CN109063535B - Pedestrian re-identification and pedestrian gender classification method based on joint deep learning - Google Patents

Pedestrian re-identification and pedestrian gender classification method based on joint deep learning Download PDF

Info

Publication number
CN109063535B
CN109063535B CN201810541294.7A CN201810541294A CN109063535B CN 109063535 B CN109063535 B CN 109063535B CN 201810541294 A CN201810541294 A CN 201810541294A CN 109063535 B CN109063535 B CN 109063535B
Authority
CN
China
Prior art keywords
pedestrian
identification
gender classification
parameter
same structure
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810541294.7A
Other languages
Chinese (zh)
Other versions
CN109063535A (en
Inventor
朱建清
曾焕强
陈婧
蔡灿辉
杜永兆
傅玉青
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huaqiao University
Original Assignee
Huaqiao University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huaqiao University filed Critical Huaqiao University
Priority to CN201810541294.7A priority Critical patent/CN109063535B/en
Publication of CN109063535A publication Critical patent/CN109063535A/en
Application granted granted Critical
Publication of CN109063535B publication Critical patent/CN109063535B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a pedestrian re-identification and pedestrian gender classification method based on joint deep learning, which can simultaneously predict the identity and gender of a pedestrian. Firstly, constructing two depth networks with the same structure, and respectively using the two depth networks for pedestrian re-identification and pedestrian gender classification; secondly, parameter sets of each layer in the two depth networks with the same structure are constrained by using parameter related regular terms, so that the parameters of the two depth networks are not greatly deviated in the optimization process, overfitting is avoided, and the accuracy of pedestrian re-identification and pedestrian gender classification is improved at the same time.

Description

Pedestrian re-identification and pedestrian gender classification method based on joint deep learning
Technical Field
The invention relates to intelligent video monitoring, machine vision and machine learning, in particular to a pedestrian re-identification and pedestrian gender classification method based on joint deep learning.
Background
In recent years, a large number of pedestrian re-identification or pedestrian gender classification algorithms based on deep networks have emerged. However, these algorithms often use pedestrian re-identification or pedestrian gender classification as two mutually independent tasks, and do not perform joint learning on the two tasks, thus naturally leaving room for improving the accuracy of pedestrian re-identification and pedestrian gender classification.
Some deep networks that have been successfully applied to face recognition, such as ResNet, *** net, DenseNet, are not necessarily directly applicable to pedestrian re-recognition or pedestrian gender classification tasks, because the database size on the pedestrian side is much lower than the database size on the face side. The deep network is easy to generate an overfitting phenomenon on a small-scale pedestrian database, so that the accuracy of a pedestrian re-identification or pedestrian gender classification algorithm is limited.
Disclosure of Invention
The invention aims to provide a pedestrian re-identification and pedestrian gender classification method based on combined deep learning, which can be used for simultaneously predicting the identity and the gender of a pedestrian and avoiding the overfitting of a deep network, thereby being beneficial to simultaneously improving the accuracy of the re-identification and the gender classification of the pedestrian.
In order to achieve the purpose, the invention adopts the technical scheme that:
a pedestrian re-identification and pedestrian gender classification method based on joint deep learning specifically comprises the following steps:
step 1, constructing two depth networks with the same structure, wherein the two depth networks are respectively used for pedestrian re-identification and pedestrian gender classification;
and 2, utilizing a parameter related regular term to constrain parameter sets of each layer in the two depth networks with the same structure.
Further, the two depth networks with the same structure used for pedestrian re-identification and pedestrian gender classification in the step 1 respectively have corresponding parameter sets of H1＝{H11,H12,…,H1nH and2＝{H21,H22,…,H2nwhere n is the number of convolutional layers in the deep network, the parameters of the two are related to a regular term R (H)1,H2) The definition is as follows:
Figure BDA0001678881330000021
wherein,
Figure BDA0001678881330000022
represents a dot product operation, H1iRepresents a parameter set H1Parameter of the ith convolution layer, H2iRepresents a parameter set H2The parameters of the ith convolutional layer in (c),
Figure BDA0001678881330000023
a Frobenius norm (Frobenius norm) representing a matrix;
gradient descent algorithm based pair H1And H2Optimization is carried out, namely the updating rule of the parameters in back propagation is as follows:
Figure BDA0001678881330000024
wherein L is1And L2Representing the Softmax loss functions for pedestrian re-identification and pedestrian gender classification, respectively; lambda is more than or equal to 0 and is used for controlling parameter related regular term R (H)1,H2) The weight of (c); t +1 and t represent two adjacent iterations in the gradient descent algorithm;
Figure BDA0001678881330000025
representing a partial derivative calculation.
As is apparent from the formula (2), after the parameter-related regularization term defined by the formula (1) is introduced, the parameter sets of the two depth networks with the same structure, which are respectively used for pedestrian re-identification and pedestrian gender classification, are interacted, and the correlation between the parameter sets corresponding to the two depth networks (Deep networks) with the same structure is increased, so that the deviation degree of the corresponding parameter sets in the optimization process is limited, and thus, the overfitting of the depth networks is avoided.
It is considered that pedestrians having the same identification information necessarily have the same sex information therebetween, whereas pedestrians having different sex information are unlikely to have the same identification information therebetween. This indicates that there is a correlation between the two tasks of pedestrian re-identification and pedestrian gender classification. By adopting the scheme, the invention simultaneously learns two depth networks with the same structure for pedestrian re-identification and pedestrian gender classification respectively. In the process of network learning, parameter sets in the two deep networks are constrained by using parameter-related regular terms, so that overfitting of the deep networks is avoided, and accuracy of pedestrian re-identification and pedestrian gender classification is improved.
Drawings
FIG. 1 is a schematic diagram of a method for pedestrian re-identification and pedestrian gender classification based on joint deep learning according to the present invention;
FIG. 2 is a schematic structural diagram of a CBLR unit of the present invention;
fig. 3 is a schematic structural diagram of the deep network of the present invention.
Detailed Description
The invention discloses a pedestrian re-identification and pedestrian gender classification method based on joint deep learning, which is characterized in that two depth networks with the same structure are constructed and are respectively used for pedestrian re-identification and pedestrian gender classification, and parameter sets of each layer in the two depth networks with the same structure are restrained by using parameter related regular terms in the network learning process, so that overfitting of the depth networks is avoided, and the accuracy of pedestrian re-identification and pedestrian gender classification is improved.
As a preferred embodiment, as shown in fig. 1, a method for re-identifying pedestrians and classifying gender of pedestrians based on joint deep learning includes the following steps:
step 1, constructing two depth networks with the same structure, wherein the two depth networks are respectively used for pedestrian re-identification and pedestrian gender classification;
as shown in fig. 2, for convenience of description, the present invention integrates the convolutional layer (Conv), the batch normalization layer (BatchNorm), and the leakage ReLU activation function into CBLR units, in which a filter of 3 × 3 size is uniformly used in all CBLR units, and the negative semi-axis slope of the leakage ReLU activation function is set to 0.15, and a 1-pixel zero padding operation is used. Based on the CBLR unit in fig. 2, the present embodiment employs the VGGNet structure to construct a deep network (i.e., the deep network in fig. 1). As shown in fig. 3, the constructed deep network includes four maximum pooling layers (MP1, MP2, MP3 and MP4), and 4 CBLR units (CBLR1-4), 3 CBLR units (CBLR5-7), 3 CBLR units (CBLR8-10) and 3 CBLR units (CBLR11-13) are respectively placed before MP1, MP2, MP3 and MP 4.
Step 2, utilizing parameter related regular terms to constrain parameter sets of each layer in two depth networks with the same structure;
assuming that the parameter sets corresponding to the two depth networks (Deep networks) with the same structure for pedestrian re-identification and pedestrian gender classification in step 1 are respectively H1＝{H11,H12,…,H1nH and2＝{H21,H22,…,H2nand n is the number of convolution layers in the deep network, and the parameter-related regular terms of the two are defined as follows:
Figure BDA0001678881330000041
wherein,
Figure BDA0001678881330000045
represents a dot product operation, H1iRepresents a parameter set H1Parameter of the ith convolution layer, H2iRepresents a parameter set H2The parameters of the ith convolutional layer in (c),
Figure BDA0001678881330000042
a Frobenius norm (Frobenius norm) representing a matrix; at present, in the process of deep network training, a gradient descent algorithm is commonly used for optimization, and H is subjected to gradient descent algorithm based1And H2Optimization is performed, i.e. the rule for updating parameters in back propagation, as follows:
Figure BDA0001678881330000043
wherein L is1And L2Representing the Softmax loss functions for pedestrian re-identification and pedestrian gender classification, respectively; lambda is more than or equal to 0 and is used for controlling parameter related regular term R (H)1,H2) The weight of (c); t +1 and t represent two adjacent iterations in the gradient descent method;
Figure BDA0001678881330000044
representing a partial derivation operation; as is apparent from the formula (2), after the parameter-related regularization term defined by the formula (1) is introduced, the parameter sets of the two depth networks with the same structure respectively used for pedestrian re-identification and pedestrian gender classification interact with each other, and the purpose is to increase the correlation between the parameter sets corresponding to the two depth networks (Deep networks) with the same structure, so as to limit the corresponding parameter setsThe degree of deviation in the optimization process, thereby avoiding overfitting of the deep network.
It is considered that pedestrians having the same identification information necessarily have the same sex information therebetween, whereas pedestrians having different sex information are unlikely to have the same identification information therebetween. This indicates that there is a correlation between the two tasks of pedestrian re-identification and pedestrian gender classification. By adopting the scheme, the invention simultaneously learns two depth networks with the same structure for pedestrian re-identification and pedestrian gender classification respectively. In the process of network learning, parameter sets in the two deep networks are constrained by using parameter-related regular terms, so that overfitting of the deep networks is avoided, and accuracy of pedestrian re-identification and pedestrian gender classification is improved.
The above description is only exemplary of the present invention and is not intended to limit the technical scope of the present invention, so that any minor modifications, equivalent changes and modifications made to the above exemplary embodiments according to the technical spirit of the present invention are within the technical scope of the present invention.

Claims (1)

1. A pedestrian re-identification and pedestrian gender classification method based on joint deep learning is characterized by comprising the following steps:
step 1, constructing two depth networks with the same structure, wherein the two depth networks are respectively used for pedestrian re-identification and pedestrian gender classification;
step 2, utilizing a parameter related regular term to constrain parameter sets of each layer in the two depth networks with the same structure;
the parameter sets corresponding to the two depth networks with the same structure for pedestrian re-identification and pedestrian gender classification in the step 1 are respectively H1＝{H11,H12,…,H1nH and2＝{H21,H22,…,H2nwhere n is the number of convolutional layers in the deep network, the parameters of the two are related to a regular term R (H)1,H2) The definition is as follows:
Figure FDA0003004215900000011
wherein,
Figure FDA0003004215900000012
represents a dot product operation, H1iRepresents a parameter set H1Parameter of the ith convolution layer, H2iRepresents a parameter set H2The parameters of the ith convolutional layer in (c),
Figure FDA0003004215900000013
a Frobenius norm representing a matrix;
gradient descent algorithm based pair H1And H2Optimization is carried out, namely the updating rule of the parameters in back propagation is as follows:
Figure FDA0003004215900000014
wherein L is1And L2Representing the Softmax loss functions for pedestrian re-identification and pedestrian gender classification, respectively; lambda is more than or equal to 0 and is used for controlling parameter related regular term R (H)1,H2) The weight of (c); t +1 and t represent two adjacent iterations in the gradient descent algorithm;
Figure FDA0003004215900000015
representing a partial derivative calculation.
CN201810541294.7A 2018-05-30 2018-05-30 Pedestrian re-identification and pedestrian gender classification method based on joint deep learning Active CN109063535B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810541294.7A CN109063535B (en) 2018-05-30 2018-05-30 Pedestrian re-identification and pedestrian gender classification method based on joint deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810541294.7A CN109063535B (en) 2018-05-30 2018-05-30 Pedestrian re-identification and pedestrian gender classification method based on joint deep learning

Publications (2)

Publication Number Publication Date
CN109063535A CN109063535A (en) 2018-12-21
CN109063535B true CN109063535B (en) 2021-09-28

Family

ID=64819809

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810541294.7A Active CN109063535B (en) 2018-05-30 2018-05-30 Pedestrian re-identification and pedestrian gender classification method based on joint deep learning

Country Status (1)

Country Link
CN (1) CN109063535B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111435452B (en) * 2019-01-11 2023-11-03 百度在线网络技术（北京）有限公司 Model training method, device, equipment and medium
CN110516569B (en) * 2019-08-15 2022-03-08 华侨大学 Pedestrian attribute identification method based on identity and non-identity attribute interactive learning
CN111695415B (en) * 2020-04-28 2024-04-12 平安科技（深圳）有限公司 Image recognition method and related equipment

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101290658A (en) * 2007-04-18 2008-10-22 中国科学院自动化研究所 Gender recognition method based on gait
KR20160017152A (en) * 2014-07-31 2016-02-16 경성대학교 산학협력단 Gender Classification Based on Binary Haar Cascade
CN105718873A (en) * 2016-01-18 2016-06-29 北京联合大学 People stream analysis method based on binocular vision
CN106960181A (en) * 2017-02-28 2017-07-18 中科唯实科技（北京）有限公司 A kind of pedestrian's attribute recognition approach based on RGBD data
CN106960176A (en) * 2017-02-22 2017-07-18 华侨大学 A kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion
CN107480631A (en) * 2017-08-11 2017-12-15 中山大学 A kind of pedestrian's recognition methods again based on visual angle association with depth e-learning

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160171546A1 (en) * 2013-12-14 2016-06-16 Fly Fleet LLC Situation Aware Travel Information

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101290658A (en) * 2007-04-18 2008-10-22 中国科学院自动化研究所 Gender recognition method based on gait
KR20160017152A (en) * 2014-07-31 2016-02-16 경성대학교 산학협력단 Gender Classification Based on Binary Haar Cascade
CN105718873A (en) * 2016-01-18 2016-06-29 北京联合大学 People stream analysis method based on binocular vision
CN106960176A (en) * 2017-02-22 2017-07-18 华侨大学 A kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion
CN106960181A (en) * 2017-02-28 2017-07-18 中科唯实科技（北京）有限公司 A kind of pedestrian's attribute recognition approach based on RGBD data
CN107480631A (en) * 2017-08-11 2017-12-15 中山大学 A kind of pedestrian's recognition methods again based on visual angle association with depth e-learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
《Multi-view joint learning network for pedestrian gender classification》;Lei Cai;《2017 International Symposium on Intelligent Signal Processing and Communication System》;20171109;23-27页 *

Also Published As

Publication number Publication date
CN109063535A (en) 2018-12-21

Similar Documents

Publication Publication Date Title
CN108010514B (en) Voice classification method based on deep neural network
CN108960140B (en) Pedestrian re-identification method based on multi-region feature extraction and fusion
CN110288030B (en) Image identification method, device and equipment based on lightweight network model
CN107369166B (en) Target tracking method and system based on multi-resolution neural network
CN109063535B (en) Pedestrian re-identification and pedestrian gender classification method based on joint deep learning
US20190228268A1 (en) Method and system for cell image segmentation using multi-stage convolutional neural networks
WO2017157183A1 (en) Automatic multi-threshold characteristic filtering method and apparatus
CN109993100B (en) Method for realizing facial expression recognition based on deep feature clustering
CN110175628A (en) A kind of compression algorithm based on automatic search with the neural networks pruning of knowledge distillation
US9129222B2 (en) Method and apparatus for a local competitive learning rule that leads to sparse connectivity
CN108985236B (en) Face recognition method based on deep separable convolution model
CN108875933B (en) Over-limit learning machine classification method and system for unsupervised sparse parameter learning
CN108985457B (en) Deep neural network structure design method inspired by optimization algorithm
EP3620982B1 (en) Sample processing method and device
Yang et al. Channel pruning based on convolutional neural network sensitivity
CN107451594B (en) Multi-view gait classification method based on multiple regression
US11475236B2 (en) Minimum-example/maximum-batch entropy-based clustering with neural networks
WO2022012668A1 (en) Training set processing method and apparatus
CN103065158A (en) Action identification method of independent subspace analysis (ISA) model based on relative gradient
CN117201122B (en) Unsupervised attribute network anomaly detection method and system based on view level graph comparison learning
CN109726662A (en) Multi-class human posture recognition method based on convolution sum circulation combination neural net
CN109886342A (en) Model training method and device based on machine learning
US20230041338A1 (en) Graph data processing method, device, and computer program product
CN115169465A (en) Transformer data anomaly detection method, device, equipment, medium and product
CN110543816B (en) Self-adaptive face image clustering method based on spectral clustering and reinforcement learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant