CN104166981A - Human body movement learning method based on multigraph expression - Google Patents

Abstract

The invention provides a human body movement learning method based on multigraph expression. The method includes the steps of: dividing a characteristic space into several mutually independent subspaces, and representing the subspaces through a plurality of graphs; generating different incidence matrixes from each subspace, and performing spectrum insertion on each subspace; zooming the insertion and joining the insertion together, providing a single expression for each point of information, so as to generate a feature gather spectrum multigraph (FGSM), the FGSM being capable of achieving the least data loss from the original feature space; and applying the FGSM to algorithms of clustering, information retrieval and recognition, so that human body movement learning is performed. The human body movement learning method based on multigraph expression which is provided by the invention facilitates anti-counterfeiting verification operation of a product, improves verification efficiency, and is practical.

Description

Human action learning method based on many graph expressions

Technical field

The present invention relates to technical field of computer vision, particularly a kind of human action learning method based on many graph expressions.

Background technology

Algorithm based on graph theory is the potential structure that a kind of strong mode is used for excavating data, thereby and improves the performance of non-supervisory and semi-supervised task.For explanation this point, associate current three kinds of methods based on graph theory the most successful: spectral clustering method can be used to find the cluster of arbitrary structures, and popular ranking method is successfully applied to information retrieval task, and Laplce's characteristic pattern (LE) is applied to dimensionality reduction.As a rule, use the method based on graph theory to find the potential structure of high dimensional data, so can improve the accuracy of non-supervisory and semi-supervised task.

The first step of these methods based on graph theory is to produce an incidence matrix, W, and its meaning is the degree of association between every pair of point in data group X.Word bag (BoF) histogram of studying emphatically for the present invention, is applied in every couple of point set xi with a thermonuclear, and the card side of xj ∈ X is apart from χ ²on.Or, use histogram intersection method.After W produces, then it is carried out to multiple operation obtain net result.In some method, LE for example, W can carry out rarefaction by the method for kNN or ∈ neighborhood, but other method is all by completely connected method.

$\mathrm{Wij}=\exp (-\frac{x^2(\mathrm{xi},\mathrm{xj})}{{\mathrm{\σ}}^2})---\left(1\right)$

But, all representations based on graph theory have a common drawback: when producing W from X, if every pair of point only produces a single relating value, can from primitive character space, have obvious information dropout.Particularly in small-sized data group, the primitive character space of (because every row W be low-dimensional) or data group has when more high-dimensional, and information can be lost especially seriously.

High dimensional data as histogram, produces a single figure by a single incidence matrix and be often not enough to obtain the entire infrastructure presenting in primitive character space.When reality represents picture or video, in a histogram, may have a plurality of statistically separate features.So, single figure also just cannot distinguish these features.

Therefore, be necessary to propose a kind of by building polygraphic method for expressing, thereby the information dropout while greatly reducing data representation and has retained the structural information of each histogram inside.

Summary of the invention

Object of the present invention aims to provide a kind of human action learning method based on many graph expressions, by building polygraphic method to make up the blank of current diagram technology.

Human action learning method based on many graph expressions provided by the invention comprises: feature space is divided into several separate subspaces, and represents described subspace by a plurality of figure; From every sub spaces, produce different incidence matrix, and compose embedding in every sub spaces; These are embedded to convergent-divergent and link together, for each information point obtains a single representation, with generating feature set spectrum multigraph FGSM, described FGSM can obtain from original feature space minimum loss of data; FGSM is applied to cluster, information retrieval and recognizer, to carry out human action study.

Further, described feature space is higher dimensional space.

Further, described characteristic set can be divided into several disjoint subsets, and between all subsets, should have the independence of height and between its inside, should have the dependence of high altitude.

Further, described feature space is divided into several separate subspaces, comprises:

On the associated diagram of Xi Er baud-Shi Mite independence criterion HSIC value, calculate every pair of feature, to obtain by frequency spectrum cluster, described HSIC can obtain whole non-linear dependence of two stochastic variable x and y.

Further, describedly on the associated diagram of Xi Er baud-Shi Mite independence criterion HSIC value, calculate every pair of feature, comprising:

By following steps, from limited (xi, yi) tuple, by experimental formula, estimate:

${\mathrm{\ρ}}_n(x,y)=\frac{1}{{(1-n)}^2}\mathrm{tr}\left({\mathrm{Hk}}_x{\mathrm{Hk}}_y\right)---\left(2\right)$

Wherein, Hij=δ _ij-n ^-1, Hi refers to HKx, and Hj refers to HKy, and Kx and Ky are respectively the vector products of vector x and y, and n is the number of sampling.

Further, describedly on the associated diagram of Xi Er baud-Shi Mite independence criterion HSIC value, calculate every pair of feature, also comprise:

Because calculating K x and Ky will spend O (n ²) time and space, by incomplete Cholesky factorization, find Lx and Ly, like this, Kx and Ky Kx '=LxLx ^tand Ky '=LyLy ^testimate, afterwards, approximate HSIC is just by calculating as follows:

ρ _n(x,y)＝tr((Lx ^TLy))((Ly ^TLx)) (3)

This calculates at O (nf ²) time in complete, wherein f is to be the selected columns of L, and in very large data group, HSIC estimation is taked to raise the efficiency from the mode of original crowd evacuation sampling, the loss of a part of accuracy is accepted in such sampling because the HSIC of estimation and real HSIC be with speed approach.

Further, also comprise:

After obtaining m disjoint subspace, according to every sub spaces, find respectively the embedding of the data group of m.For every sub spaces sm, with associated diagram W of following Formula:

Wij＝sum(min(Pm(xi),Pm(xj))) (4)

Wherein Pm (x) is mapped to x the formula of m sub spaces.Compared with using kNN neighborhood or ∈-Neighborhood Graph, what used here is typical Laplce's characteristic pattern, and W builds with complete connection layout.

Human action learning method based on many graph expressions provided by the invention, by building polygraphic method in order to make up the blank of current diagram technology, and each is heavy corresponding to the different characteristic in original image or data, its reason is, these multi-medium datas of picture or video are comprised of multiple different feature, therefore the relation that will fully obtain between these parts just need to represent with a plurality of figure, thereby makes up the various defects of existing figure method for expressing.

The aspect that the present invention is additional and advantage in the following description part provide, and these will become obviously from the following description, or recognize by practice of the present invention.

Accompanying drawing explanation

Fig. 1 has expressed effect schematic diagram according to an embodiment of the present invention;

Fig. 2 has expressed schematic flow sheet according to an embodiment of the present invention.

Embodiment

Describe embodiments of the present invention below in detail, the example of described embodiment is shown in the drawings, and wherein same or similar label represents same or similar assembly or has the assembly of identical or similar functions from start to finish.Below by the embodiment being described with reference to the drawings, be exemplary, only for explaining the present invention, and can not be interpreted as limitation of the present invention.

Those skilled in the art of the present technique understand, unless specially statement, singulative used herein " ", " one ", " described " and " being somebody's turn to do " also can comprise plural form.Should be further understood that, the wording of using in instructions of the present invention " comprises " and refers to and have described feature, integer, step, operation, assembly and/or assembly, but do not get rid of, do not exist or adds one or more further features, integer, step, operation, assembly, assembly and/or their group.Should be appreciated that, when claiming assembly to be " connected " or " coupling " arrives another assembly, it is directly connected or coupled to other assembly, or also has intermediate module.In addition, " connection " used herein or " coupling " comprise wireless connections or couple.Wording "and/or" used herein comprises arbitrary unit of listing item and all combinations that one or more is associated.

Those skilled in the art of the present technique understand, unless otherwise defined, all terms used herein (comprising technical term and scientific terminology) have with the present invention under the identical meaning of the general understanding of those of ordinary skill in field.Should also be understood that such as those terms that define in general dictionary and should be understood to have the consistent meaning of meaning in the context with prior art, unless and definition as here, can not explain by idealized or too formal implication.

The feature of this technology is to produce multigraph from the independent subset of feature space.First, feature space is divided into several separate subspaces, and represents them with a plurality of figure.Then, from every sub spaces, produce different incidence matrix.Afterwards, in every sub spaces, compose embedding.Finally, these are embedded to convergent-divergent and link together, for each information point obtains a single representation.This represents to be exactly the Feature Combination spectrum multigraph (FGSM) of mentioning before.Compare general spectrum embedding grammar, FGSM can obtain from original feature space minimum loss of data.Finally, FGSM can be applied to cluster, information retrieval and identification scheduling algorithm and embody its superiority.

Fig. 1 has expressed the effect schematic diagram of implementing according to the present invention.As shown in Figure 1, the histogrammic vertical bar of some word bag is more relevant than other parts for an ingredient in video.In the example of these simple six entries, entry 1 and 2 with strong being associated above the waist, what entry 3 and semi-finals were strong is associated with background, entry 5 and 6 is strong to be associated with the lower part of the body.The height segmenting of this three part makes this histogram represent with FGSM can be more desirable.

Fig. 2 has expressed the schematic flow sheet of implementing according to the present invention.As shown in Figure 2, the feature of this technology is to produce multigraph from the independent subset of feature space.First, feature space is divided into several separate subspaces, and represents them with a plurality of figure.Then, from every sub spaces, produce different incidence matrix.Afterwards, in every sub spaces, compose embedding.Finally, these are embedded to convergent-divergent and link together, for each information point obtains a single representation.This represents to be exactly the Feature Combination spectrum multigraph (FGSM) of mentioning before.Compare general spectrum embedding grammar, FGSM can obtain from original feature space minimum loss of data.Finally, FGSM can be applied to cluster, information retrieval and identification scheduling algorithm and embody its superiority.

When FGSM is applied to data group, the feature space of original data set should have two specific characters: 1) feature space must be higher-dimension, 2) characteristic set must be able to be divided into several disjoint subsets, and between all subsets, should have height independence and between its inside, should have height dependence.The histogram that why proposes these characteristics to be applied to video represents it is because the locality of feature is because histogrammic each entry is to be all associated from each different formation of video in essence.This concept can be with reference to the description of upper figure.

In the general learning method based on graph theory, much from the information in primitive character space, can in the process that creates associated diagram, lose, yet FGSM overcomes this problem.Because it finds a plurality of independent visual angle (namely subspace) in original expression, and produces respectively an associated diagram for each visual angle.

Algorithm 1:FGSM-multigraph represents

Data:

The histogram of an X-mono-data group represents

The characteristic number that m-subspace will be looked for

The proper vector number of k-each proper subspace

Result: the multigraph of a Y-mono-data group represents

1. calculate each row HSIC incidence matrix between any two in X, wherein (formula 3)

2. use the method for Ng to W spectral clustering, find m feature clustering: C1..Cm;

3. defining equation P1..Pm is in order to be mapped to X in proper subspace C1..Cm;

4.fori ← 1 is to m;

5. calculate T ← Pi (X);

6. calculate wjk ← sum (min (T _j, T _k));

7. calculate S ← D ^-1/2lD ^-1/2, wherein to also have D be the capable sum of l that diagonal matrix and Du equal W to L ← D-W;

8. k proper vector e1..ek before finding in S, and their are listed as to being connected: Ei ← [e1..ek];

By every column criterionization of Ei to and be 1;

10. find λ _ias the mean distance between every row Mi: Mi: λ i ← σ (dist (Ei));

11. row obtain Y to the E1..Em connecting after convergent-divergent:

$Y\←[\left({\mathrm{\λ}}_{-1}^1{E}_1\right)\·\·\left({\mathrm{\λ}}_m^1{E}_m\right)].$

1, feature grouping

The first step is from primitive character space, to propose several separate subspaces.Accomplish this point, need to, on the associated diagram of Xi Er baud-Shi Mite independence criterion (HSIC) value, calculate every pair of feature, thereby obtain object by frequency spectrum cluster.

HSIC can obtain whole non-linear dependence of two stochastic variable x and y.As the people such as Gretton are described, so long as about the problem of reproducing kernel, Xi Er baud space is exactly very general.The method that this method is measured independence than other, for example related coefficient, is more suitable for this object.For illustrating that this method can measure independence really, when the proof such as Gretton and if only if x and y independence, HSIC value is 0.For object, it estimates by experimental formula from limited (xiyi) tuple by following steps:

${\mathrm{\ρ}}_n(x,y)=\frac{1}{{(1-n)}^2}\mathrm{tr}\left({\mathrm{Hk}}_{x}h{k}_y\right)---\left(2\right)$

Wherein, H _ij=σ _ij-n ^-1, Kx and Ky are respectively the vector products of vector x and y, n is the number of sampling.Yet calculating K x and Ky will spend O (n ²) time and space, this is very expensive for larger data group.So the full Cholesky factorization that toos many or too much for use finds Lx and Ly.Like this, Kx and Kx '=LxLx for Ky ^tand Ky '=LyLy ^testimate.Afterwards, approximate HSIC is just by calculating as follows:

${\mathrm{\ρ}}_n(x,y)=\mathrm{tr}\left(\left(L_x^T{L}_y\right)\right)\left(\left(L_y^T{L}_x\right)\right)---\left(3\right)$

This calculates at O (nf ²) time in complete, wherein f is to be the selected columns of L.In very large data group, HSIC estimation is taked to raise the efficiency from the mode of original crowd evacuation sampling.The loss of a part of accuracy is accepted in such sampling because the HSIC of estimation and real HSIC be with speed approach.

Carry out Feature Combination, will calculate ρ to the every couple of feature i of the luv space of data group X and j _nobtain like this associated diagram Wij=ρ _n(xi, xj).According to the people's such as NG research, W is carried out to frequency spectrum cluster and find m disjoint proper subspace afterwards as below shown in experiment, m can obtain reasonable result while getting on a large scale numerical value, so this selection problem is not strict.But, m >=20 o'clock result is best conventionally.

2, multiple collection of illustrative plates embeds

After obtaining m disjoint subspace, according to every sub spaces, find respectively the embedding of the data group of m.For every sub spaces sm, with associated diagram W of following Formula:

Wij＝sum(min(Pm(xi),Pm(xj)))(4)

Wherein Pm (x) is mapped to x the formula of m sub spaces.Compared with using kNN neighborhood or ∈-Neighborhood Graph, what used here is typical Laplce's characteristic pattern.W builds with complete connection layout.Use complete connection layout to select according to experience, all good because the result that in experiment before, connection layout draws is completely got any k than kNN Neighborhood Graph.Afterwards, according to each step in the 2-4 such as Ng, on W, compose embedding.These steps are:

1. find L=D ^-1/2wD ^-1/2, wherein D is that diagonal matrix and Dii equal the result to the capable summation of the i of W.

2. find k proper vector e1...ek the highest in L, and to building matrix E, be [e1...ek] by row

3. each row of E is normalized to and is 1

Notice, also some is different for these processes and Laplce's characteristic pattern, because step 3 does not adopt the people's such as Ng way.Unit standardization is very important for reducing that m minute dimensional variation between other embedding.The optimal selection of k seems to embed and change for difference spectrum, yet for simplicity, just for all embedding unifications, has selected a k.In following work, exploring be each embedding find one independently k improve performance.

Final step is to produce FGSM to merge this m embedding.This is all linked up and obtains by the simple simple E1...Em whole, shows: X=[E1..Em with list].Then the i of X is capable is exactly m * k length, is used for describing the vector of sampling xi.Although this scheme effect is fine, yet further by convergent-divergent suitable before merging, each embeds to improve its performance.Between every a line, calculate Euclidean distance, then with it, find λ ₁:

dists _i，jk＝||e _i，j-e _i，k|| ²

λ _i＝a(dists _i)

Wherein σ (x) is the standard deviation in x.Then, for obtaining final expression, use each λ ₁... λ _me1...Em is carried out to convergent-divergent, then connects in column: as a result, each embeds and to zoom to that to make total range rate be 1.

In the present invention, do not consider the expansion outside sample, although perhaps these be important for the task of looking like to identify identification or retrieval time.Yet each embedding is applied respectively to Nystrom estimation and obtain sample external expansion.

The method---for human action, improve the accuracy of cluster, retrieval and identification mission especially.Based on forefathers, in the work aspect spectrum embedding, on each subspace separating in primitive character space, produce several spectrums and embed, and suppose that this way can farthest keep the information in primitive character space.By the Comprehensive Experiment in four data groups, the new algorithm showing---FGSM surmounts current state-of-the-art algorithm in the performance of the tasks such as cluster, retrieval and relevant feedback in whole data groups.And FGSM also can surmount the identification accuracy of current most advanced algorithm in specific data group.

Those skilled in the art of the present technique understand, and the present invention relates to for carrying out the equipment of the one or more operation of operation described in the application.Described equipment is required object and specialized designs and manufacture, or also comprises the known device in multi-purpose computer, and described multi-purpose computer has storage procedure Selection within it and activates or reconstruct.Such computer program (is for example stored in equipment, computing machine), in computer-readable recording medium or be stored in the medium of any type that is suitable for store electrons instruction and is coupled to respectively bus, described computer-readable medium includes but not limited to the dish (comprising floppy disk, hard disk, CD, CD-ROM and magneto-optic disk) of any type, storer (RAM), ROM (read-only memory) (ROM), electrically programmable ROM, electric erasable ROM (EPROM), electrically erasable ROM (EEPROM), flash memory, magnetic card or light card immediately.Computer-readable recording medium comprises for any mechanism with for example, by the storage of the readable form of equipment (, computing machine) or transmission information.For example, computer-readable recording medium comprises storer (RAM) immediately, ROM (read-only memory) (ROM), magnetic disk storage medium, optical storage medium, flash memory device, the signal (such as carrier wave, infrared signal, digital signal) propagated with electricity, light, sound or other form etc.

Those skilled in the art of the present technique understand, and with computer program instructions, realize each frame in these structural drawing and/or block diagram and/or flow graph and the combination of the frame in these structural drawing and/or block diagram and/or flow graph.The processor that these computer program instructions is offered to multi-purpose computer, special purpose computer or other programmable data disposal route generates machine, thereby the instruction of carrying out by the processor of computing machine or other programmable data disposal route has created for the frame of implementation structure figure and/or block diagram and/or flow graph or the method for a plurality of frame appointments.

Those skilled in the art of the present technique understand, and the step in the various operations of having discussed in the present invention, method, flow process, measure, scheme are replaced, change, combine or delete.Further, have other step in the various operations discussed in the present invention, method, flow process, measure, scheme also by alternately, change, reset, decompose, combination or delete.Further, of the prior art have with the present invention in step in disclosed various operations, method, flow process, measure, scheme also by alternately, change, reset, decompose, combination or delete.

The above is only part embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (7)

1. the human action learning method based on many graph expressions, is characterized in that, comprising:

Feature space is divided into several separate subspaces, and represents described subspace by a plurality of figure;

From every sub spaces, produce different incidence matrix, and compose embedding in every sub spaces;

These are embedded to convergent-divergent and link together, for each information point obtains a single representation, with generating feature set spectrum multigraph FGSM, described FGSM can obtain from original feature space minimum loss of data;

FGSM is applied to cluster, information retrieval and recognizer, to carry out human action study.

2. the human action learning method based on many graph expressions as claimed in claim 1, is characterized in that, described feature space is higher dimensional space.

3. the human action learning method based on many graph expressions as claimed in claim 1, is characterized in that, described characteristic set can be divided into several disjoint subsets, independent of one another between all subsets, and relies on each other between subset inside.

4. the human action learning method based on many graph expressions as claimed in claim 1, is characterized in that, described feature space is divided into several separate subspaces, comprising:

On the associated diagram of Xi Er baud-Shi Mite independence criterion HSIC value, calculate every pair of feature, to obtain in described HSIC by frequency spectrum cluster, whole non-linear dependence of two stochastic variable x and y.

5. the human action learning method based on many graph expressions as claimed in claim 4, is characterized in that, describedly on the associated diagram of Xi Er baud-Shi Mite independence criterion HSIC value, calculates every pair of feature, comprising:

By following steps, from limited (xi, yi) tuple, by experimental formula, estimate:

${\mathrm{\ρ}}_n(x,y)=\frac{1}{{(1-n)}^2}\mathrm{tr}\left({\mathrm{Hk}}_x{\mathrm{Hk}}_y\right)---\left(2\right)$

Wherein, Hij=δ _ij-n ^-1, Hi refers to HKx, and Hj refers to HKy, and Kx and Ky are respectively the vector products of vector x and y, and n is the number of sampling.

6. the human action learning method based on many graph expressions as claimed in claim 5, is characterized in that, describedly on the associated diagram of Xi Er baud-Shi Mite independence criterion HSIC value, calculates every pair of feature, also comprises:

Because calculating K x and Ky will spend O (n ²) time and space, by incomplete Cholesky factorization, find Lx and Ly, like this, Kx and Ky Kx '=LxLx ^tand Ky '=LyLy ^testimate, afterwards, approximate HSIC is just by calculating as follows:

ρ _n(x,y)＝tr((Lx ^TLy))((Ly ^TLx)) (3)

This calculates at O (nf ¹in the time of defending, complete, wherein f is to be the selected columns of L.

7. the human action learning method based on many graph expressions as claimed in claim 1, is characterized in that, also comprises:

After obtaining m disjoint subspace, according to every sub spaces, find respectively the embedding of the data group of m, for every sub spaces sm, with associated diagram W of following Formula:

Wij＝sum(min(Pm(xi),Pm(xj))) (4)

Wherein Pm (x) is mapped to x the formula of m sub spaces.