	89.41	78.95	88.34	84.81	86.76	72.61	84.71
本文算法	90.11	85.32	85.39	85.43	92.63	88.6	82.32	88.97	86.82	88.06	78.33	88.54

下载: 导出CSV

表 6 Market 1501数据集行人再识别结果

Table 6 Re-id results of the Market 1501 dataset

方法	rank-1	rank-5	rank-10	rank-20	mAP
DADM^[17]	39.4	-	-	-	19.6
MBC^[18]	45.56	67	76	82	26.11
SML^[19]	45.16	68.12	76	84	-
DLDA^[20]	48.15	-	-	-	29.94
DNS^[21]	55.43	-	-	-	29.87
LSTM^[8]	61.6	-	-	-	35.3
S-CNN^[22]	65.88	-	-	-	39.55
2Stream^[23]	79.51	90.91	94.09	96.23	59.87
GAN^[13]	79.33	-	-	-	55.95
Pose^[24]	78.06	90.76	94.41	96.52	56.23
Deep^[25]	83.7	-	-	-	65.5
APR	84.29	93.2	95.19	97	64.67
本文-$F{{C}_{0}}$	85.37	94.05	96.13	97.31	65.11
本文-归一化	84.92	93.75	95.92	97.46	64.82
本文-权重	85.67	94.69	96.75	97.94	65.23
本文-归一化+权重	86.31	94.97	97.10	98.01	65.46
本文	86.90	95.37	97.03	98.17	65.87

下载: 导出CSV

表 7 DukeMTMC数据集行人再识别结果

Table 7 Re-id results of the DukeMTMC dataset

方法	rank-1	mAP
BoW+kissme^[12]	25.13	12.17
LOMO+XQDA^[16]	30.75	17.04
GAN^[13]	67.68	47.13
APR	70.69	51.88
本文-$F{{C}_{0}}$	71.56	52.36
本文-归一化	70.92	52.03
本文-权重	71.82	52.67
本文-归一化+权重	72.11	52.84
本文	72.83	53.42

下载: 导出CSV

表 8 PETA数据集行人再识别结果

Table 8 Re-id results of the PETA dataset

方法	rank-1	mAP	mA
ikSVM^[26]	41.12^*	26.87^*	69.5
MRFr2^[27]	51.71^*	30.77^*	75.6
ACN^[28]	59.04^*	35.89^*	81.15
DeepMAR^[10]	64.58^*	41.12^*	82.6
WPAL^[29]	68.14^*	42.68^*	85.5
APR	71.29^*	45.31^*	84.71^*
本文-$F{{C}_{0}}$	72.82	47.75	86.24
本文-归一化	73.63	47.86	87.01
本文-权重	73.14	47.51	87.28
本文-归一化+权重	74.57	49.69	88.05
本文	75.68	51.03	88.54

下载: 导出CSV

参考文献(29)

[1]	Farenzena M, Bazzani L, Perina A, Murino V, Cristani M. Person re-identification by symmetry-driven accumulation of local features. In: Proceedings of the 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. San Francisco, USA: IEEE, 2010. 2360-2367 http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=5539926
[2]	Pedagadi S, Orwell J, Velastin S, Boghossian B. Local fisher discriminant analysis for pedestrian re-identification. In: Proceedings of the 2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland, USA: IEEE, 2013. 3318-3325 http://ieeexplore.ieee.org/document/6619270
[3]	种衍文, 匡湖林, 李清泉.一种基于多特征和机器学习的分级行人检测方法.自动化学报, 2012, 38(3):375-381 http://www.aas.net.cn/CN/abstract/abstract17688.shtml Chong Yan-Wen, Kuang Hu-Lin, Li Qing-Quan. Two-stage pedestrian detection based on multiple features and machine learning. Acta Automatica Sinica, 2012, 38(3):375-381 http://www.aas.net.cn/CN/abstract/abstract17688.shtml
[4]	Zheng W S, Gong S G, Xiang T. Person re-identification by probabilistic relative distance comparison. In: Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition. Providence, RI, USA: IEEE, 2011. 649-656 http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=5995598
[5]	齐美彬, 檀胜顺, 王运侠, 刘皓, 蒋建国.基于多特征子空间与核学习的行人再识别.自动化学报, 2016, 42(2):299-308 http://www.aas.net.cn/CN/abstract/abstract18819.shtml Qi Mei-Bin, Tan Sheng-Shun, Wang Yun-Xia, Liu Hao, Jiang Jian-Guo. Multi-feature subspace and kernel learning for person re-identification. Acta Automatica Sinica, 2016, 42(2):299-308 http://www.aas.net.cn/CN/abstract/abstract18819.shtml
[6]	于雪松, 刘家锋, 唐降龙, 黄剑华.基于概率模型的行人四肢自遮挡的检测.自动化学报, 2010, 36(4):610-615 http://www.aas.net.cn/CN/abstract/abstract13712.shtml Yu Xue-Song, Liu Jia-Feng, Tang Xiang-Long, Huang Jian-Hua. Estimating the pedestrian 3D motion indoor via hybrid tracking model. Acta Automatica Sinica, 2010, 36(4):610-615 http://www.aas.net.cn/CN/abstract/abstract13712.shtml
[7]	Li W, Zhao R, Xiao T, Wang X G. Deepreid: deep filter pairing neural network for person re-identification. In: Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus, USA: IEEE, 2014. 152-159 https://ieeexplore.ieee.org/document/6909421
[8]	Varior R R, Shuai B, Lu J W, Xu D, Wang G. A siamese long short-term memory architecture for human re-identification. In: Proceedings of the 2016 European Conference on Computer Vision. Cham, Switzerland: Springer, 2016. 135-153
[9]	Liu H, Feng J S, Qi M B, Jiang J G, Yan S C. End-to-end comparative attention networks for person re-identification. IEEE Transactions on Image Processing, 2017, 26(7):3492-3506 doi: 10.1109/TIP.2017.2700762
[10]	Li D W, Chen X T, Huang K Q. Multi-attribute learning for pedestrian attribute recognition in surveillance scenarios. In: Proceedings of the 3rd IAPR Asian Conference on Pattern Recognition (ACPR). Kuala Lumpur, Malaysia: IEEE, 2015. 111-115 http://ieeexplore.ieee.org/document/7486476/
[11]	Lin Y T, Zheng L, Zheng Z D, Wu Y, Yang Y. Improving person re-identification by attribute and identity learning[Online], available: http://arxiv.org/abs/1703.07220, June 14, 2018.
[12]	Zheng L, Shen L Y, Tian L, Wang S J, Wang J D, Tian Q. Scalable person re-identification: a benchmark. In: Proceedings of the 2015 IEEE International Conference on Computer Vision. Santiago, Chile: IEEE, 2015. 1116-1124 http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=7410490
[13]	Zheng Z D, Zheng L, Yang Y. Unlabeled samples generated by GAN improve the person re-identification baseline in vitro[Online], available: http://arxiv.org/abs/1701.07717, June 14, 2018.
[14]	Zhang C L, Luo J H, Wei X S, Wu J X. In defense of fully connected layers in visual representation transfer. In: Proceedings of the 2017 Pacific-Rim Conference on Multimedia. Harbin, China: Springer-Verlag, 2017. 807-817 doi: 10.1007/978-3-319-77383-4_79
[15]	Visin F, Kastner K, Cho K, Matteucci M, Courville A, Bengio Y. ReNet: a recurrent neural network based alternative to convolutional networks[Online], available: http://arxiv.org/abs/1505.00393, June 14, 2018.
[16]	Liao S C, Hu Y, Zhu X Y, Li S Z. Person re-identification by local maximal occurrence representation and metric learning. In: Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition. Boston, USA: IEEE, 2015. 2197-2206 http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=7298832
[17]	Su C, Zhang S L, Xing J L, Gao W, Tian Q. Deep attributes driven multi-camera person re-identification. In: Proceedings of the 2016 European Conference on Computer Vision. Cham: Springer, 2016. 475-491 doi: 10.1007/978-3-319-46475-6_30
[18]	Ustinova E, Ganin Y, Lempitsky V. Multi-region bilinear convolutional neural networks for person re-identification. In: Proceedings of the 14th IEEE International Conference on Advanced Video and Signal Based Surveillance. Lecce, Italy: IEEE, 2017. 1-6 http://arxiv.org/abs/1512.05300
[19]	Jose C, Fleuret F. Scalable metric learning via weighted approximate rank component analysis. In: Proceedings of the 2016 European Conference on Computer Vision. Cham: Springer, 2016. 875-890 doi: 10.1007/978-3-319-46454-1_53
[20]	Wu L, Shen C H, van den Hengel A. Deep linear discriminant analysis on fisher networks:a hybrid architecture for person re-identification. Pattern Recognition, 2017, 65:238-250 doi: 10.1016/j.patcog.2016.12.022
[21]	Zhang L, Xiang T, Gong S G. Learning a discriminative null space for person re-identification. In: Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, USA: IEEE, 2016. 1239-1248
[22]	Varior R R, Haloi M, Wang G. Gated siamese convolutional neural network architecture for human re-identification. In: Proceedings of the 2016 European Conference on Computer Vision. Cham, Switerland: Springer, 2016. 791-808 doi: 10.1007/978-3-319-46484-8_48
[23]	Zheng Z D, Zheng L, Yang Y. A discriminatively learned CNN embedding for person reidentification. ACM Transactions on Multimedia Computing, Communications, and Applications, 2018, 14(1):Article No.13 https://arxiv.org/abs/1611.05666
[24]	Zheng L, Huang Y J, Lu H C, Yang Y. Pose invariant embedding for deep person re-identification[Online], available: http://arxiv.org/abs/1701.07732, June 14, 2018.
[25]	Geng M Y, Wang Y W, Xiang T, Tian Y H. Deep transfer learning for person re-identification[Online], available: http://arxiv.org/abs/1611.05244, June 14, 2018.
[26]	Maji S, Berg A C, Malik J. Classification using intersection kernel support vector machines is efficient. In: Proceedings of the 2008 IEEE Conference on Computer Vision and Pattern Recognition. Anchorage, USA: IEEE, 2008. 1-8 http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=4587630
[27]	Deng Y B, Luo P, Loy C C, Tang X O. Learning to recognize pedestrian attribute[Online], available: http://arxiv.org/abs/1501.00901, June 14, 2018.
[28]	Sudowe P, Spitzer H, Leibe B. Person attribute recognition with a jointly-trained holistic CNN model. In: Proceedings of the 2015 IEEE International Conference on Computer Vision Workshops. Santiago, Chile: IEEE, 2015. 329-337 http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=7406400&navigation=1
[29]	Yu K, Leng B, Zhang Z, Li D W, Huang K Q. Weakly-supervised learning of mid-level features for pedestrian attribute recognition and localization[Online], available: http://arxiv.org/abs/1611.05603, June 14, 2018.