Video-Based Person Re-Identification via Combined Multi-Level Deep Feature Representation and Ordered Weighted Distance Fusion

Rui Sun; Qiheng Huang; Weiming Lu; Jun Gao

doi:10.3788/AOS201939.0915006

Acta Optica Sinica, Volume. 39, Issue 9, 0915006(2019)

Video-Based Person Re-Identification via Combined Multi-Level Deep Feature Representation and Ordered Weighted Distance Fusion

Rui Sun^1,2, Qiheng Huang^1,2、*, Weiming Lu^1,2, and Jun Gao¹

Author Affiliations

¹ School of Computer and Information, Hefei University of Technology, Hefei, Anhui 230009, China

² Anhui Provincial Key Laboratory of Industry Safety and Emergency Technology, Hefei, Anhui 230009, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(18)

Fig. 1. Challenges and difficulties in video-based person re-identification

Download full size

Fig. 2. Architecture of video-based person re-identification algorithm combining multi-level depth feature representation and ordered weighted distance fusion

Download full size

Fig. 3. Structure of RNN

Download full size

Fig. 4. Flowchart of distance metric learning algorithm used for computing distance between persons

Download full size

Fig. 5. Algorithm of ordered weighted distance fusion

Download full size

Fig. 6. CMCs of comparison with still-image-based person re-identification methods on i-LIDS and PRID-2011 datasets. (a) i-LIDS dataset; (b) PRID-2011 dataset

Download full size

Fig. 7. CMCs of comparison with video-based person re-identification methods on i-LIDS and PRID-2011 datasets. (a) i-LIDS dataset; (b) PRID-2011 dataset

Download full size

Fig. 8. CMCs of comparison with deep-learning-based person re-identification methods on i-LIDS and PRID-2011 dataset. (a) i-LIDS dataset; (b) PRID-2011 dataset

Download full size

Table 1. Structure parameters of global and local appearance depth feature network

View table

Table 1. Structure parameters of global and local appearance depth feature network

Network layer/Parameter
Conv1/Filters(96,11×11),Stride 4
Max pooling1/Filters(3×3),Stride 2
Conv2/Filters(256,5×5),Stride 1,pad 2
Max pooling2/Filters(3×3),Stride 2
Conv3/Filters(384,3×3),Stride 1,pad 1
Conv4/Filters(384,3×3),Stride 1,pad 1
Conv5/Filters(256,3×3),Stride 1,pad 1
Max pooling5/Filters(3×3),Stride 2
Fc6 layer	Slice layer
Fc6 layer	Fc6_Lh layer	Fc6_Ll layer
Fc7 layer	Concat layer

Table 1. 0 Results of comparison with key factors on PRID 2011 dataset
View table
Table 1. 0 Results of comparison with key factors on PRID 2011 dataset
Method Matching rate of PRID-2011 dataset /%
Rank-1 Rank-5 Rank-10 Rank-20
Proposed 71.8 90.4 94.1 97.5
LAF-1 12.4 30.3 41.6 61.8
SF-1 33.7 65.2 75.3 87.6
GAF-1 40.4 64.0 79.8 91.0
LAF+ST 28.1 53.9 68.5 80.9
GAF+ST 42.7 67.4 81.8 89.9

Table 2. Public datasets for person re-identification

View table

Table 2. Public datasets for person re-identification

Experimental dataset	Characteristic	Object
PRID-2011	Viewpoint, lighting, and background changes	Outdoor person
i-LIDS	Viewpoint and lighting changes, occlusions	Airport person
VIPeR	Viewpoint, lighting, and background changes	School person
CAVIAR4REID	Lighting changes and occlusions	Shopping person
GRID	Viewpoint and lighting changes, low resolution	Subway person

Table 3. Results of comparison with still-image-based person re-identification methods on i-LIDS dataset
View table
Table 3. Results of comparison with still-image-based person re-identification methods on i-LIDS dataset
Method Matching rate of i-LIDS dataset (image-based methods) /%
Rank-1 Rank-5 Rank-10 Rank-20
Proposed 55.7 84.7 95.3 96.7
GRDL 25.7 49.9 63.2 77.6
DVDL 25.9 48.2 57.3 68.9
SRID 24.9 44.5 55.6 66.2
Salience 10.2 24.8 35.5 52.9

Table 4. Results of comparison with still-image-based person re-identification methods on PRID 2011 dataset
View table
Table 4. Results of comparison with still-image-based person re-identification methods on PRID 2011 dataset
Method Matching rate of PRID-2011 dataset (image-based methods) /%
Rank-1 Rank-5 Rank-10 Rank-20
Proposed 71.8 90.4 94.1 97.5
GRDL 41.6 76.4 84.9 89.9
DVDL 40.6 69.7 77.8 85.6
SRID 35.1 59.4 69.8 79.7
Salience 25.8 43.6 52.6 62.0

Table 5. Results of comparison with video-based person re-identification methods on i-LIDS dataset

View table

Table 5. Results of comparison with video-based person re-identification methods on i-LIDS dataset

Method	Matching rate of i-LIDS dataset (video-based methods) /%
Method	Rank-1	Rank-5	Rank-10	Rank-20
Proposed	55.7	84.7	95.3	96.7
TDL	56.3	87.6	95.6	98.3
DGM+IDE	37.2	62.6	73.4	80.8
VR	23.3	42.4	55.3	68.4
PAMM	30.3	56.3	70.3	82.7
UNKISS	35.9	63.3	74.9	83.4
ISR	11.6	22.1	27.4	36.7

Table 6. Results of comparison with video-based person re-identification methods on PRID 2011 dataset

View table

Table 6. Results of comparison with video-based person re-identification methods on PRID 2011 dataset

Method	Matching rate of PRID-2011 dataset (video-based methods) /%
Method	Rank-1	Rank-5	Rank-10	Rank-20
Proposed	71.8	90.4	94.1	97.5
TDL	56.7	80.0	87.6	93.6
DGM+IDE	61.6	89.0	94.8	98.2
VR	28.9	55.3	65.5	82.8
PAMM	45.0	72.0	85.0	92.5
UNKISS	58.1	81.9	89.6	96.0
ISR	17.6	35.8	43.0	54.4

Table 7. Results of comparison with deep-learning-based person re-identification methods on i-LIDS dataset

View table

Table 7. Results of comparison with deep-learning-based person re-identification methods on i-LIDS dataset

Method	Matching rate of i-LIDS dataset (deep learning methods) /%
Method	Rank-1	Rank-5	Rank-10	Rank-20
Proposed	55.7	84.7	95.3	96.7
RFA	49.3	76.8	85.3	90.0
Deep-RCN	42.6	70.2	86.4	92.3
CNN-KISS	48.8	75.6	-	92.6
BRNN	55.3	85.0	91.7	95.1
SRM-TAM	55.2	86.5	-	97.0
TRL	57.7	81.7	-	94.1

Table 8. Results of comparison with deep-learning-based person re-identification methods on PRID 2011 dataset

View table

Table 8. Results of comparison with deep-learning-based person re-identification methods on PRID 2011 dataset

Method	Matching rate of PRID-2011 dataset (deep learning methods) /%
Method	Rank-1	Rank-5	Rank-10	Rank-20
Proposed	71.8	90.4	94.1	97.5
RFA	58.2	85.8	93.4	97.9
Deep-RCN	49.8	77.4	90.7	94.6
CNN-KISS	69.9	90.6	-	98.2
BRNN	72.8	92.0	95.1	97.6
SRM-TAM	79.4	94.4	-	99.3
TRL	87.8	97.4		99.3

Table 9. Results of comparison with key factors on i-LIDS dataset
View table
Table 9. Results of comparison with key factors on i-LIDS dataset
Method Matching rate of i-LIDS dataset /%
Rank-1 Rank-5 Rank-10 Rank-20
Proposed 55.7 84.7 95.3 96.7
LAF-1 5.3 14.0 21.3 32.7
ST-1 38.7 71.3 82.3 88.0
GAF-1 46.0 74.0 79.3 89.3
LAF+ST 24.1 52.0 70.6 83.4
GAF+ST 31.5 61.4 71.3 84.0

Tools

Get Citation

Copy Citation Text

Rui Sun, Qiheng Huang, Weiming Lu, Jun Gao. Video-Based Person Re-Identification via Combined Multi-Level Deep Feature Representation and Ordered Weighted Distance Fusion[J]. Acta Optica Sinica, 2019, 39(9): 0915006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Jan. 14, 2019

Accepted: May. 31, 2019

Published Online: Sep. 9, 2019

The Author Email: Huang Qiheng (jchqh123@163.com)

DOI:10.3788/AOS201939.0915006

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Structure parameters of global and local appearance depth feature network

Table 1. Structure parameters of global and local appearance depth feature network

Table 1. 0 Results of comparison with key factors on PRID 2011 dataset

Table 1. 0 Results of comparison with key factors on PRID 2011 dataset

Table 2. Public datasets for person re-identification

Table 2. Public datasets for person re-identification

Table 3. Results of comparison with still-image-based person re-identification methods on i-LIDS dataset

Table 3. Results of comparison with still-image-based person re-identification methods on i-LIDS dataset

Table 4. Results of comparison with still-image-based person re-identification methods on PRID 2011 dataset

Table 4. Results of comparison with still-image-based person re-identification methods on PRID 2011 dataset

Table 5. Results of comparison with video-based person re-identification methods on i-LIDS dataset

Table 5. Results of comparison with video-based person re-identification methods on i-LIDS dataset

Table 6. Results of comparison with video-based person re-identification methods on PRID 2011 dataset

Table 6. Results of comparison with video-based person re-identification methods on PRID 2011 dataset

Table 7. Results of comparison with deep-learning-based person re-identification methods on i-LIDS dataset

Table 7. Results of comparison with deep-learning-based person re-identification methods on i-LIDS dataset

Table 8. Results of comparison with deep-learning-based person re-identification methods on PRID 2011 dataset

Table 8. Results of comparison with deep-learning-based person re-identification methods on PRID 2011 dataset

Table 9. Results of comparison with key factors on i-LIDS dataset

Table 9. Results of comparison with key factors on i-LIDS dataset