CN102663374A - Multi-class Bagging gait recognition method based on multi-characteristic attribute - Google Patents

Abstract

The invention relates to a multi-class Bagging gait recognition method based on a multi-gait characteristic attribute, which belongs to the technical field of pattern recognition. According to the method, a nearest neighbor classifier is used as a weak classifier, and an integration classifier is constructed by expanding a two-class attribute Bagging method to a plurality of classes on the basis of 20 gait attribute characteristic sets after wavelet packets are decomposed and principal components are completely analyzed so as to carry out gait identity identification. The method comprises the following steps of: preprocessing, extracting characteristics and finally classifying test samples by using a method combining a nearest neighbor classifying principle and an MCAB algorithm. According to the multi-class Bagging gait recognition method based on the multi-gait characteristic attribute, a method fusing wavelet packet decomposition (WPD) and (2D) 2 principal component analysis (PCA) is adopted for the first time to extract and also select gait characteristics. The problem of loss of high-frequency components in a traditional gait recognition method based on wavelet transformation or overlarge dimensionality caused by simply adopting all data is solved. The multi-class Bagging gait recognition method based on the multi-gait characteristic attribute has higher recognition rate and visual angle change robustness.

Description

Multi-class Bagging gait recognition method based on many characteristic attributes

Technical field

The invention belongs to mode identification technology; Be specifically related to a kind of multi-class Bagging Gait Recognition new method based on multistep attitude characteristic attribute; Being a kind of automatic analysis of the gait that computer technology, Digital image processing technique, pattern-recognition etc. realize people and method of differentiation utilized, is about the algorithm of Method of Gait Feature Extraction and identification in the living things feature recognition field.

Background technology

Biometrics identification technology is meant the technology of utilizing human physiological characteristic that had, that can identify its identity own or behavioural characteristic to carry out authentication.Compare with traditional identity validation technology, biometrics identification technology has fundamentally been stopped to forge and steal, and has higher reliability, security, more and more widely the authentication that is applied to some security systems.

The Gait Recognition technology is as a kind of novel biometrics identification technology, and it is the biological identification technology that carries out identification according to the posture that people in the video sequence walks.Compare with other biometrics identification technology; The Gait Recognition technology is with its non-infringement property, remote identity and advantage such as be difficult to hide and received the people's favor, has a wide range of applications in fields such as national public safety, financial security, authentication, video monitorings.

Feature Extraction Technology about gait; There is document to adopt WAVELET PACKET DECOMPOSITION to solve this problem preferably; But the characteristics of image dimension after the WAVELET PACKET DECOMPOSITION is higher; And it adopts classical PCA algorithm to carrying out feature extraction, when promptly adopting the method for svd to ask eigenwert and the proper vector of correlation matrix, calculates wasteful.Two dimension principal component analysis (PCA) (2DPCA) can directly be calculated image data matrix, and calculated amount is few relatively a lot, but needs n*k behind the 2DPCA (wherein; N is an image resolution ratio; K is a selected characteristic column vector number after the conversion, and k＜n) individual data are come presentation video, and the dimension of proper vector is still higher.Principal component analysis (PCA) ((2D) fully ²PCA) can further reduce the dimension of proper vector, expend thereby reduce identification, and suitable on its recognition performance with 2DPCA, even be superior to 2DPCA.

The someone proposes a kind of attribute Bagging algorithm AB about two types of problems at present; The present invention is on the basis of the attribute Bagging of two types of problems algorithm AB, proposes a kind of multi-class Bagging Gait Recognition new method based on multistep attitude characteristic attribute---MCAB (Multi-class Attribute Bagging) algorithm.

Summary of the invention

Content of the present invention is to have proposed multi-class Bagging (MCAB) Gait Recognition new method based on multistep attitude characteristic attribute.This method with 1NN as Weak Classifier, through two generic attribute Bagging methods being expanded to the multi-class integrated classifier MCAB (Multi-class Attribute Bagging) that makes up.We evaluate and test this method on NLPR gait data storehouse, and the result shows, with simple employing wavelet packet with (2D) ²Recognition methodss such as PCA are compared, and this method has higher discrimination and visual angle change robustness.

Technology contents of the present invention is following:

For utilizing the MCAB algorithm to experimentize, we need to extract the gait energygram, to overcome gait data amount problems of too at first through pretreated normalization gait image sequence is carried out cycle detection; Again the gait energygram is carried out WAVELET PACKET DECOMPOSITION and principal component analysis (PCA) fully, the result images that obtains has been represented the not characteristic of ipsilateral of gait image respectively; Last classification performance according to aforementioned different characteristic is also classified with the MCAB algorithm through the different attribute that each characteristic is regarded as gait.

Multi-class Bagging (MCAB) Gait Recognition new method based on multistep attitude characteristic attribute; The step of this method comprises: the pre-service of body gait sequence, feature extraction; Utilize the MCAB algorithm to be grouped into specimen in corresponding type at last; And recognition effect estimated, its concrete steps are following:

Step 1, pre-service

(1) morphology is handled

The human motion target image of background separation is carried out morphology handle,, obtain more excellent segmentation effect to remove the cavity that binary image exists;

(2) target is extracted

Utilize the method for 8 connected component analyses to extract a simply connected moving target, promptly people's silhouette is removed residual noise, thereby obtains more excellent two-value profile diagram;

(3) image normalization

Cut out the gait image of standard according to the human body contour outline coordinate, obtain the size normalized image, wherein, the size of image is unified to be the 64*64 pixel.

Step 2, feature extraction

(1) detection of gait cycle

Utilize the profile width of human body that the characteristic that sexually revises synchronizing cycle takes place in time, the change width signal through human body contour outline comes dividing gait cycles;

(2) set up gait energygram (GEI), with of the representative of gait energygram as different gait sequences;

Gait energygram: after carrying out the gait cycle detection, be through the gait sequence Flame Image Process in the one-period is generated GEI:

$G(x,y)=\frac{1}{N_1}\underset{t=1}{\overset{N_1}{\mathrm{\Σ}}}{B}_t(x,y)$

In the formula, (x y) is the gait energygram of this periodic sequence, N to G ₁Be the length of complete gait cycle sequence, B _t(x y) is t gait image in the one-period, and x, y represent the two dimensional image plane coordinate.

(3) merge WPD+ (2D) ²The PCA feature selecting

At first, adopt secondary WAVELET PACKET DECOMPOSITION (WPD) gait energygram, after decomposing, every width of cloth gait energygram obtains 20 images after the decomposition; Then, adopt complete principal component analysis (PCA) ((2D) respectively ²PCA) (the two-dimentional principal component analysis (PCA) of column direction at once) further extract validity feature; Discern respectively with nearest neighbor classifier at last, respectively each characteristic is carried out the calculating of discrimination.

Step 3, integrated Classification and Identification

(1) primitive attribute of confirming double sampling is gathered

In the MCAB algorithm, need carry out that at first the double sampling of putting back to is arranged for n time to the community set of original training instance, n is process WPD+ (2D) ²Discrimination constitutes new training instance by this n attribute, again by these new training instance constructs Weak Classifiers more than or equal to 50% attribute number behind the PCA.Because the primitive attribute collection is had the double sampling of putting back to, the attribute that therefore in new training example is described, has possibly occur repeatedly, and the attribute that has possibly once not occur yet.We select those discriminations to constitute community set AttributeSet more than or equal to 50% attribute.

(2) utilize the MCAB algorithm to carry out Classification and Identification

Have document to propose a kind of attribute Bagging algorithm AB about two types of problems, be provided with two classifications, the classification space is Y={1 ,-1}, and instance space is ￠, the training example set is S={ (x ₁, y ₁), (x ₂, y ₂) ..., (x _m, y _m), x wherein _i∈ ￠, y _i∈ Y, instance x _iWith n attribute representation, property set A is { a ₁, a ₂..., a _n.Algorithm AB carries out T wheel altogether, and concentrating from primitive attribute during each is taken turns has a double sampling n attribute of putting back to, and constitutes new training example by the training example of this n new attribute description and collects, and will newly train instance to go out basic sorter h with basic classification algorithm training again _t(x) → { 1,1} distributes to h _t(x) weights are a _tAt last the instance of the unknown result with T basic sorter is voted, the simple vote function does

H(x)＝sign(f(x)) (1)

Wherein, $f\left(x\right)=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{a}_t{h}_t\left(x\right),$ a _t=1.

But the Gait Recognition problem is a multi-class classification problem, and for this reason, we introduce the M1 algorithm above-mentioned algorithm is expanded to the multi-class attribute Bagging of multistep attitude algorithm MCAB.If multi-class space is Y={1,2 ..., m}, y _i∈ Y.For multi-class M1 algorithm, establish every the wheel and train the training set that uses to be S _t, t=1 wherein ..., T, wherein T is an iterations, at S _tLast with weak learning algorithm training generation Weak Classifier---anticipation function h _t(x) → and Y, distribute to h _t(x) weights are a _t, last strong anticipation function does

$H\left(x\right)=\arg \underset{y\∈Y}{\max }\underset{t:h_t\left(x\right)=y}{\mathrm{\Σ}}{a}_t\left|\right|h_t\left(x\right)=y\left|\right|---\left(2\right)$

Wherein, when the category label of assigning to when basic sorter is identical with y, || h _t(x)=y|| is 1, otherwise be 0;

r _tBe h _t(x) the correct number of classification deducts the number of branch classification error, and t is an iterations, t=1 ..., T.

Following formula is appreciated that establishing y has weight w for test sample book x can assign to a classification y ∈ Y on each basic sorter, if basic sorter h _tAssigned on the y, then corresponding weight w is this base sorter h _tWeights a _t, the class that maximum type of final weights sum assigned to for test sample book.

MCAB carries out the T wheel, generally gets 10=＜T＜=100, and each is taken turns has a resampling n attribute of putting back to from aforesaid primitive attribute set A ttributeSet, constitutes new training example collection by the individual training example of attribute description newly of this n; On new training example collection, go out different basic sorter h with the 1NN algorithm training _t(x); Constitute integrated classifier by above-mentioned a plurality of basic sorters by formula (2) at last.

Beneficial effect of the present invention is: 1. through gait cycle is calculated the gait energygram; This width of cloth gait energygram has comprised gait information such as the profile of each gait image, frequency, phase place in the gait cycle; Reduce pending gait image data volume when can guarantee not abandon gait feature, to reduce calculation consumption; 2. utilize based on WPD+ (2D) ²The PCA method is extracted characteristic; Solve and existingly to lose or simply adopt dimension problems of too due to the total data based on the gait recognition method medium-high frequency component of wavelet transformation, the method more exactly abstraction reaction movement human walking characteristic effective information with reduce the intrinsic dimensionality that is used to discern gait; 3. utilize the MCAB algorithm to carry out Classification and Identification, and visual angle change is had good robustness with raising Gait Recognition accuracy.

Description of drawings

Fig. 1 is the process flow diagram of algorithm of the present invention.

Fig. 2 is a pretreated process flow diagram in the algorithm of the present invention

Fig. 3 is the process flow diagram that algorithm characteristics of the present invention is extracted

Fig. 4 is secondary WPD+ (2D) in the algorithm of the present invention ²PCA feature extraction process flow diagram

Fig. 5 is a wavelet packet exploded view in the algorithm of the present invention

Embodiment

Provide the explanation of each related in this invention technical scheme detailed problem below in detail:

Step 1, preprocessing process are following:

The database that we adopt is a gait data storehouse of Institute of Automation Research of CAS, and this database is with background separation, and the work that the present invention will do is on this basis, to carry out pre-service, thereby carries out operations such as cycle detection and calculating gait energygram.

(1) morphology is handled

Because can there be noise unavoidably in the influence of other extraneous factors such as weather, illumination, shadow in the image after the background separation, therefore also need do further processing, to obtain best segmentation effect to image.The present invention uses morphologic filtering to eliminate the noise in the bianry image and fills up the disappearance of moving target.As a kind of image commonly used filter method for de-noising, the fundamental operation that morphology is used for image filtering is to expand and corrosion, by expanding and the mutually combining and derive other two kinds of computings of corrosion: opening operation and closed operation.But the cam contour of opening operation smooth object breaks off narrow connection, removes tiny jut; But the concave contour of closed operation smooth object connects into tiny curved mouthful with long and narrow breach.Utilize this attributes can realize the purpose of filtering and filling cavity.

(2) target is extracted

After morphology is handled, still possibly exist the part clutter noise to form piece not of uniform size, and real moving target is maximum in these pieces often.Therefore image is further carried out the connected domain analysis, promptly utilize the method for 8 connected component analyses to extract a simply connected moving target, purpose is only to keep the moving target in the image, thereby obtains more excellent two-value profile diagram.

(3) image normalization

For of the influence of removal of images size to identification, should at first make human body placed in the middle, the size unification with image is the 64*64 pixel then.

Step 2, characteristic extraction procedure are following:

(1) detection of gait cycle

People's walking is the behavior of one-period property, and the definition gait cycle is: to the time that the corresponding leg heel lands and experienced once more, comprise two stance phases and two shaking peroids from heel contact.In order to raise the efficiency, the present invention utilizes the profile width of human body that the characteristic that sexually revises synchronizing cycle takes place in time, and the change width signal through human body contour outline comes dividing gait cycles.

(2) gait energygram

After the cycle detection of carrying out gait; Directly adopting gait cycle to carry out Gait Recognition exists data volume huge; The difficulty and the calculation consumption of Method of Gait Feature Extraction have been aggravated; For guaranteeing that not abandoning gait feature reduces pending gait image data volume simultaneously; The present invention has adopted the method for gait energygram, and the some gait images that are about in one-period synthesize piece image through weighted-average method, and this width of cloth image has comprised gait information such as the profile of each gait image, frequency, phase place in the gait cycle.This method need not considered the gait size at interval of every frame, and avoids the influence of some accidentalia.For given two-value gait cycle image sequence B _t(x, y), the definition of GEI is following:

$G(x,y)=\frac{1}{N_1}\underset{t=1}{\overset{N_1}{\mathrm{\Σ}}}{B}_t(x,y)$

Wherein, (x y) is the gait energygram of this periodic sequence, N to G ₁Be the length of complete gait cycle sequence, B _t(x y) is t gait image in the one-period, and x, y represent the two dimensional image plane coordinate.

(3) merge WPD+ (2D) ²The PCA feature selecting

Two-dimensional discrete wavelet conversion is divided into low frequency component A and high fdrequency component with every two field picture, and comprising horizontal component H, vertical component V and diagonal components D, the low frequency component A that this method has only kept picture signal has lost high fdrequency component.WAVELET PACKET DECOMPOSITION (WPD) has not only kept the low frequency component of picture signal, has also kept high fdrequency component.

Characteristics of image dimension after the WAVELET PACKET DECOMPOSITION is higher, adopts classical PCA algorithm when carrying out feature extraction, and when promptly adopting the method for svd to ask eigenwert and the proper vector of correlation matrix, calculated amount is very big.And two-dimentional principal component analysis (PCA) (2DPCA) is directly calculated matrix, and calculated amount is few relatively a lot.But 2DPCA also has its drawback aspect image characteristics extraction.Suppose that the image size is 64*64, come presentation video through needing the individual data of 64*k (selected characteristic column vector number after wherein k is conversion, and k＜64) behind the 2DPCA; In order to satisfy accuracy requirement; Usually k is greater than 5, needs 320 data to come presentation video so at least, and the dimension of proper vector is still higher.Principal component analysis (PCA) ((2D) fully ²PCA) can further reduce the dimension of proper vector, expend thereby reduce identification, and suitable on its recognition performance with 2DPCA, even be superior to 2DPCA.For the information of more effectively utilizing image is carried out Gait Recognition, based on above-mentioned consideration, this paper has proposed a kind of fusion WAVELET PACKET DECOMPOSITION (WPD) and (2D) ²The method of PCA.

Concrete implementation procedure is following:

At first (x y) carries out secondary WPD and decomposes, and obtains 20 images after the decomposition to every width of cloth gait energygram G.That is: utilize wavelet packet character that original gait energygram is carried out first order WPD and decompose, get the image of low frequency A1, horizontal high frequency H1, vertical high frequency V1, diagonal angle high frequency D1, respectively these four images are carried out second level WPD again and decompose, obtain low frequency A2 successively; Horizontal high frequency H2, vertical high frequency V2, diagonal angle high frequency D2, low frequency HA2; Horizontal high frequency HH2, vertical high frequency HV2, diagonal angle high frequency HD2, low frequency VA2; Horizontal high frequency VH2, vertical high frequency VV2, diagonal angle high frequency VD2, low frequency DA2; Horizontal high frequency DH2, vertical high frequency DV2, diagonal angle high frequency DD2.

Secondly, the image after respectively WPD being decomposed (A1 ..., DD2) carry out (2D) ²The PCA conversion, (2D) ²The detailed process of PCA conversion is following:

I) calculated population sample average matrix

$\stackrel{\‾}{A}=\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{A}_k$

Wherein, A _kBe k sample image, M is a total sample number

Ii) calculate sample covariance matrix

$G_1=\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\Σ}}}(A_k-\stackrel{\‾}{A}){(A_k-\stackrel{\‾}{A})}^T$

$G_2=\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{(A_k-\stackrel{\‾}{A})}^T(A_k-\stackrel{\‾}{A})$

Calculate eigenwert and the standard orthogonal characteristic vector of covariance matrix G1, G2 respectively, the matrix that preceding d1, the proper vector of a d2 bigger nonzero eigenvalue are formed is U, V.Wherein, $U={[u_1,u_2,\·\·\·,u_{d_1}]}^T,V=[v_1,v_2,\·\·\·,v_{d_2}]$

Iii) generate (2D) ²Eigenmatrix Y ' after the PCA conversion

Y′＝UAV

Wherein, A is sample matrix arbitrarily.

At last, respectively 20 image sequences after the conversion are carried out the arest neighbors classification, utilize 1NN to calculate discrimination to each image sequence respectively.

Step 3, utilize the MCAB algorithm to classify

(1) primitive attribute of confirming double sampling is gathered

With 0 ° be example, in the selection in the step last discrimination discrimination constitute the primitive character set A ttribute Set={A1 that treats double sampling, H1, V1, A2, H2, V2, HA2, HH2} more than or equal to 50% characteristic.Be provided with 20 people's gait image, promptly the classification number is 20.

(2) utilize the MCAB algorithm to carry out Classification and Identification

Step1: confirm that the number in the double sampling community set is 8, confirm iterations 10;

Step2:For t=1: 10, iteration 10 times;

Iteration for the first time: from Attribute Set, have with putting back to and sample 8 times, extract one at every turn, obtain property set A1={A1, H1, HH2; V2, H2, V1, H1; A1} owing to all be that the extraction of putting back to is arranged randomly at every turn, so some attribute may occur repeatedly, for example: A1; H1, some attribute maybe not can occur, for example: A2, HA2.Each training sample instance X _kAttribute only get A ₁Obtain vectorial S _1k=[A1 ^(k)H1 ^(k)HH2 ^(k)V2 ^(k)H2 ^(k)V1 ^(k)H1 ^(k)A1 ^(k)], X _kUse S _1kRepresent, go out basic sorter h with the 1NN algorithm training ₁(x) → and Y, calculate weights:

$a_1=\frac{1}{2}\ln \left|\frac{1+r_1}{1-r_1}\right|$

Wherein, r ₁=h ₁(x) number of correct classification deducts h ₁(x) number of mis-classification.

...

The tenth iteration: from Attribute Set, put back to the ground double sampling and obtain property set A 8 times ₁₀={ V2, HH1, A2, H2, HH2, HA2, H1, A1}, each instance sample X _kAttribute only get A ₁₀Obtain vectorial S _10k=[V2 ^(k)HH1 ^(k)A2 ^(k)H2 ^(k)HH2 ^(k)HA2 ^(k)H1 ^(k)A1 ^(k)]; X _iUse S _10kRepresent, go out basic sorter h with the 1NN algorithm training ₁₀(x) → and Y, calculate weights

$a_{10}=\frac{1}{2}\ln \left|\frac{1+r_{10}}{1-r_{10}}\right|$

Wherein, r ₁₀=h ₁₀(x) number of correct classification deducts h ₁₀(x) number of mis-classification.

Step3: for any test sample book x,

$H\left(x\right)=\arg \underset{y\∈Y}{\max }\underset{t:h\left(x\right)=y}{\mathrm{\Σ}}{a}_t\left|\right|h_t\left(x\right)=y\left|\right|$

Wherein, Y is the classification set, y ∈ Y; T is an iterations, and value is 1...T.

Formula further specifies as follows among the Step3: for any test sample book x, the classification situation of each basic sorter is following:

h ₁(x)-＞1，h ₂(x)-＞2，h ₃(x)-＞2，h ₄(x)-＞3，h ₅(x)-＞1，h ₆(x)-＞2，h ₇(x)-＞5，h ₈(x)-＞19，h ₉(x)-＞1，h ₁₀(x)-＞3。

The weights sum of classification 1 is: a ₁+ a ₅+ a ₉

The weights sum of classification 2 is: a ₂+ a ₃+ a ₆

The weights sum of classification 3 is: a ₄+ a ₁₀

The weights sum of classification 5 is: a ₇

The weights sum of classification 19 is: a ₈

All the other category label weights are: 0.

The category label of final test sample is: the maximum corresponding class label of weights sum,

That is: max{a ₁+ a ₅+ a ₉, a ₂+ a ₃+ a ₆, a ₄+ a ₁₀, a ₇, a ₈Pairing category label.As the weights sum of two or more maximums, then give any category label in these several kinds to test sample book.

Following detailed description experimental result of the present invention:

The database that experiment of the present invention is adopted is the CASIA Dataset A database that Institute of Automation, CAS provides; This database comprises 20 people altogether; Everyone has 3 different visual angles (0 °, 45 ° and 90 °) respectively, and 4 sequences are taken at each visual angle respectively, amounts to comprise 240 sequences.These color image sequences are taken with the speed of 25 frame per seconds, and original size is the 352*240 pixel, and average length is about 100 frames.

In order to verify the MCAB algorithm, we select three sequences as training data in 4 sequences of everyone same view angle arbitrarily in the NLPR storehouse, and a remaining sequence is done test.

Table 1 is 20 people under three kinds of different visual angles (0 °, 45 °, 90 °) condition, adopts the basis function of Haar small echo as WAVELET PACKET DECOMPOSITION, and each subimage of WAVELET PACKET DECOMPOSITION is carried out (2D) respectively ²The aforesaid result who discerns based on the nearest neighbor classifier of L2 normal form is adopted in the PCA conversion at last, and we have added up the asynchronous discrimination of iterations at last.

Table 1WPD+ (2D) ²The discrimination of PCA

Table 2 is that the recognition result with table 1 is the basis, calls the result that the MCAB algorithm is discerned then, and we have added up the asynchronous discrimination of iterations at last.

The discrimination of table 2MCAB algorithm

Can find out from table 2: each the subimage coefficient than simple use wavelet decomposition is discerned, and when iterations T was big, the recognition performance of MCAB algorithm was higher, and can both obtain high recognition to different visual angles (0 °, 45 °, 90 °).

Claims (2)

1. multi-class Bagging gait recognition method based on multistep attitude characteristic attribute is characterized in that step is following:

Step 1, pre-service

To the human motion target image of background separation carries out the morphology processing successively, target is extracted and image normalization is handled;

Step 2, feature extraction

Through the detection of gait cycle, set up the gait energygram, with of the representative of gait energygram, utilize WPD+ (2D) as different gait sequences ²The PCA method is calculated 20 Feature Recognition rates;

Step 3, Classification and Identification

Select discrimination in the step 2 more than or equal to 50% characteristic as the primitive character set A ttribute Set that treats double sampling, utilize the MCAB algorithm that test sample book is finally discerned; Specific as follows

(1) in the MCAB algorithm; Need carry out that at first the double sampling of putting back to is arranged for n time to the community set of original training instance; N for through discrimination after the step 2 more than or equal to 50% attribute number, this n attribute may occur repeatedly, what have possibly once also can not occur; Constitute new training instance by this n attribute, again by these new training instance constructs Weak Classifiers;

(2) utilize the MCAB algorithm to carry out Classification and Identification

I: confirm the number n of double sampling attribute, confirm iterations T; Wherein n is a discrimination more than or equal to the number of 50% attribute, and the T value is 10～100;

3 steps below ii:For t=1: T carries out;

A): a double sampling n attribute obtains property set A from Attribute Set _t,, only get A to each sample among the training set S _tIn attribute, obtain new attribute S _t

B): at attribute S _tUpward go out basic sorter h with the 1NN algorithm training _t(x) → Y;

C): calculate weights

$a_1=\frac{1}{2}\ln \left|\frac{1+r_1}{1-r_1}\right|$

Wherein, r _t=h _t(x) number of correct classification deducts h _t(x) number of mis-classification;

Iii: for any test sample book x, integrated classifier H (x) is:

$H\left(x\right)=\arg \underset{y\∈Y}{\max }\underset{t:h\left(x\right)=y}{\mathrm{\Σ}}{a}_t\left|\right|h_t\left(x\right)=y\left|\right|$

Wherein, Y is the classification set, y ∈ Y; T is an iterations, and value is 1...T.

2. the multi-class Bagging gait recognition method based on multistep attitude characteristic attribute according to claim 1 is characterized in that the extraction step of described gait energygram is following:

After carrying out the gait cycle detection, be through the gait energygram that the gait sequence Flame Image Process in the one-period is generated:

$G(x,y)=\frac{1}{N_1}\underset{t=1}{\overset{N_1}{\mathrm{\Σ}}}{B}_t(x,y)$

In the formula, (x y) is the gait energygram of this periodic sequence, N to G ₁Be the length of complete gait cycle sequence, B _t(x y) is t gait image in the one-period, and x, y represent the two dimensional image plane coordinate.

Images