CN113496149B - Cross-view gait recognition method for subspace learning based on joint hierarchy selection - Google Patents
Cross-view gait recognition method for subspace learning based on joint hierarchy selection Download PDFInfo
- Publication number
- CN113496149B CN113496149B CN202010198551.9A CN202010198551A CN113496149B CN 113496149 B CN113496149 B CN 113496149B CN 202010198551 A CN202010198551 A CN 202010198551A CN 113496149 B CN113496149 B CN 113496149B
- Authority
- CN
- China
- Prior art keywords
- gait
- view
- target
- visual angle
- matrix
- 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
Links
- 230000005021 gait Effects 0.000 title claims abstract description 156
- 238000000034 method Methods 0.000 title claims abstract description 56
- 239000011159 matrix material Substances 0.000 claims abstract description 100
- 230000000007 visual effect Effects 0.000 claims abstract description 75
- 238000012360 testing method Methods 0.000 claims abstract description 56
- 238000012549 training Methods 0.000 claims abstract description 36
- 230000009977 dual effect Effects 0.000 claims abstract description 11
- 230000006870 function Effects 0.000 claims description 10
- 239000013598 vector Substances 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 7
- 238000000513 principal component analysis Methods 0.000 claims description 6
- 230000000717 retained effect Effects 0.000 claims description 6
- 230000000903 blocking effect Effects 0.000 claims description 5
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 4
- 238000009499 grossing Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 abstract description 6
- 230000000694 effects Effects 0.000 abstract description 4
- 230000002708 enhancing effect Effects 0.000 abstract 1
- 238000000638 solvent extraction Methods 0.000 abstract 1
- 238000005457 optimization Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 208000003443 Unconsciousness Diseases 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000007480 spreading Effects 0.000 description 1
- 210000003462 vein Anatomy 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
A cross-view gait recognition method of subspace learning based on joint hierarchy selection comprises the steps of firstly dividing gait samples of a target view and a registration view into a training set and a test set, simultaneously carrying out hierarchy partitioning on gait data of two views, respectively vectorizing multi-level gait energy image blocks, then carrying out feature selection and cascading; and then, simultaneously projecting the registration visual angle data and the target visual angle data to a public subspace, enhancing the relation between the registration visual angle data and the target visual angle data in a mode of constructing a cross-visual angle dual graph, performing effective feature selection and effective gait energy pattern block selection in the projection process, removing redundancy to form a registration sample set in the public subspace, projecting the test target visual angle data to the public subspace through a trained target visual angle projection matrix to form a target sample set in the public subspace, and performing gait recognition on the two sample sets in a Euclidean distance nearest neighbor mode. According to the invention, the gait data of the registered visual angle is introduced into the field of target visual angles, so that the cross-visual angle gait recognition effect is enhanced.
Description
Technical Field
The invention relates to a joint hierarchy selection-based cross-view gait recognition method for subspace learning, and belongs to the technical field of pattern recognition.
Background
Gait is one of the remotely perceptible biometric features, which may be captured from an unconscious and uncooperative subject, as compared to other biometric features (e.g., face, fingerprint, palm, vein, etc.). Gait recognition has the advantages of being non-contact, remote and not prone to camouflage. Research on gait recognition has been extensively conducted in recent decades. Recent gait recognition techniques are largely classified into two categories, namely model-based methods and motion-based methods. The model-based approach robustly extracts gait features and avoids noise interference problems. The motion-based method can represent the motion pattern of the human body without fitting model parameters. However, since factors such as clothes, shoes, conditions for carrying ornaments, ground surface for walking, walking speed, etc. vary, the recognition performance may be deteriorated. More seriously, the change of the visual angle will cause the appearance of the gait to change greatly, which brings great difficulty to the gait recognition. Therefore, cross-perspective gait recognition is very challenging and worthy of further investigation. One effect that hinders the good performance of gait recognition systems is that the appearance of gait is more susceptible to perspective than identity, especially when the perspective direction of the test gait is different from the direction of the registration gait. Conventional projection learning methods cannot solve this problem because these methods can only learn one projection matrix, and therefore for the same individual, it cannot convert cross-perspective gait features into similar gait features.
The existing cross-visual angle gait recognition methods are divided into two types: a feature description based method and a machine learning based method. Feature description based methods typically describe gait attributes that are robust to changes in viewing angle. The latter studies how to infer and understand the underlying relationship between trans-perspective gaits. The feature description based approach includes a 3D gait model and trajectory estimation. However, the 3D gait model requires high computational complexity and the trajectory estimation is very difficult from a front view, so that the feature description based method has certain limitations. However, methods based on machine learning can well overcome these problems. Most machine learning-based methods assume that the target perspective can be generated from the enrollment perspective, and that the target perspective and the enrollment perspective share common features in the common space. However, the method is complex in calculation, the dimensionality of the gait image converted into the vector is often up to ten thousand, the calculation amount is large, and the gait recognition efficiency is greatly reduced. Furthermore, since model-based methods are data-constrained, such methods suffer a large degradation in performance when the change in viewing angle is large. In addition, some depth models exist, but deep learning is data-driven in nature, and requires a large amount of training data, while the existing database has a relatively small data volume, which greatly limits the performance of the depth models. Therefore, aiming at the defects, the invention provides a cross-view gait recognition method based on subspace learning of joint hierarchy selection, and the recognition effect is improved.
Disclosure of Invention
The appearance of gait is more susceptible to perspective than identity, especially when the perspective direction of the test gait is different from the registration gait. Conventional projection learning methods cannot solve this problem because these methods can only learn one projection matrix, and therefore for the same individual, it cannot convert cross-perspective gait features into similar gait features. In order to effectively improve the cross-visual angle gait recognition effect, the invention adopts the registered visual angle gait data as the auxiliary domain to carry out cross-visual angle gait recognition. The invention provides a cross-view gait recognition method for subspace learning based on joint hierarchy selection, which can simultaneously perform block selection and feature selection on registered view angle and target view angle gait data of hierarchy blocks, remove the influence of redundant features, project the registered view angle and the target view angle gait data into a common subspace, and enable the gaits of different view angles of the same person to be close enough in the common subspace.
Summary of the invention:
a cross-perspective gait recognition method based on joint hierarchy selection subspace learning comprises the following steps: the method comprises the steps of hierarchical block division, data processing, subspace learning of joint hierarchical selection and gait recognition.
The technical scheme of the invention is as follows:
a cross-perspective gait recognition method based on joint hierarchy selection subspace learning comprises the following steps:
1) Firstly, dividing gait samples of a target visual angle and a registered visual angle into a training set and a testing set, wherein the training set comprises the gait samples of the target visual angle and the gait samples of the registered visual angle, and the testing set is the gait samples of the target visual angle;
acquiring gait energy images of two gait samples, namely a target view angle and a registration view angle, and dividing the gait energy images by adopting a hierarchical block division scheme;
2) In the training stage, after the gait samples of the target visual angle and the registered visual angle in the training set are divided into layers in blocks, vectorizing each divided gait energy image respectively, and performing feature selection to form a d-dimensional vector; thus, for example, using a gait sample, the cascade of features for all blocks of a registered perspective gait sample can be expressed as
Wherein d is a = d × n, d denotes a characteristic dimension of each block, n denotes a total number of blocks divided by hierarchical blocks, d a The total characteristic dimension of a cascaded gait sample of a registered visual angle is indicated; the registered perspective data matrix is therefore represented as +>
Wherein N is a Representing the number of gait samples of the registered visual angle; the subscript a hereafter indicates that the registered viewing angle is relevant. Similarly, the same-level block division is carried out on the target visual angle gait sample, the same gait energy image vectorization and the data processing of the characteristic vector cascade are carried out on the basis of the division, and finally the target visual angle data matrix is obtained>
Wherein d is b Representing a characteristic dimension, N, of a target perspective gait sample b Represents the number of target view angle samples, wherein d a ＝d b (ii) a The subscript b hereinafter indicates that the target viewing angle is relevant.
3) Subspace learning based on joint hierarchy selection is carried out on the training set target visual angle data matrix and the registration visual angle data matrix, the training set target visual angle data matrix and the registration visual angle data matrix are both projected to a public subspace, convergence conditions are met through multiple iterations, and an ideal target visual angle projection matrix U is obtained through respective learning b And registration view projection matrix U a (ii) a Projecting the matrix U from the target perspective b And the registered view projection matrix U a Forming a registration sample set in a public subspace;
4) For theAfter the gait samples of the target visual angle in the test set are divided into layers and blocks, each divided gait energy image is subjected to gait energy image vectorization and characteristic cascading data processing, and finally a target visual angle data matrix X in the test set is obtained b test (ii) a (superscript test refers to test set data, and no superscript refers to training data);
target view data matrix X for test set b test Performing subspace learning based on joint hierarchy selection, and testing the target visual angle data matrix X of the set b test The target visual angle projection matrix U obtained by learning in the step 3) b And projecting the data to the public subspace to form a target sample set in the public subspace, and performing gait recognition on the registration sample set in the public subspace and the target sample set in the public subspace by adopting a nearest neighbor mode of Euclidean distance.
Preferably, in step 1), the hierarchical blocking scheme includes gradually dense grids with different densities, and the gait energy map blocks with different sizes are generated through the grids with different densities, which ensures that the ideal blocks most related to the gait information are most likely and properly included.
Preferably, in step 2), the gait energy map after hierarchical segmentation is arranged in columns to generate column vectors; and carrying out feature selection by adopting a Principal Component Analysis (PCA) method to unify all vectorized features of each block into a consistent dimension, wherein the dimension of the minimum block is taken.
Preferably, in step 3), the subspace learning algorithm based on the joint hierarchy selection is mathematically described as follows:
wherein,
respectively representing a registered view projection matrix and a target view projection matrix, lambda 1 ,λ 2 And λ 3 Are all balance parameters, (| | U) a || 2,1 +||U b || 2,1 ) Selecting item for characteristics | | |, the calness 2,1 Represents L 21 A norm, minimizing the constraint can control the feature selection degree of the projection matrix for the original data; />
Selecting terms for blocks, | | - | luminance F The Frobenius norm is expressed, and the constraint is minimized for each block of the hierarchy blocks, so that the block can be selected; omega (U) a ,U b ) For a map regularization term, ->
Representing the public subspace, c representing the number of subjects (people) that the perspective gait belongs to, i.e. the identity tag, N a Number of gait samples representing registered view angle, N b Representing the number of gait samples of a target visual angle; in fact N is present a ＝N b The common subspace is defined as follows:
here, projecting the registered perspective data matrix and the target perspective data matrix to a common subspace helps to enhance the association of the two perspective data. At the same time, it is desirable to perform a correlation constraint for each projection matrix during the projection process so that the selection of blocks and basis features can be controlled. The algorithm automatically learns in the iterative optimization process, eliminates the blocks and the features with small contribution by minimizing the given objective function, and selects the blocks and the features with large contribution.
Preferably, in step 3), the regularization term Ω (U) of the graph of formula (1) a ,U b ) The cross-view dual graph can be constructed to further constrain the relationship between the registered view data and the target view data, on one hand, the common geometric structure of the registered view and the target view sample is reserved by utilizing the graph similarity relationship between views, and on the other hand, the cross-view dual graph further constrains the relationship between the registered view data and the target view sampleOn one hand, the manifold structure of each view angle can be well preserved by utilizing the intra-view graph similarity relationship, and firstly, the two graph similarity relationships are defined as follows:
graph similarity between views: if the samples of the two views belong to the same subject (person), there must be some similarity relationship, i.e., assume if the view data sample x is registered i And target perspective data sample y j Belonging to the same subject (person), or target perspective sample y i And registering view data sample x j Belonging to the same subject (person), if they all have similar relationship, then the similar relationship is retained to learn the public space, and the similarity matrix between the view angles is defined as follows:
graph similarity relationship within view angle: for the target view and the registered view, a certain manifold structure also exists in the view, that is, gait samples belonging to the same subject (person) and different angles in the view should have a certain similarity relationship, similarly, local structure information is retained by constructing a similarity graph and named as the view similarity relationship, and a registered view similarity matrix W is used for determining the similarity relationship between the gait samples and the local structure information a And the target view angle similarity matrix W b The definition is as follows:
where σ is a constraint factor, taken here as 1;
according to the similarity relationship between the two graphs, the similarity matrix of the final cross-view dual graph is integrated by the two graph relationships and defined as follows:
wherein β > 0 is a parameter for balancing the influence of the inter-view graph similarity relationship and the intra-view graph similarity relationship, and based on the cross-view dual graph similarity matrix, the following items can be defined:
wherein N = N a +N b Representing the total number of gait samples of the target view and the registered view,
l = D-W is the graph Laplacian matrix, D denotes the diagonal matrix, where the ith diagonal element is given by D ii ＝∑ j W ij Computing ； Considering that the data of the target view and the registered view are completely aligned, the data can be regarded as the data of two modalities under the same theme to establish a similar graph relation. Order U a ＝U 1 ，U b ＝U 2 ，X a ＝X 1 ,X b ＝X 2 Equation (8) can therefore be abbreviated as:
likewise, equation (1) can be abbreviated as follows:
further preferably, in step 3), the objective function of the formula (10) is iteratively solved by using a half-quadratic minimization method, and first, the objective function formula (10) can be transformed into:
wherein R is i ＝Diag(r i ) The t-th element thereof
Is defined as->
Theta is added as a smoothing term, and is a small integer to prevent the occurrence of a non-convergence condition when the denominator is 0. />
Each->
(g =1, 2.. Multidot.n) are each diagonal matrices, the kth diagonal element of which is defined as follows,
deriving 0 for the rewritten (11) yields the following equation:
the above equation (13) is rewritten as:
in practice, the optimization of the problem of equation (13) can be solved by solving a linear problem, updating U for the t +1 th iteration by the following equation i t+1
Obtaining a final target view projection matrix U b (U b ＝U 2 ) And a registered view projection matrix U a (U a ＝U 1 )。
Preferably, in the step 4), the training gait data and the testing gait data are measured by a nearest neighbor method based on euclidean distance, and a specific method for gait recognition by using the nearest neighbor method of euclidean distance is as follows:
by solving the linear optimization problem of the objective function, a trained registered visual angle projection matrix U can be obtained a Projection matrix U from the target view angle b We adopt U b And projecting the test target visual angle data matrix into the public subspace and comparing and identifying the data representation in the public subspace in a nearest neighbor mode. Firstly, a projection matrix is obtained, and M = [ U ] is set a T X a ,U b T X b ]Representing all training data in a common subspace, and at the same time, using U b T X b test Projecting the test target visual angle data matrix into a public subspace for identification;
given a test sample
N P Indicates the number of test samples, then any test sample after projection is->
Nearest neighbor comparison based on Euclidean distance is adopted>
Wherein, pi i Refers to the subject (person) of the ith training sample; the topic of the training sample closest to the person (i.e. the identity label of the person) is assigned to the sample under test to complete the identification of the identity.
The invention has the beneficial effects that:
the invention provides a cross-view gait recognition method for subspace learning based on joint hierarchy selection, which introduces relevant data in a registered view field into a target view field, fully constructs common characteristic factors between the registered view and the target view and enhances the relation; meanwhile, the same block selection and feature selection can be carried out on the registered visual angle and the target visual angle data in the projection process, so that the most effective gait information is extracted and transferred, the cross-visual angle gait recognition performance is improved, and the robustness is high.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2a is a schematic representation of a gait energy plot of a gait sample from a target perspective according to the invention;
FIG. 2b is a graphical representation of gait energy of a registered perspective gait sample in accordance with the invention;
FIG. 3 is a diagram illustrating an example of a hierarchical block division method according to the present invention.
Detailed Description
The present invention will be further described by way of examples, but not limited thereto, with reference to the accompanying drawings.
Example 1:
a cross-perspective gait recognition method based on joint hierarchy selection subspace learning comprises the following steps:
1) Firstly, dividing gait samples of a target visual angle and a registered visual angle into a training set and a testing set, wherein the training set comprises the gait samples of the target visual angle and the gait samples of the registered visual angle, and the testing set is the gait samples of the target visual angle;
the gait energy images of two gait samples, namely a target view angle and a registration view angle, are obtained, the gait energy images are divided by adopting a hierarchical blocking dividing scheme, the hierarchical blocking dividing scheme comprises gradually-dense grids with different densities, and gait energy image blocks with different sizes are generated through the grids with different densities, so that ideal blocks most related to gait information are most probably and properly contained, the dividing mode is shown in figure 3, and the actual operation can be specifically divided into different densities according to different data sets.
2) In the training stage, after layered blocking division is carried out on target view angle and registered view angle gait samples in a training set, vectorizing each divided gait energy image, carrying out feature selection to form a d-dimensional vector, and arranging and spreading each layered blocked gait energy image into column vectors in columns; and carrying out feature selection by adopting a Principal Component Analysis (PCA) method to unify all vectorized features of each block into a consistent dimension, wherein the dimension of the minimum block is taken. Thus, for example, using a gait sample, the cascade of features for all blocks of a registered perspective gait sample can be expressed as
Wherein d is a = d × n, d denotes a characteristic dimension of each block, n denotes a total number of blocks divided by hierarchical blocks, d a The total characteristic dimension of a cascaded gait sample of a registered visual angle is indicated; the registered view angle data matrix is therefore denoted as pick>
Wherein N is a Representing the number of gait samples of the registered visual angles; the subscript a hereafter indicates that the registered viewing angle is relevant. Similarly, the same-level block division is carried out on the target visual angle gait sample, the same gait energy image vectorization and the data processing of the characteristic vector cascade are carried out on the basis of the division, and finally the target visual angle data matrix is obtained>
Wherein d is b Representing a characteristic dimension, N, of a target perspective gait sample b Represents the number of target view angle samples, wherein d a ＝d b (ii) a The subscript b hereinafter indicates that the target viewing angle is relevant.
3) Subspace learning based on joint hierarchy selection is carried out on the training set target visual angle data matrix and the registration visual angle data matrix, the training set target visual angle data matrix and the registration visual angle data matrix are both projected to a public subspace, convergence conditions are met through multiple iterations, and an ideal target visual angle projection matrix U is obtained through respective learning b And registration view projection matrix U a (ii) a Projecting the matrix U from the target perspective b And registration view projection matrix U a Forming a registration sample set in a public subspace;
the subspace learning algorithm based on the joint hierarchy selection is mathematically described as follows:
wherein,
respectively representing a registered view projection matrix and a target view projection matrix, lambda 1 ,λ 2 And λ 3 Are all balance parameters, (| | U) a || 2,1 +||U b || 2,1 ) Selecting a term for a feature, | | lighter 2,1 Represents L 21 A norm, minimizing the constraint can control the feature selection degree of the projection matrix for the original data; />
Selecting terms for blocks, | | - | luminance F The Frobenius norm is expressed, and the constraint is minimized for each block of the hierarchy blocks, so that the block can be selected; omega (U) a ,U b ) For a map regularization term, ->
Representing a common subspace, c representing the number of subjects (persons) representing that the perspective gait belongs to a person, i.e. an identity tag, N a Number of gait samples representing registered view angle, N b Representing the number of gait samples of a target visual angle; in fact N is present a ＝N b The common subspace is defined as follows: />
Here, projecting the registered perspective data matrix and the target perspective data matrix into a common subspace helps to enhance the association of the two perspective data. At the same time, it is desirable to perform a correlation constraint for each projection matrix during the projection process so that the selection of blocks and basis features can be controlled. The algorithm automatically learns in the iterative optimization process, eliminates the blocks and the features with small contribution by minimizing the given objective function, and selects the blocks and the features with large contribution.
Regularization term omega (U) of formula (1) diagram a ,U b ) The cross-view dual graph can be constructed to further constrain the relationship between the registered view angle data and the target view angle data, on one hand, the common geometric structure of the registered view angle and the target view angle sample is reserved by utilizing the inter-view image similarity relationship, on the other hand, the manifold structure of each view angle can be well reserved by utilizing the intra-view image similarity relationship, and firstly, the two image similarity relationships are defined as follows:
graph similarity between views: if the samples of the two views belong to the same subject (person), there must be some similarity relationship, i.e., assume if the view data sample x is registered i And target perspective data sample y j Belonging to the same subject (person), or target perspective sample y i And registering view data sample x j Belonging to the same subject (person), if they all have similar relationship, then the similar relationship is retained to learn the public space, and the similarity matrix between the view angles is defined as follows:
similar relationship of images within view angle: for the target view and the registered view, a certain manifold structure also exists in the view, that is, gait samples belonging to the same subject (person) and different angles in the view should have a certain similarity relationship, similarly, local structure information is retained by constructing a similarity graph and named as view similarity relationship, and the registered viewAngular similarity matrix W a And the target view angle similarity matrix W b The definition is as follows:
wherein σ is a constraint factor, here 1;
according to the similarity relation of the two graphs, the similarity matrix of the final cross-view dual graph is integrated by the two graph relations, and the similarity matrix is defined as follows:
wherein β > 0 is a parameter for balancing the influence of inter-view graph similarity relationship and intra-view graph similarity relationship, and based on the cross-view dual graph similarity matrix, the following can be defined:
wherein N = N a +N b Representing the total number of target perspective and registered perspective gait samples,
l = D-W is the graph Laplacian matrix, D denotes the diagonal matrix, where the ith diagonal element is given by D ii ＝∑ j W ij Calculating; considering that the data of the target view and the registered view are completely aligned, the data can be regarded as the data of two modalities under the same theme to establish a similar graph relation. Let U a ＝U 1 ，U b ＝U 2 ，X a ＝X 1 ,X b ＝X 2 Equation (8) can therefore be abbreviated as:
likewise, equation (1) can be abbreviated as follows:
the objective function of equation (10) is solved iteratively in a half-quadratic minimization mode, and first the objective function equation (10) can be transformed into:
wherein R is i ＝Diag(r i ) The t-th element thereof
Is defined as->
Theta is added as a smoothing term, and takes a small integer to prevent the occurrence of a non-convergence condition when the denominator is 0. />
Each->
(g =1, 2...., n) are each diagonal matrices, whose k-th diagonal element is defined as follows,
deriving 0 for the rewritten (11) yields the following equation:
the above equation (13) is rewritten as:
in practice, the optimization of the problem of equation (13) can be solved by solving a linear problem, updating U for the t +1 th iteration by the following equation i t+1
Obtaining a final target view projection matrix U b (U b ＝U 2 ) And a registered view projection matrix U a (U a ＝U 1 )。
4) After the gait samples of the target visual angle in the test set are divided into layers and blocks, each divided gait energy image is subjected to gait energy image vectorization and characteristic cascade data processing, and finally a target visual angle data matrix X in the test set is obtained b test (ii) a (superscript test refers to test set data, and no superscript refers to training data);
target view data matrix X for test set b test Performing subspace learning based on joint hierarchy selection, and testing the target visual angle data matrix X of the set b test The target visual angle projection matrix U obtained by learning in the step 3) b And projecting the data to the public subspace to form a target sample set in the public subspace, and performing gait recognition on the registration sample set in the public subspace and the target sample set in the public subspace by adopting a nearest neighbor mode of Euclidean distance.
The training gait data and the testing gait data distance are measured through a nearest neighbor mode based on Euclidean distance, and a specific method for carrying out gait recognition through the nearest neighbor mode of Euclidean distance is as follows:
by solving the linear optimization problem of the objective function, a trained registered visual angle projection matrix U can be obtained a Projection matrix of target view angleU b We adopt U b And projecting the test target view angle data matrix into the public subspace and comparing and identifying the data representation in the public subspace in a nearest neighbor mode. Firstly, a projection matrix is obtained, and M = [ U ] a T X a ,U b T X b ]Representing all training data in a common subspace, and at the same time, using U b T X b test Projecting the test target visual angle data matrix into a public subspace in the same way and then identifying;
given a test sample
N P Indicates the number of test samples, then any test sample after projection is->
Nearest neighbor comparison based on Euclidean distance is adopted>
Wherein, pi i Subject (person) referring to the ith training sample; and (3) assigning the theme of the training sample (namely the identity label of a person) closest to the training sample to the sample to be tested to complete the identification of the identity. />
Claims (7)
1. A cross-perspective gait recognition method based on subspace learning of joint hierarchy selection is characterized by comprising the following steps:
1) Firstly, dividing gait samples of a target view angle and a registered view angle into a training set and a testing set, wherein the training set comprises the gait samples of the target view angle and the gait samples of the registered view angle, and the testing set is the gait samples of the target view angle;
acquiring gait energy maps of two gait samples, namely a target visual angle and a registered visual angle, and dividing the gait energy maps by adopting a hierarchical block division scheme;
2) In the training stage, after the gait samples of the target visual angle and the registered visual angle in the training set are divided into blocks in a hierarchy wayVectorizing each divided gait energy image, and performing feature selection to form a d-dimensional vector; the cascade of characteristics of all blocks of a registered perspective gait sample is expressed as
Wherein d is a = d × n, d denotes a characteristic dimension of each block, n denotes a total number of blocks divided by hierarchical blocks, d a The total characteristic dimension of a cascaded gait sample of a registered visual angle is indicated; the registered perspective data matrix is therefore represented as +>
Wherein N is a Representing the number of gait samples of the registered visual angle; similarly, the same-level block division is carried out on target visual angle gait samples, the same gait energy image vectorization and feature vector cascading data processing are carried out on the basis of the division, and finally a target visual angle data matrix->
Wherein d is b Representing a characteristic dimension, N, of a target perspective gait sample b Represents the number of target view angle samples, wherein d a ＝d b ；
3) Subspace learning based on joint hierarchy selection is carried out on the training set target visual angle data matrix and the registration visual angle data matrix, the training set target visual angle data matrix and the registration visual angle data matrix are both projected to a public subspace, convergence conditions are met through multiple iterations, and an ideal target visual angle projection matrix U is obtained through respective learning b And registration view projection matrix U a (ii) a Projecting the matrix U from the target perspective b And registration view projection matrix U a Forming a registration sample set in a public subspace;
4) After the gait samples of the target visual angle in the test set are divided into layers and blocks, each divided gait energy image is subjected to gait energy image vectorization and characteristic cascadeFinally obtaining a target visual angle data matrix X in the test set by data processing b test ；
Target perspective data matrix X for test set b test Performing subspace learning based on joint hierarchy selection, and testing the target visual angle data matrix X of the set b test The target visual angle projection matrix U obtained by learning in the step 3) b And projecting the data to the public subspace to form a target sample set in the public subspace, and performing gait recognition on the registration sample set in the public subspace and the target sample set in the public subspace by adopting a nearest neighbor mode of Euclidean distance.
2. The joint hierarchy selection-based cross-perspective gait recognition method for subspace learning according to claim 1, wherein in the step 1), the hierarchical blocking scheme comprises gradually dense grids with different densities, and gait energy map blocks with different sizes are generated through the grids with different densities.
3. The gait recognition method for subspace learning based on joint hierarchy selection according to claim 1, characterized in that in the step 2), each block of the layered and partitioned gait energy map is arranged and spread into column vectors; and carrying out feature selection by adopting a Principal Component Analysis (PCA) method to unify all vectorized features of each block into a consistent dimension, wherein the dimension of the minimum block is taken.
4. The joint hierarchy selection-based subspace learning cross-perspective gait recognition method according to claim 1, wherein in the step 3), the mathematical description of the joint hierarchy selection-based subspace learning algorithm is as follows:
wherein,
respectively representing a registered view projection matrix and a target view projection matrix, lambda 1 ,λ 2 And λ 3 Are all balance parameters, (| | U) a || 2,1 +||U b || 2,1 ) Selecting a term for a feature, | | lighter 2,1 Represents L 21 A norm, minimizing the constraint can control the feature selection degree of the projection matrix for the original data; />
Selecting terms for blocks, | | - | luminance F Representing a Frobenius norm for minimizing the constraint for each of the hierarchical chunks; omega (U) a ,U b ) For a graph regularization term>
Representing a common subspace, c representing the number of topics, i.e. identity tags, N a Number of gait samples representing registered view angle, N b Representing the number of gait samples of the target view angle; in fact N is present a ＝N b The common subspace is defined as follows:
5. the joint hierarchy selection-based cross-perspective gait recognition method for subspace learning according to claim 4, wherein in the step 3), the regularization term Ω (U) of formula (1) is a ,U b ) A cross-view dual graph can be constructed, on one hand, a common geometric structure of a registered view and a target view sample is reserved by using an inter-view graph similarity relation, and on the other hand, a manifold structure of each view is reserved by using an intra-view graph similarity relation, and firstly, the two graph similarity relations are defined as follows:
graph similarity between views: if the samples of two perspectives belong to the same subject (person), there must be some similarity relationship, i.e.Suppose if the view data sample x is registered i And target perspective data sample y j Belonging to the same subject (person), or target perspective sample y i And registering view data sample x j Belonging to the same subject (person), they all have a similar relation, the subject (person) indicates that the perspective gait belongs to a person, and the similar relation is retained to learn the public space, and define the similarity matrix between perspectives, which is defined as follows:
similar relationship of images within view angle: for the target view and the registered view, a certain manifold structure also exists in the view, that is, gait samples belonging to the same subject (person) and different angles in the view should have a certain similarity relationship, similarly, local structure information is retained by constructing a similarity graph and named as the view similarity relationship, and a registered view similarity matrix W is used for determining the similarity relationship between the gait samples and the local structure information a And the target view angle similarity matrix W b The definition is as follows:
wherein σ is a constraint factor, here 1;
according to the similarity relation of the two graphs, the similarity matrix of the final cross-view dual graph is integrated by the two graph relations, and the similarity matrix is defined as follows:
wherein β > 0 is a parameter for balancing the influence of the inter-view graph similarity relationship and the intra-view graph similarity relationship, and based on the cross-view dual graph similarity matrix, the following items can be defined:
wherein N = N a +N b Representing the total number of target perspective and registered perspective gait samples,
l = D-W is the graph Laplacian matrix, D denotes the diagonal matrix, where the ith diagonal element is given by D ii ＝∑ j W ij Calculating; let U a ＝U 1 ，U b ＝U 2 ，X a ＝X 1 ,X b ＝X 2 Equation (8) can therefore be abbreviated as:
likewise, equation (1) can be abbreviated as follows:
6. the joint hierarchy selection-based cross-perspective gait recognition method for subspace learning according to claim 5, wherein in the step 3), the objective function of the formula (10) is iteratively solved by using a semi-quadratic minimization method, and firstly, the objective function formula (10) can be transformed into:
wherein R is i ＝Diag(r i ) The tth element r thereof i t Is defined as
By adding theta as a smoothing term, theta being taken as a smaller integer, and>
each->
Are diagonal matrices, whose k-th diagonal elements are defined as follows,
deriving 0 for the rewritten (11) yields the following equation:
the above equation (13) is rewritten as:
update the U of the t +1 th iteration by the following equation i t+1
Obtaining a final target view projection matrix U b (U b ＝U 2 ) And a registered view projection matrix U a (U a ＝U 1 )。
7. The joint hierarchy selection-based subspace learning cross-perspective gait recognition method according to claim 1, wherein the distance between the training gait data and the test gait data is measured in the step 4) by a nearest neighbor method based on Euclidean distance, and a specific method for performing gait recognition by using the nearest neighbor method of Euclidean distance is as follows:
firstly, a projection matrix is obtained, and M = [ U ] is set a T X a ,U b T X b ]Representing all training data in a common subspace, and at the same time, using U b T X b test Projecting the test target visual angle data matrix into a public subspace for identification;
given a test sample
N P Indicating the number of test samples, then any test sample after projection is->
Nearest neighbor comparison based on Euclidean distance is adopted>
Wherein, pi i Refers to the subject of the ith training sample; and assigning the theme of the training sample closest to the training sample to the sample to be tested to complete identity recognition. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010198551.9A CN113496149B (en) | 2020-03-20 | 2020-03-20 | Cross-view gait recognition method for subspace learning based on joint hierarchy selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010198551.9A CN113496149B (en) | 2020-03-20 | 2020-03-20 | Cross-view gait recognition method for subspace learning based on joint hierarchy selection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113496149A CN113496149A (en) | 2021-10-12 |
| CN113496149B true CN113496149B (en) | 2023-04-07 |
Family
ID=77993612
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010198551.9A Active CN113496149B (en) | 2020-03-20 | 2020-03-20 | Cross-view gait recognition method for subspace learning based on joint hierarchy selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113496149B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114677755B (en) * | 2022-03-14 | 2024-04-26 | 山东大学 | Cross-view gait recognition method based on layering and blocking feature fusion |
| CN115205983B (en) * | 2022-09-14 | 2022-12-02 | 武汉大学 | Cross-perspective gait recognition method, system and equipment based on multi-feature aggregation |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9530052B1 (en) * | 2013-03-13 | 2016-12-27 | University Of Maryland | System and method for sensor adaptation in iris biometrics |
| CN108921114A (en) * | 2018-07-10 | 2018-11-30 | 福州大学 | The two visual angle gait recognition methods based on multi-non-linear locality preserving projections |
| CN109558834A (en) * | 2018-11-28 | 2019-04-02 | 福州大学 | A kind of multi-angle of view gait recognition method based on similarity study and kernel method |
| CN110619367A (en) * | 2019-09-20 | 2019-12-27 | 哈尔滨理工大学 | Joint low-rank constraint cross-view-angle discrimination subspace learning method and device |
-
2020
- 2020-03-20 CN CN202010198551.9A patent/CN113496149B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9530052B1 (en) * | 2013-03-13 | 2016-12-27 | University Of Maryland | System and method for sensor adaptation in iris biometrics |
| CN108921114A (en) * | 2018-07-10 | 2018-11-30 | 福州大学 | The two visual angle gait recognition methods based on multi-non-linear locality preserving projections |
| CN109558834A (en) * | 2018-11-28 | 2019-04-02 | 福州大学 | A kind of multi-angle of view gait recognition method based on similarity study and kernel method |
| CN110619367A (en) * | 2019-09-20 | 2019-12-27 | 哈尔滨理工大学 | Joint low-rank constraint cross-view-angle discrimination subspace learning method and device |
Non-Patent Citations (2)
| Title |
|---|
| xianye ben.Coupled partch Alignment for matching Cross-View Gaits.《IEEE》.2019,全文. * |
| 郭艳菊 ; 陈雷 ; 于明 ; 李锵 ; 陈国鹰.基于细菌趋药性算法的盲图像分离方法.光电子·激光.2011,第22卷(第3期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113496149A (en) | 2021-10-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Choy et al. | Deep global registration | |
| Hong et al. | Image-based three-dimensional human pose recovery by multiview locality-sensitive sparse retrieval | |
| Luo et al. | Hierarchical face parsing via deep learning | |
| CN108647583B (en) | Face recognition algorithm training method based on multi-target learning | |
| CN111652124A (en) | Construction method of human behavior recognition model based on graph convolution network | |
| Dambreville et al. | A framework for image segmentation using shape models and kernel space shape priors | |
| Prisacariu et al. | Simultaneous monocular 2D segmentation, 3D pose recovery and 3D reconstruction | |
| US9141871B2 (en) | Systems, methods, and software implementing affine-invariant feature detection implementing iterative searching of an affine space | |
| CN111738143B (en) | Pedestrian re-identification method based on expectation maximization | |
| McIntosh et al. | Medial-based deformable models in nonconvex shape-spaces for medical image segmentation | |
| CN113920472B (en) | Attention mechanism-based unsupervised target re-identification method and system | |
| Sedai et al. | A Gaussian process guided particle filter for tracking 3D human pose in video | |
| CN113496149B (en) | Cross-view gait recognition method for subspace learning based on joint hierarchy selection | |
| Abdelsamea et al. | A SOM-based Chan–Vese model for unsupervised image segmentation | |
| CN109829353A (en) | A kind of facial image stylizing method based on space constraint | |
| CN112232134A (en) | Human body posture estimation method based on hourglass network and attention mechanism | |
| Cai et al. | Detecting, grouping, and structure inference for invariant repetitive patterns in images | |
| CN107784284A (en) | Face identification method and system | |
| CN118135660A (en) | Cross-view gait recognition method for joint multi-view information bottleneck under view-angle deficiency condition | |
| Haindl et al. | Model-based texture segmentation | |
| CN110135435A (en) | A kind of conspicuousness detection method and device based on range learning system | |
| Tang et al. | General scheme of region competition based on scale space | |
| Tong et al. | Semi-supervised facial landmark annotation | |
| CN109815440A (en) | The Dimensionality Reduction method of the optimization of joint figure and projection study | |
| Langs et al. | Learning deformation and structure simultaneously: In situ endograft deformation analysis |
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 |