CN111783526A - Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment - Google Patents

Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment Download PDF

Info

Publication number
CN111783526A
CN111783526A CN202010434344.9A CN202010434344A CN111783526A CN 111783526 A CN111783526 A CN 111783526A CN 202010434344 A CN202010434344 A CN 202010434344A CN 111783526 A CN111783526 A CN 111783526A
Authority
CN
China
Prior art keywords
domain
pedestrian
representing
identity
attribute
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.)
Granted
Application number
CN202010434344.9A
Other languages
Chinese (zh)
Other versions
CN111783526B (en
Inventor
李华锋
庞健
严双林
欧洋汛
张亚飞
余正涛
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.)
Kunming University of Science and Technology
Original Assignee
Kunming University of Science and Technology
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 Kunming University of Science and Technology filed Critical Kunming University of Science and Technology
Priority to CN202010434344.9A priority Critical patent/CN111783526B/en
Publication of CN111783526A publication Critical patent/CN111783526A/en
Application granted granted Critical
Publication of CN111783526B publication Critical patent/CN111783526B/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
    • G06V40/20—Movements or behaviour, e.g. gesture recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries

Abstract

The invention provides a cross-domain pedestrian re-identification method by utilizing posture invariance and graph structure alignment, belonging to the field of computer vision. The invention provides a dictionary learning algorithm based on matrix decomposition to eliminate the influence of domain information and pedestrian attitude information among data sets on cross-domain pedestrian re-identification. Specifically, the method is divided into two parts: (1) decomposing original visual features into attitude invariant components, domain information components and interference information components based on the idea of matrix decomposition, and aiming at extracting visual components which are not influenced by domain information and pedestrian attitude information; (2) in order to further improve the generalization capability of the model, the relation between the posture invariant feature and the semantic attribute is established by introducing hypergraph structure alignment constraint so as to accurately predict the pedestrian attribute of the target data set at the later stage, and finally the pedestrian similarity measurement can be carried out by combining the posture invariant feature and the semantic attribute of the pedestrian so as to further improve the recognition performance.

Description

Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment
Technical Field
The invention relates to a cross-domain pedestrian re-identification method by utilizing posture invariance and graph structure alignment, belonging to the field of computer vision.
Background
With the rapid development of artificial intelligence, it is a need to apply pedestrian re-identification technology based on high-dimensional features to real life. Therefore, the scholars at home and abroad make a series of great research progresses in the aspect of pedestrian re-identification, and a plurality of methods are developed. Some methods design discriminative artifact features robust to changes in illumination, viewing angle, etc. for a target data set, or cluster unmarked target data. However, the performance of this kind of method is poor, mainly because the target data has no label, and the model is very difficult to mine the discriminant information. Some more advanced approaches view pedestrian re-identification as an unsupervised domain adaptation problem, which focuses on source domain to target domain knowledge migration. Compared to traditional unsupervised domain adaptation methods, pedestrian labels are completely different in the source domain and the target domain, and therefore the challenge is greater. Such methods still suffer from poor performance compared to supervised methods.
Disclosure of Invention
The invention aims to provide a cross-domain pedestrian re-recognition method by utilizing posture invariance and graph structure alignment, which is used for solving the problem that the existing pedestrian re-recognition algorithm is difficult to deploy; introducing an effective hypergraph structure alignment constraint, establishing a conversion relation between the posture invariant feature and the semantic attribute, and fully combining the advantages of the posture invariant feature and the semantic attribute to carry out joint measurement, wherein the specific flow is shown in figure 1. Compared with the existing method, the method can perform cross-domain re-recognition task, namely, the trained model is deployed to a brand-new camera network for pedestrian recognition.
A cross-domain pedestrian re-recognition method using posture invariance and graph structure alignment comprises the following steps:
1) defining data set variables and characteristics and attributes of pedestrians;
2) a design feature decomposition module for determining a target function containing a posture invariant component dictionary, a domain information component dictionary, an interference component dictionary and a conversion matrix;
3) designing a hypergraph structure alignment module by utilizing semantic attribute information;
4) designing a domain adaptation module capable of reducing domain offset;
5) merging the proposed loss functions into a final optimization function;
6) obtaining a dictionary and a conversion matrix by using an alternative optimization algorithm, thereby further obtaining a target domain data coding coefficient;
7) predicting the identity and the attribute of the pedestrian through the target domain coding coefficient;
8) and calculating the similarity between the pedestrians by using the cosine similarity and combining the predicted identity and the attribute.
The method comprises the following specific steps:
step 1, defining that K pedestrians exist in a source data set,
Figure BDA0002501673230000021
wherein
Figure BDA0002501673230000022
Representing the ith pedestrian feature of the source domain s, d representing the feature dimension,
Figure BDA0002501673230000023
representing the ith pedestrian attribute, c represents the attribute dimension,
Figure BDA0002501673230000024
indicates the i-th pedestrian label, NsIndicating the number of samples. Xs，As，YsRespectively representing a source domain feature set, a source domain attribute set and a source domain label set. Defining a target dataset
Figure BDA0002501673230000025
Contains N in totaltThe number of the samples is one,
Figure BDA0002501673230000026
the ith pedestrian feature of the target domain t is represented, and d represents the feature dimension. And using GOG pedestrian features at a feature level, and using the attributes of the existing data set as the attributes of the pedestrians.
Step 2: the following loss function Feature Decomposition term (FD) L is designedFDThe purpose of (1) decomposing a source domain feature set into an attitude invariant component, a domain component and an interference component:
Figure BDA0002501673230000027
wherein, VsDenotes the total number of source domain views, Xs,v,iAnd (3) representing the features of the ith identity at the v view angle in the training set s. Dp，Dd，DrRespectively representing an attitude invariant component dictionary, a domain information component dictionary, and an interference component dictionary. While
Figure BDA0002501673230000028
Represents Xs,v,iCorresponding to the coding coefficients of the three component dictionaries, respectively. I | · | purple wind*Represents the kernel norm, | ·| non-woven phosphor of the matrix2,1Expressing the structured sparse norm η, λ1，λ2A regularization parameter is represented. Wherein phi (D)r,Cp,Cr) Regular terms that promote domain separation are represented, specifically as follows:
Figure BDA0002501673230000029
wherein C isp，CrRepresenting the data set as a whole coding coefficients. Lambda [ alpha ]3And λ4Representing a regularization parameter. I and Q represent the identity matrix and identity matrix, respectively.
And step 3: in order to enhance the robustness and the domain invariance of the semantic attributes, the semantic attributes are introduced to assist cross-domain pedestrian re-identification. Loss function hypergraph structure alignment term (Hypergr)aph Structure Alignment，HSA)LHSAIs represented as follows:
Figure BDA0002501673230000031
firstly, a hypergraph G (X, E) is constructed through image samples of a source domain and the identity of a pedestrian, and comprises a group of vertexes
Figure BDA0002501673230000032
And a set of super edges
Figure BDA0002501673230000033
Wherein | NjI and | NrAnd | respectively represents the number of vertexes and super edges. For any given hypergraph, its hyper-edges can be easily converted into a correlation matrix
Figure BDA0002501673230000034
α1，α2，β1The representation of the hyper-parameter is,
Figure BDA0002501673230000035
representing two hypergraph laplacian regularizations, P and E represent linear transformation coefficient matrices, L-I-W represent hypergraph laplacian matrices,
Figure BDA0002501673230000036
a weight matrix representing a hypergraph to measure the degree of correlation between two vertices;
Figure BDA0002501673230000037
Dxand DeDiagonal matrices representing the degrees of the super edge and the degrees of the vertex, respectively. WeA diagonal matrix representing super-edge weights.
And 4, step 4: in order to solve the Domain deviation, a Domain Adaptation item is introduced, part of unlabeled data of the target Domain participates in the training of a characteristic decomposition model, and a Domain Adaptation (DA) L is lostDATo representThe following were used:
Figure BDA0002501673230000038
wherein, VtRepresents the total number of views of the target domain, NtRepresenting the number of samples, X, of the target domaint,v,iAnd (3) representing the pedestrian image feature sequence of the ith identity at the v view angle in the target data set t. While
Figure BDA0002501673230000039
Represents Xt,v,iCorresponding to three component dictionaries D respectivelyp，Dd，DrThe coding coefficients of (1). Lambda [ alpha ]2Is a regularization parameter. Finally, the entire objective function is represented as:
L＝LFD+LHSA+LDA(6)。
and 5: the proposed functions are then consolidated and merged, and the overall loss function L in step 4 can be expanded into the following form:
Figure BDA0002501673230000041
step 6: in the step 5, 9 variables need to be solved, each variable is solved by using an alternating iterative optimization algorithm, and other variables need to be fixed in the process of solving one variable. Obtaining an attitude invariant component dictionary D by solvingpDomain information component dictionary DdDictionary of interference components DrAnd transformation matrices P and E. With these dictionaries, the corresponding coding coefficients can be calculated by the following formula
Figure BDA0002501673230000042
Figure BDA0002501673230000043
ζ represents a regularization parameter.
And 7: when calculated, get
Figure BDA0002501673230000044
Then, using the transformation matrices P, E obtained in step 6, h can be obtained by equations (9) and (10)t,iAnd at,i：
Figure BDA0002501673230000045
Figure BDA0002501673230000046
In the above formula, ht,iAnd E can be considered constant by finding the optimum at,iThe minimum value is taken after the F norm of the right term is squared, and the a at the moment is obtainedt,i. With predicted identity representation h for the test samplet,iAnd semantic Attribute at,i。α2The regularization parameters are represented.
And 8: finally, the similarity achievement sim of the pedestrian image pair in the identity space and the semantic space can be respectively calculated through the cosine distance calculation formula of the equation (11)hAnd sima。
Figure BDA0002501673230000051
Wherein z isaAnd zbRespectively representing the current pedestrian identity expression vector and the semantic attribute vector and h obtained in the step 7t,iAnd at,iAre identical, except that zaAnd zbBroadly refers to the identity representation and semantic attributes of the current pedestrian, and ht,i，at,iAn identity representation and semantic attributes representing the ith pedestrian. Is constant 0.0000001. And (4) weighting and summing the similarity scores respectively obtained by the identity space and the semantic attribute space, and taking the weighted similarity score as a final pedestrian to perform similarity measurement on the similarity score.
simfinal＝τsima+(1-τ)simh(12)
Where τ > 0 represents the weight occupied by each space. In the present invention, τ is set to 0.2. Through the method, the similarity of the pedestrians in the target data set can be finally measured by using the solved variable.
The invention has the following beneficial effects:
(1) by the aid of the proposed decomposition model, influence of domain information and pedestrian posture information among data sets on cross-domain pedestrian re-identification is eliminated, and differences among different domains are reduced. The method is beneficial to the model to extract the more robust characteristics of the pedestrian in the real scene.
(2) By introducing an effective hypergraph structure alignment constraint, a conversion relation between the posture invariant feature and the semantic attribute is established, and the model is more discriminative for different pedestrians by combining a similarity measurement method performed by the two, for example, the appearances of two pedestrians are very similar, but the two pedestrians can be prevented from being identified as the same pedestrian through attribute information, so that misjudgment is avoided.
Drawings
FIG. 1 is a flow chart of a cross-domain pedestrian re-identification method using gesture invariance and graph structure alignment according to the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific examples.
Example 1: as shown in fig. 1, a cross-domain pedestrian re-identification method using posture invariance and graph structure alignment includes the following steps:
1) defining data set variables and characteristics and attributes of pedestrians;
2) a design feature decomposition module for determining a target function containing a posture invariant component dictionary, a domain information component dictionary, an interference component dictionary and a conversion matrix;
3) designing a hypergraph structure alignment module by utilizing semantic attribute information;
4) designing a domain adaptation module capable of reducing domain offset;
5) merging the proposed loss functions into a final optimization function;
6) obtaining a dictionary and a conversion matrix by using an alternative optimization algorithm, thereby further obtaining a target domain data coding coefficient;
7) predicting the identity and the attribute of the pedestrian through the target domain coding coefficient;
8) and calculating the similarity between the pedestrians by using the cosine similarity and combining the predicted identity and the attribute.
The method comprises the following specific steps:
step 1, defining that K pedestrians exist in a source data set,
Figure BDA0002501673230000061
wherein
Figure BDA0002501673230000062
Representing the ith pedestrian feature of the source domain s, d representing the feature dimension,
Figure BDA0002501673230000063
representing the ith pedestrian attribute, c represents the attribute dimension,
Figure BDA0002501673230000064
indicates the i-th pedestrian label, NsIndicating the number of samples. Xs，As，YsRespectively representing a source domain feature set, a source domain attribute set and a source domain label set. Defining a target dataset
Figure BDA0002501673230000065
Contains N in totaltThe number of the samples is one,
Figure BDA0002501673230000066
the ith pedestrian feature of the target domain t is represented, and d represents the feature dimension. And using GOG pedestrian features at a feature level, and using the attributes of the existing data set as the attributes of the pedestrians.
Step 2: the following loss function Feature Decomposition term (FD) L is designedFDIs to set the source domain features
Figure BDA0002501673230000067
Decomposition into pose-invariant components, domain components, interference components:
Figure BDA0002501673230000068
wherein, VsDenotes the total number of source domain views, Xs,v,iAnd (3) representing the features of the ith identity at the v view angle in the training set s. Dp，Dd，DrRespectively representing an attitude invariant component dictionary, a domain information component dictionary, and an interference component dictionary. While
Figure BDA0002501673230000069
Represents Xs,v,iCorresponding to the coding coefficients of the three component dictionaries, respectively. I | · | purple wind*Represents the kernel norm, | ·| non-woven phosphor of the matrix2,1Expressing the structured sparse norm η, λ1，λ2A regularization parameter is represented. Wherein Φ Dr,Cp,Cr) Regular terms that promote domain separation are represented, specifically as follows:
Figure BDA00025016732300000610
wherein C isp，CrRepresenting the data set as a whole coding coefficients. Lambda [ alpha ]3And λ4Representing a regularization parameter. I and Q represent the identity matrix and identity matrix, respectively.
And step 3: in order to enhance the robustness and the domain invariance of the semantic attributes, the semantic attributes are introduced to assist cross-domain pedestrian re-identification. Loss function Hypergraph Structure Alignment (HSA) LHSAIs represented as follows:
Figure BDA0002501673230000071
firstly, a hypergraph G (X, E) is constructed through image samples of a source domain and the identity of a pedestrian, and comprises a group of vertexes
Figure BDA0002501673230000072
And a set of super edges
Figure BDA0002501673230000073
Wherein | NjI and | NrAnd | respectively represents the number of vertexes and super edges. For any given hypergraph, its hyper-edges can be easily converted into a correlation matrix
Figure BDA0002501673230000074
α1，α2，β1The representation of the hyper-parameter is,
Figure BDA0002501673230000075
representing two hypergraph laplacian regularizations, P and E represent linear transformation coefficient matrices, L-I-W represent hypergraph laplacian matrices,
Figure BDA0002501673230000076
a weight matrix representing a hypergraph to measure the degree of correlation between two vertices;
Figure BDA0002501673230000077
Dxand DeDiagonal matrices representing the degrees of the super edge and the degrees of the vertex, respectively. WeA diagonal matrix representing super-edge weights.
And 4, step 4: in order to solve the Domain deviation, a Domain Adaptation item is introduced, part of unlabeled data of the target Domain participates in the training of a characteristic decomposition model, and a Domain Adaptation (DA) L is lostDAIs represented as follows:
Figure BDA0002501673230000078
wherein, VtRepresents the total number of views of the target domain, NtRepresenting the number of samples, X, of the target domaint,v,iAnd (3) representing the pedestrian image feature sequence of the ith identity at the v view angle in the target data set t. While
Figure BDA0002501673230000079
Represents Xt,v,iCorresponding to three component dictionaries D respectivelyp，Dd，DrThe coding coefficients of (1). Lambda [ alpha ]2Is a regularization parameter. Finally, the entire objective function is represented as:
L＝LFD+LHSA+LDA(6)。
and 5: the proposed functions are then consolidated and merged, and the overall loss function L in step 4 can be expanded into the following form:
Figure BDA0002501673230000081
step 6: in the step 5, 9 variables need to be solved, each variable is solved by using an alternating iterative optimization algorithm, and other variables need to be fixed in the process of solving one variable. Obtaining an attitude invariant component dictionary D by solvingpDomain information component dictionary DdDictionary of interference components DrAnd transformation matrices P and E. With these dictionaries, the corresponding coding coefficients can be calculated by the following formula
Figure BDA0002501673230000082
Figure BDA0002501673230000083
ζ represents a regularization parameter.
And 7: when calculated, get
Figure BDA0002501673230000084
Then, using the transformation matrices P, E obtained in step 6, h can be obtained by equations (9) and (10)t,iAnd at,i：
Figure BDA0002501673230000085
Figure BDA0002501673230000086
In the above formula, ht,iAnd E can be considered constant by finding the optimumat,iThe minimum value is taken after the F norm of the right term is squared, and the a at the moment is obtainedt,i. With predicted identity representation h for the test samplet,iAnd semantic Attribute at,i。α2The regularization parameters are represented.
And 8: finally, the similarity achievement sim of the pedestrian image pair in the identity space and the semantic space can be respectively calculated through the cosine distance calculation formula of the equation (11)hAnd sima。
Figure BDA0002501673230000091
Wherein z isaAnd zbRespectively representing the current pedestrian identity expression vector and the semantic attribute vector and h obtained in the step 7t,iAnd at,iAre identical, except that zaAnd zbBroadly refers to the identity representation and semantic attributes of the current pedestrian, and ht,i，at,iAn identity representation and semantic attributes representing the ith pedestrian. Is constant 0.0000001. And (4) weighting and summing the similarity scores respectively obtained by the identity space and the semantic attribute space, and taking the weighted similarity score as a final pedestrian to perform similarity measurement on the similarity score.
simfinal＝τsima+(1-τ)simh(12)
Where τ > 0 represents the weight occupied by each space. In the present invention, τ is set to 0.2. Through the method, the similarity of the pedestrians in the target data set can be finally measured by using the solved variable.
In the model proposed above, there are 11 parameters to be set, including dictionary Dp，Dd，DrAtom size d ofp,dd,drAnd the regularization term parameter λ1,λ2,λ3,λ4,α1,α2β ζ in the experiment, these parameters were set to d, respectivelyp＝600,dd＝180,dr＝180，λ1＝0.0001,λ2＝0.0001,λ3＝0.01,λ4＝1,α1＝0.1,α2＝0.1,β＝0.1,ζ＝0.1。
The GOG features are used as visual features of pedestrians, and standard semantic attributes which are already represented are used as attributes of the pedestrians. To demonstrate that the algorithm can be deployed in real life, experiments were performed on the VIPeR dataset. The data set contains two cameras, each capturing one image per person. The data set has various pedestrian attitude changes, as well as viewing angle and illumination changes. And taking prid2011 and grid as source data sets, and averagely dividing the model into training and testing. Training was repeated 10 times to obtain the average as the final performance. The comparison results are shown in table 1. The experiment proves that the method can directly deploy the trained model to the VIPer scene for recognition and keep good recognition rate.
TABLE 1 VIPeR data set
Figure BDA0002501673230000092
Figure BDA0002501673230000101
The invention also carries out experiments on the CUHK01 data set, the data set is collected from the campus scene of Chinese university in hong Kong, the cameras are respectively arranged in a teaching building and an outdoor scene, and the visual angle is wide step by step. Tests were performed with VIPeR as the source data set and CUHK01 as the target data set. The results are shown in table 2, which also shows the performance of other processes, from which it can be seen that the process achieves a relatively high performance.
TABLE 2 CUHK01 dataset
Figure BDA0002501673230000102
While the present invention has been described in detail with reference to the embodiments shown in the drawings, the present invention is not limited to the embodiments, and various changes can be made without departing from the spirit and scope of the present invention.

Claims (2)

1. A cross-domain pedestrian re-recognition method using posture invariance and graph structure alignment is characterized in that: the method comprises the following steps:
1) defining data set variables and characteristics and attributes of pedestrians;
2) a design feature decomposition module for determining a target function containing a posture invariant component dictionary, a domain information component dictionary, an interference component dictionary and a conversion matrix;
3) designing a hypergraph structure alignment module by utilizing semantic attribute information;
4) designing a domain adaptation module capable of reducing domain offset;
5) merging the proposed loss functions into a final optimization function;
6) obtaining a dictionary and a conversion matrix by using an alternative optimization algorithm, thereby further obtaining a target domain data coding coefficient;
7) predicting the identity and the attribute of the pedestrian through the target domain coding coefficient;
8) and calculating the similarity between the pedestrians by using the cosine similarity and combining the predicted identity and the attribute.
2. The method of claim 1, wherein the method comprises the following steps: the method comprises the following specific steps:
step 1, defining that K pedestrians exist in a source data set,
Figure FDA0002501673220000011
wherein
Figure FDA0002501673220000012
Representing the ith pedestrian feature of the source domain s, d representing the feature dimension,
Figure FDA0002501673220000013
representing the ith pedestrian attribute, c represents the attribute dimension,
Figure FDA0002501673220000014
indicates the i-th pedestrian label, NsDenotes the number of samples, Xs，As，YsRespectively representing a source domain feature set, a source domain attribute set, a source domain label set and defining a target data set
Figure FDA0002501673220000015
Contains N in totaltThe number of the samples is one,
Figure FDA0002501673220000016
representing the ith pedestrian feature of the target domain t, using the GOG pedestrian feature on the feature level, and using the attribute of the existing data set as the attribute of the pedestrian;
step 2: the loss function characteristic decomposition term L is designed as followsFDIs to set the source domain features
Figure FDA0002501673220000017
Decomposition into pose-invariant components, domain components, interference components:
Figure FDA0002501673220000018
wherein, VsDenotes the total number of source domain views, Xs,v,iFeatures representing the ith identity at the v view in the training set s, Dp，Dd，DrRespectively represent an attitude invariant component dictionary, a domain information component dictionary, and an interference component dictionary, and
Figure FDA0002501673220000021
represents Xs,v,iCorresponding to the coding coefficients of the three-component dictionary, | | · | | non-woven phosphor*Represents the kernel norm, | ·| non-woven phosphor of the matrix2,1Representing a structured sparse norm, η, λ1，λ2Represents a regularization parameter, where Φ (D)r,Cp,Cr) Regular terms that promote domain separation are represented, specifically as follows:
Figure FDA0002501673220000022
wherein C isp，CrRepresenting the overall coding coefficient, λ, of the data set3And λ4Representing a regular parameter, wherein I and Q respectively represent an identity matrix and an identity matrix;
and step 3: in order to enhance the robustness and the domain invariance of semantic attributes, the semantic attributes are introduced to assist cross-domain pedestrian re-identification, and a loss function hypergraph structure alignment item LHSAIs represented as follows:
Figure FDA0002501673220000023
firstly, a hypergraph G (X, E) is constructed through image samples of a source domain and the identity of a pedestrian, and comprises a group of vertexes
Figure FDA0002501673220000024
And a set of super edges
Figure FDA0002501673220000025
Wherein | NjI and | NrL respectively represents the number of vertexes and super edges, and for any given super graph, the super edge can be easily converted into a correlation matrix
Figure FDA0002501673220000026
α1，α2，β1Denotes a hyperparameter, tr (C)pLCpT) Representing two hypergraph laplacian regularizations, P and E represent linear transformation coefficient matrices, L-I-W represent hypergraph laplacian matrices,
Figure FDA0002501673220000027
a weight matrix representing a hypergraph to measure the degree of correlation between two vertices;
Figure FDA0002501673220000028
Dxand DeDiagonal matrices, W, representing the degrees of the super-edges and the degrees of the vertices, respectivelyeA diagonal matrix representing super-edge weights;
and 4, step 4: in order to solve the domain deviation, a domain adaptation term is introduced, part of unlabeled data of the target domain participates in the training of the characteristic decomposition model, and the function domain adaptation term L is lostDAIs represented as follows:
Figure FDA0002501673220000029
Figure FDA00025016732200000210
Figure FDA0002501673220000031
wherein, VtRepresents the total number of views of the target domain, NtRepresenting the number of samples, X, of the target domaint,v,iA sequence of pedestrian image features representing the ith identity at the v view angle in the target data set t, and
Figure FDA0002501673220000032
represents Xt,v,iCorresponding to three component dictionaries D respectivelyp，Dd，DrA coding coefficient of (a)2To regularize the parameters, finally, the entire objective function is expressed as:
L＝LFD+LHSA+LDA(6)
and 5: the proposed functions are then consolidated and merged, and the overall loss function L in step 4 can be expanded into the following form:
Figure FDA0002501673220000033
Figure FDA0002501673220000034
step 6: in the step 5, 9 variables need to be solved, each variable is solved by using an alternative iterative optimization algorithm, in the process, one variable needs to be fixed with other variables, and the attitude invariant component dictionary D is obtained by solvingpDomain information component dictionary DdDictionary of interference components DrAnd transformation matrices P and E, with these dictionaries, whose corresponding coding coefficients can be calculated by the following formula
Figure FDA0002501673220000035
Figure FDA0002501673220000036
ζ represents a regularization parameter;
and 7: when calculated, get
Figure FDA0002501673220000037
Then, using the transformation matrices P, E obtained in step 6, h can be obtained by equations (9) and (10)t,iAnd at,i：
Figure FDA0002501673220000038
Figure FDA0002501673220000039
In the above formula, ht,iAnd E can be considered constant by finding the optimum at,iThe minimum value is taken after the F norm of the right term is squared, and the a at the moment is obtainedt,iFor the test sample, there is a predicted identity representation ht,iAnd semantic Attribute at,i，α2Representing a regularization parameter;
and 8: finally, the similarity between the pedestrian image pair in the identity space and the semantic space can be respectively calculated through the cosine distance calculation formula of equation (11)Achievement simhAnd sima，
Figure FDA0002501673220000041
Wherein z isaAnd zbRespectively representing the current pedestrian identity expression vector and the semantic attribute vector and h obtained in the step 7t,iAnd at,iAre identical, except that zaAnd zbBroadly refers to the identity representation and semantic attributes of the current pedestrian, and ht,i，at,iThe identity representation and the semantic attribute of the ith pedestrian are represented and are constant 0.0000001, the similarity scores obtained from the identity space and the semantic attribute space are weighted and summed, and the weighted similarity score is taken as the final pedestrian to perform similarity measurement on the similarity score:
simfinal＝τsima+(1-τ)simh(12)
wherein tau > 0 represents the weight occupied by each space, and tau is set to be 0.2 in the invention, and finally the similarity of pedestrians in the target data set can be measured by using the solved variable.
CN202010434344.9A 2020-05-21 2020-05-21 Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment Active CN111783526B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010434344.9A CN111783526B (en) 2020-05-21 2020-05-21 Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010434344.9A CN111783526B (en) 2020-05-21 2020-05-21 Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment

Publications (2)

Publication Number Publication Date
CN111783526A true CN111783526A (en) 2020-10-16
CN111783526B CN111783526B (en) 2022-08-05

Family

ID=72754356

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010434344.9A Active CN111783526B (en) 2020-05-21 2020-05-21 Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment

Country Status (1)

Country Link
CN (1) CN111783526B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112528756A (en) * 2020-11-20 2021-03-19 华南理工大学 Unsupervised pedestrian re-identification method based on heteromorphic image

Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020068842A1 (en) * 1999-01-29 2002-06-06 Brundage Scott R. Blending of economic, reduced oxygen, winter gasoline
US20030019787A1 (en) * 2001-05-15 2003-01-30 Doherty Helen M. Reduced emissions transportation fuel
CN104268583A (en) * 2014-09-16 2015-01-07 上海交通大学 Pedestrian re-recognition method and system based on color area features
CN105160312A (en) * 2015-08-27 2015-12-16 南京信息工程大学 Recommendation method for star face make up based on facial similarity match
CN106778464A (en) * 2016-11-09 2017-05-31 深圳市深网视界科技有限公司 A kind of pedestrian based on deep learning recognition methods and device again
CN107563328A (en) * 2017-09-01 2018-01-09 广州智慧城市发展研究院 A kind of face identification method and system based under complex environment
CN109101865A (en) * 2018-05-31 2018-12-28 湖北工业大学 A kind of recognition methods again of the pedestrian based on deep learning
CN109190470A (en) * 2018-07-27 2019-01-11 北京市商汤科技开发有限公司 Pedestrian recognition methods and device again
CN109214442A (en) * 2018-08-24 2019-01-15 昆明理工大学 A kind of pedestrian's weight recognizer constrained based on list and identity coherence
CN110046870A (en) * 2019-04-16 2019-07-23 中山大学 A kind of method and system of registering based on geographical location and face characteristic
CN110349240A (en) * 2019-06-26 2019-10-18 华中科技大学 It is a kind of it is unsupervised under based on posture conversion pedestrian's picture synthetic method and system
CN110826417A (en) * 2019-10-12 2020-02-21 昆明理工大学 Cross-view pedestrian re-identification method based on discriminant dictionary learning
CN111177447A (en) * 2019-12-26 2020-05-19 南京大学 Pedestrian image identification method based on depth network model

Patent Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020068842A1 (en) * 1999-01-29 2002-06-06 Brundage Scott R. Blending of economic, reduced oxygen, winter gasoline
US20030019787A1 (en) * 2001-05-15 2003-01-30 Doherty Helen M. Reduced emissions transportation fuel
CN104268583A (en) * 2014-09-16 2015-01-07 上海交通大学 Pedestrian re-recognition method and system based on color area features
CN105160312A (en) * 2015-08-27 2015-12-16 南京信息工程大学 Recommendation method for star face make up based on facial similarity match
CN106778464A (en) * 2016-11-09 2017-05-31 深圳市深网视界科技有限公司 A kind of pedestrian based on deep learning recognition methods and device again
CN107563328A (en) * 2017-09-01 2018-01-09 广州智慧城市发展研究院 A kind of face identification method and system based under complex environment
CN109101865A (en) * 2018-05-31 2018-12-28 湖北工业大学 A kind of recognition methods again of the pedestrian based on deep learning
CN109190470A (en) * 2018-07-27 2019-01-11 北京市商汤科技开发有限公司 Pedestrian recognition methods and device again
CN109214442A (en) * 2018-08-24 2019-01-15 昆明理工大学 A kind of pedestrian's weight recognizer constrained based on list and identity coherence
CN110046870A (en) * 2019-04-16 2019-07-23 中山大学 A kind of method and system of registering based on geographical location and face characteristic
CN110349240A (en) * 2019-06-26 2019-10-18 华中科技大学 It is a kind of it is unsupervised under based on posture conversion pedestrian's picture synthetic method and system
CN110826417A (en) * 2019-10-12 2020-02-21 昆明理工大学 Cross-view pedestrian re-identification method based on discriminant dictionary learning
CN111177447A (en) * 2019-12-26 2020-05-19 南京大学 Pedestrian image identification method based on depth network model

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
YANG E: "Shared predictive cross-modal deep quantization", 《IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS》, 31 December 2018 (2018-12-31), pages 5292 - 5303 *
张耿宁等: "基于特征融合与核局部Fisher判别分析的行人重识别", 《计算机应用》, no. 09, 10 September 2016 (2016-09-10), pages 2597 - 2600 *
詹敏等: "基于自适应度量学习的行人再识别", 《电脑知识与技术》, no. 10, 5 April 2017 (2017-04-05), pages 159 - 164 *
郑伟诗等: "非对称行人重识别:跨摄像机持续行人追踪", 《中国科学:信息科学》, no. 05, 20 May 2018 (2018-05-20), pages 545 - 563 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112528756A (en) * 2020-11-20 2021-03-19 华南理工大学 Unsupervised pedestrian re-identification method based on heteromorphic image
CN112528756B (en) * 2020-11-20 2023-06-23 华南理工大学 Unsupervised pedestrian re-identification method based on different composition

Also Published As

Publication number Publication date
CN111783526B (en) 2022-08-05

Similar Documents

Publication Publication Date Title
CN106919920B (en) Scene recognition method based on convolution characteristics and space vision bag-of-words model
CN111160533B (en) Neural network acceleration method based on cross-resolution knowledge distillation
CN105160310A (en) 3D (three-dimensional) convolutional neural network based human body behavior recognition method
CN109299707A (en) A kind of unsupervised pedestrian recognition methods again based on fuzzy depth cluster
CN107463954B (en) A kind of template matching recognition methods obscuring different spectrogram picture
CN109255289B (en) Cross-aging face recognition method based on unified generation model
CN108256486B (en) Image identification method and device based on nonnegative low-rank and semi-supervised learning
CN111339343A (en) Image retrieval method, device, storage medium and equipment
CN107862680B (en) Target tracking optimization method based on correlation filter
CN109635726B (en) Landslide identification method based on combination of symmetric deep network and multi-scale pooling
CN112257741B (en) Method for detecting generative anti-false picture based on complex neural network
CN105760879A (en) Fourier-Mellin transform-based image geometric matching method
CN111783521A (en) Pedestrian re-identification method based on low-rank prior guidance and based on domain invariant information separation
CN109492570A (en) A kind of SAR image target recognition method based on multiple dimensioned rarefaction representation
CN111695455B (en) Low-resolution face recognition method based on coupling discrimination manifold alignment
CN103714331A (en) Facial expression feature extraction method based on point distribution model
Pratama et al. Face recognition for presence system by using residual networks-50 architecture
CN110458235A (en) Movement posture similarity comparison method in a kind of video
CN109766748A (en) A kind of pedestrian based on projective transformation and dictionary learning knows method for distinguishing again
CN110135435B (en) Saliency detection method and device based on breadth learning system
CN114972904A (en) Zero sample knowledge distillation method and system based on triple loss resistance
CN111783526B (en) Cross-domain pedestrian re-identification method using posture invariance and graph structure alignment
Liu et al. Loop closure detection based on improved hybrid deep learning architecture
CN108960013B (en) Pedestrian re-identification method and device
Lei et al. Student action recognition based on multiple features

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