CN101034441A - Human motion date recognizing method based on integrated Hidden Markov model leaning method - Google Patents

Abstract

The invention opens a identification method of human motion data based on integrated hidden Markov model learning method. This method extracts two-dimensional geometric features for capturing data from human motion, and then effectively reducts the dimensionality of movement features data by the introduction of dimension reduction methods of non-linear flow pattern learning, and finally learns the movement of sports database by adoption of Hidden Markov integrated learning based on self-adaptive advance algorithm for achieving fast retrieval of conventional movement. Two-dimensional geometric features extracted by the method well expresses the essential attribute of movement, dimensionality reduction methods of the expansion of nonlinear manifold will successfully map features of high-dimensional movement to low-dimensional space that can reflect the inherent links between data, thus greatly eliminates data redundancy. While this invention can study through drop dimensional data used by methods of integrated Hidden Markov Model learning, makes movement automatically identificate and classificate on the basis of high-precision.

Description

Recognition methods based on the human body movement data of integrated hidden Markov model learning method

Technical field

The present invention relates to multimedia human body three-dimensional animation field, relate in particular to a kind of recognition methods of the human body movement data based on integrated hidden Markov model learning method.

Background technology

Since nineteen nineties, along with the rise of capturing movement technology, and the progress of equipment and technology, a large amount of 3 d human motions is caught data and is generated, and is applied in computer animation widely, recreation, medical science emulation, fields such as film trick.Appearance along with magnanimity 3 d human motion acquisition database, make and how from the human motion of complexity, to find the correctly essential characteristic of expressive movement information, how exercise data correctly being discerned efficiently and the magnanimity exercise data is efficiently handled, is a new challenge thereby effectively utilize motion capture data Kucheng.

Motion is the harmony combination of each articulation point signal, in retrieving, need rational motion feature to describe mechanism, therefore, extract which type of motion feature, and which type of expression way to come the expressive movement feature to have great influence with to the effect and the efficient of motion process.The two-dimensional geometry characterization method that is adopted among the present invention is different from traditional motion feature extracting method, considers that the geometry of motion can reflect the motion inherent attribute more truly.

The intrinsic dimensionality that extracts from exercise data is all very high usually, and it is intimate identical that the distance between each data will become because of central limiting law, can't distinguish each other, produces higher-dimension disaster (Curse ofDimensionality) problem.Therefore from high dimensional feature, seek the low dimensional feature of " meaningful ", can avoid causing higher time and space complexity, improve recall precision.And non-linearity manifold study (ISOMAP) algorithm uses the geodesic line distance that shortest path obtains being similar among the arest neighbors figure, replacement can not be represented Euclidean (Euclidean) distance of inherent manifold structure, be input to then in the multi-dimentional scale analysis (MDS) and handle, and then find to be embedded in the low dimension coordinate of higher dimensional space.

And by the learning method of hidden Markov model (HMM) motion in the human body movement data storehouse is learnt, can set up a continuous hidden Markov model for each two-dimensional geometry feature of the common type of sports of the overwhelming majority automatically.And because the precision of single weak learner is very low, so the present invention introduces the method for integrated study.Integrated study (ensemble learning) comes a problem is found the solution by a plurality of versions of the basic learner of training (base learner), can improve the generalization ability of learning system significantly.Recent years, the Many researchers in fields such as neural network, machine learning, statistics is all put in the research of integrated study in the world, makes this field become a suitable active research focus.By adopting adaptive propelling algorithm (Adaboost) integrated to weak hidden Markov learner, we have improved the precision of learner, thereby have improved the accuracy of Motion Recognition.

Summary of the invention

The object of the present invention is to provide a kind of recognition methods of the human body movement data based on integrated hidden Markov model learning method.

Recognition methods based on the human body movement data of integrated hidden Markov model learning method comprises the steps:

(1) from 3 d human motion data, extracts a kind of two-dimensional geometry feature that can reflect the local geometric relation;

(2) the dimensionality reduction algorithm of employing non-linearity manifold study projects to the higher-dimension raw data in the subspace of a low-dimensional, discloses the immanent structure of human motion by this lower dimensional space, realizes the data dimensionality reduction;

(3) adopt study major component feature kernel function to come dimensionality reduction is approached, realize the expansion of non-linearity manifold study algorithm, enable to handle the new exercise data outside the training set;

(4) adopt hidden Markov model to learn for the exercise data after the dimensionality reduction, obtain hidden Markov model parameter based on the common type of sports of two-dimensional geometry feature;

(5) adopt adaptive propelling algorithm, weak hidden Markov model has been set up a reinforced integrated study device, finish identification motion.

Described from 3 d human motion data, extract a kind of two-dimensional geometry feature that can reflect the local geometric relation: represent with Boolean function: F: ∏ → 0,1}; For the vector that f Boolean function formed, obtain mixed function F a: ∏ → 0,1} ^fF is regarded as a fundamental function, vector F (p) then is proper vector or briefly is the feature of posture p ∈ ∏, be applied to motion capture data D:[1:T by combination F ο D] → ∏ on, wherein, F ο D has expressed two peak values of motion cycle, can obtain the speed of moving very easily by peak value, feature F is for global position with towards, bone size, and the different local space distortion and the vertical moving of shank have unchangeability.

The dimensionality reduction algorithm of described employing non-linearity manifold study, the higher-dimension raw data is projected in the subspace of a low-dimensional, disclose the immanent structure of human motion by this lower dimensional space, realize the data dimensionality reduction: adopt non-linear manifold learning arithmetic to make up an adjacent map earlier, each point only links to each other with own adjacent point, 2 manifold distance is approached by their Euclidean distance in the neighborhood, 2 manifold distance in same neighborhood is not approached by their the shortest Dijkstra distance on adjacent map, non-subsequently linear manifold learning is based on this adjacency matrix, directly calculate the embedding of low-dimensional stream shape, make following objective function minimize:

$E={\left|\right|D-\tilde{D}\left|\right|}_{L_2}$

D wherein _IjCertain distance measure of point on the expression stream shape,

The statement embedded space point between distance, L ₂The Frobenius norm of representing matrix, ${\left|\right|A\left|\right|}_{L_2}=\sqrt{{\mathrm{\Σ}}_{i,j}{A}_{\mathrm{ij}}^2}.$

The step of described non-linearity manifold study algorithm is as follows:

1) set up adjacency matrix G., mainly contain the limit that two kinds of ways are determined adjacency matrix, epsilon neighborhood and K neighborhood. when a j at an i, or some i is when dropping on the ε of a j or K neighborhood, setting up a weight between i and j is d _x(i, limit j).

2) utilize Dijkstra or Flyod algorithm, calculation level between bee-line, obtain matrix D _G, all elements wherein all is the shortest path of every pair of point among the G

D _G＝{d _G(i，j)}

3) setting up d n-dimensional subspace n embeds: allow λ _pBe matrix τ (D _G) preceding p eigenwert (eigenwert by descending sort), v _p ⁱBe i component of p proper vector; The coordinate vector y of d n-dimensional subspace n _iP component just equal

Described employing study major component feature kernel function comes dimensionality reduction is approached, and realizes the expansion of non-linearity manifold study algorithm, enables to handle the new exercise data outside the training set: allow D={x ₁, x ₂..., x _nBe a sampled data set of unknown distribution, its continuous density is p, allows the P be that its corresponding experience distributes.Consider a Hilbert space

The inner product function be expressed as follows:

<f，g> _p＝∫f(x)g(x)p(x)dx

Wherein, p (x) is a weighting function, and kernel function K is in the space like this

A linear operator K neutralizes _pGet in touch as follows:

(K _pf)(x)＝∫K(x，y)f(y)p(y)ddy

Replace unknown distribution p with experience distribution P then and redefine " experience " Hilbert space

Kernel function

Produce a symmetric matrix based on D Wherein ${\tilde{M}}_{\mathrm{ij}}=\tilde{K}(x_i,x_j);$ Allow proper vector, eigenwert is to (v _k, λ _i) conduct ${\tilde{K}}_p{f}_k={\mathrm{\&lambda;}}_k^{\&prime;}{f}_k$ Separate; Allow $e_k\left(x\right)=y_k\left(x\right)\sqrt{{\mathrm{\&lambda;}}_k}$ Be used for representing the mapping of new data x to lower dimensional space, the expansion non-linearity manifold study makes it can apply on the new data:

$e_k\left(x\right)=\frac{1}{2\sqrt{{\mathrm{\&lambda;}}_k}}\underset{i}{\mathrm{\&Sigma;}}{v}_{\mathrm{ik}}\left\{E_{x^{\&prime;}}\right[{\tilde{D}}^2(x^{\&prime;},x_i)]-{\tilde{D}}^2(x_i,x)\}$

Exercise data with sufficient amount removes estimation space

Last K _pFundamental function, just obtain a non-training data mapping.

Describedly adopt hidden Markov model to learn for the exercise data after the dimensionality reduction, obtain hidden Markov model parameter:, be hidden Markov model λ of each type of sports study for j two-dimensional geometry feature based on the common type of sports of two-dimensional geometry feature _i(i=1,2 ..., M); A given observation sequence O is for each hidden Markov model utilizes forward-backward algorithm algorithm computation P (O| λ); Motion Recognition based on j two-dimensional geometry feature is just as follows by the expression formula that searching has the type of sports i of maximum P (O| λ) value to realize:

$\mathrm{action}\left(O\right)=\underset{i:i=1,\&CenterDot;\&CenterDot;\&CenterDot;,M}{\arg \max }(P\left(O\right|{\mathrm{\&lambda;}}_i)$

By above expression formula, M type of sports and N two-dimensional geometry feature have been formed the latent Markov models matrix of a M * N, and the hidden Markov model of j articulation point of i kind type of sports is expressed as HMM _{I, j}, corresponding parameter is λ _{I, j}The pairing Hidden Markov Model (HMM) of row j is gathered the sorter of corresponding j two-dimensional geometry feature;

Need to calculate given state s at t moment observed events O _tProbability P (O _t| s _t=s); Hidden Markov model calculates P (O with probability density function continuously _t| s _t=s), expression formula is as follows:

$\underset{m=1}{\overset{3}{\mathrm{\&Sigma;}}}\left(w_{\mathrm{sm}}\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{\frac{d}{2}}{\left|{\mathrm{\&Sigma;}}_{\mathrm{sm}}\right|}^{\frac{1}{2}}}{e}^{-\frac{1}{2}(O_t-{\mathrm{\&mu;}}_{\mathrm{sm}}){\mathrm{\&Sigma;}}_{\mathrm{sm}}^{-1}{(O_t-{\mathrm{\&mu;}}_{\mathrm{sm}})}^T}\right).$

The adaptive propulsion method of described employing is strengthened weak hidden Markov model, obtains a reinforced integrated classifier that can reflect the human body integral movable information, realizes correctly Motion Recognition efficiently: given n training sample (x ₁, y ₁) ..., (x _n, y _n) training set, y wherein _i=| 0,1|, (i=1,2 ..., n) mistake of corresponding sample identification and correct has u error sample in the sample, v correct sample; The space characteristics for the treatment of each articulation point of branch type games is expressed as f _j(), wherein 1≤j≤16; For i training sample x _i, it be characterized as f _j(x _i); The Weak Classifier h of the space characteristics of j articulation point _j(x) by a feature f _j, a threshold values θ _jBias p with an indication inequality direction _jConstitute:

$h_j\left(x\right)=\left\{\begin{array}{c}1,p_j{f}_j<p_j{\mathrm{\&theta;}}_{\mathrm{<figure-callout\; id="0"\; label="j"\; filenames="A20071006768400231.png"\; state="\{\{state\}\}">j</figure-callout>}}\\ 0,\mathrm{otherwise}\end{array}\right.$

By aligning counter-example analysis, select T minimum Weak Classifier of error rate, optimal combination becomes a strong classifier.

The useful effect that the present invention has is:

(1) extracted the two-dimensional geometry feature that the fine reflection human body integral of energy moves, and the dimensionality reduction algorithm of employing non-linearity manifold study, in the subspace of high dimensional feature data projection to a low-dimensional, disclose the immanent structure of human motion by this lower dimensional space, realize the data dimensionality reduction, eliminate data redundancy, reduce follow-up calculated amount;

(2) adopt study major component feature kernel function to come dimensionality reduction is approached, realize the expansion of non-linearity manifold study algorithm, enable to handle the new exercise data outside the training set, improved the general applicability of non-linearity manifold study method;

(3) adopt hidden Markov model to learn to the exercise data after the dimensionality reduction, obtain hidden Markov model based on the common type of sports of two-dimensional geometry feature, adopt adaptive propulsion method that weak hidden Markov model is strengthened, obtain a reinforced integrated classifier that can reflect the human body integral movable information, realization is automatic to human motion, fast identification.

Description of drawings

Fig. 1 is that example has shown the characteristics of motion two-dimensional geometry feature among the present invention with the sports style of walking;

Fig. 2 is the motion projection of the dimension reduction method of the non-linearity manifold study that adopts among the present invention at three n-dimensional subspace ns;

The hidden Markov model matrix that Fig. 3 the present invention trains human body movement data to obtain, each hidden Markov model have three hiding states, and each state comprises three mixed Gaussian components;

Fig. 4 is that the motion of walking is adopted the method among the present invention and adopted the comparison of the Motion Recognition precision that the method for weak hidden Markov model obtains at human body;

Fig. 5 is at the method among human body boxing employing the present invention and the comparison of the Motion Recognition precision that the method for hidden Markov model obtains a little less than adopting.

Embodiment

The present invention adopts the method for the hidden Markov model of integrated study, the dimension reduction method of the non-linearity manifold study by expansion is realized that the human motion two-dimensional geometry feature of characteristic dimensionality reduction learns, obtain the learner of different motion type, thereby realize the high-precision fast automatic identification of human motion.

In order to realize goal of the invention, this method adopts following technical scheme:

Step 1: from 3 d human motion data, extract a kind of two-dimensional geometry feature that can reflect the local geometric relation:

In order to express human motion effectively, we extract the two-dimensional geometry feature of motion.Therefore utilize the geometric relationship of each articulation point in the motion posture to extract motion characteristics, the two-dimensional geometry feature that obtains can solve the similar problem of logic of exercise data well, we introduce the notion of boolean's feature, and represent: F: ∏ → { 0 with a Boolean function, 1}, obviously any Boolean expression of Boolean function all is an itself.For the vector that f Boolean function formed, we can obtain mixed function F a: ∏ → 0,1} ^fAs mentioned above, we regard F as a fundamental function, and vectorial F (p) then is proper vector or briefly is the feature of posture p ∈ ∏.Fundamental function is applied to a motion capture data D:[1:T by combination F ο D] → ∏ on,

Set j ₁=' root ', j ₂=' lankle ', j ₃=' lhip ', and j ₄=' rtoes ', the two-dimensional structure feature can be expressed as like this $F^r:=F_{\mathrm{plane}}^{(j_1,j_2,j_3;j_4)}.$ j ₁, j ₂, j ₃The plane that is determined, the posture of standing naturally clearly, feature F for a people ^r(P) value is 1.And,, then be characterized as 0 if right crus of diaphragm moves to after one's death or left foot moves to the place ahead for walking or road-work.Exchange features F ^rThe position, the left and right sides of defining point and the plane of overturning towards, we have obtained another one feature F ^lLet us is studied the F:=F of fundamental function combination now ^r∧ F ^l, the value of F is 1 and if only if F ^r, F ^lAll be 1, promptly left foot toe and right crus of diaphragm toe all are positioned at the place ahead of institute's corresponding flat.As can be seen, function F is fit to run and wait this class step shuttling movement type to express walking very much.If exercise data D:[1:T] → ∏ describes is such motion, F ο D has then expressed two peak values of motion cycle so, can obtain the speed of moving very easily by peak value.On the other hand, feature F is for global position with towards, bone size, is constant the different local space distortion (such as leaning to one side) and the vertical moving of shank,

F _Plane ^{J1, j2, j3; J4}In four points can select by various method.Such as, work as j ₁=' root ', j ₂=' l shoulder ', j ₃=' rshoulder ', j ₄=' during lwrist ', whether the character representation left hand is positioned at the front or the back of health.By some suitable compensation, we can make more robust of feature.We by experiment and priori have also defined the two-dimensional geometry feature of a series of other types.With three points define plane different be to obtain reference plane by the given normal vector of two points.Such as, with the vectorial plane orthogonal in from " chest " to " neck ", by this plane, can judge very easily head on neck still below,

An other class geometric properties is then checked two articulation points, two body segment, and perhaps in other words whether articulation point and body segment are exactly close enough whether in a very near zone.What need here to consider contacts with each other such as two hands, and perhaps hand removes the head of body contact, and the situation of leg is so need guarantee to avoid as much as possible mistake to determining an enough big threshold values near the zone.The privileged site that the two-dimensional geometry feature also is used to judge health is such as arm, and whether leg or trunk be crooked or stretch.Such geometric properties is to utilize the angle of suitable body section to represent, such as thigh and shank, and the angle between upper arm and forearm or spine and the left and right sides thigh.Experimental result shows that 120 angle is a good threshold values, and whether the different piece that can be used for distinguishing limbs crooked (＜120) or stretch (＞120).At last, also need the feature with some non-geometry to compensate the full feature set, such as the absolute velocity of some articulation points and relative velocity or the like, two-dimensional geometry feature synoptic diagram is as shown in table 1:

The signal of table 1 two-dimensional geometry feature

Feature geometries	Feature and corresponding expression implication
		F l	F l 1/F l 2Represent a left side/right crus of diaphragm in front, F l 3/F l 4An expression left side/right crus of diaphragm lifts F l 5/F l 6Expression left and right sides knee bends, F l 7/F l 8An expression left side/right crus of diaphragm side is stretched,
F u	Fu 1/Fu 2Represent a left side/right hand in front, F u 3/F u 4An expression left side/right hand lifts F u 5/F u 6Expression right-hand man bending, F u 7/F u 8An expression left side/right-hand side is stretched,
		F m	F m 1/F m 2An expression left side/right hand contacts any pin, F m 3/F m 4An expression left side/right hand contact head or neck, F m 5/F m 6The expression right-hand man contacts buttocks, F m 7The expression comptocormia

Step 2: the dimensionality reduction algorithm that adopts non-linearity manifold study, the higher-dimension raw data is projected in the subspace of a low-dimensional, disclose the immanent structure of human motion by this lower dimensional space, realize the data dimensionality reduction: adopt non-linear manifold learning arithmetic to make up an adjacent map earlier, each point only links to each other with own adjacent point, 2 manifold distance is approached by their Euclidean distance in the neighborhood, 2 manifold distance in same neighborhood is not approached by their the shortest Dijkstra distance on adjacent map, non-subsequently linear manifold learning is based on this adjacency matrix, directly calculate the embedding of low-dimensional stream shape, make following objective function minimize:

$E={\left|\right|D-\tilde{D}\left|\right|}_{L_2}$

D wherein _IjCertain distance measure of point on the expression stream shape,

The statement embedded space point between distance, L ₂The Frobenius norm of representing matrix, ${\left|\right|A\left|\right|}_{L2}=\sqrt{{\mathrm{\&Sigma;}}_{i,j}{A}_{\mathrm{ij}}^2},$ This optimization can solve by classical multi-dimentional scale analytical approach,

The step of non-linearity manifold study algorithm is as follows:

1) set up adjacency matrix G., mainly contain the limit that two kinds of ways are determined adjacency matrix, epsilon neighborhood and K neighborhood. when a j at an i, or some i is when dropping on the ε of a j or K field, setting up a weight between i and j is d _x(i, limit j);

2) utilize Dijkstra or Flyod algorithm, calculation level between bee-line, obtain matrix D _G, all elements wherein all is the shortest path of every pair of point among the G

D _G＝{d _G(i，j)}；

3) setting up d n-dimensional subspace n embeds: allow λ _pBe matrix τ (D _G) preceding p eigenwert (eigenwert by descending sort), v _p ⁱBe i component of p proper vector; The coordinate vector y of d n-dimensional subspace n _iP component just equal

Step 3: adopt study major component feature kernel function to come dimensionality reduction is approached, realize the expansion of non-linearity manifold study algorithm, enable to handle the new exercise data outside the training set:

For the non-linearity manifold study algorithm can be handled new data, we come dimensionality reduction is approached by study major component feature kernel function, allow D={x ₁, x ₂..., x _nBe a sampled data set of unknown distribution, its continuous density is p, allows the P be that its corresponding experience distributes.Consider a Hilbert space

The inner product function be expressed as follows:

<f，g> _p＝∫f(x)g(x)p(x)dx

Here p (x) is a weighting function,

Kernel function K can be in the space like this

A linear operator K neutralizes _pGet in touch as follows:

(K _pf)(x)＝∫K(x，y)f(y)p(y)ddy

We redefine " experience " Hilbert space with experience distribution P replacement unknown distribution p then

Kernel function

Produce a symmetric matrix based on D

Wherein ${\tilde{M}}_{\mathrm{ij}}=\tilde{K}(x_i,x_j).$ Allow proper vector, eigenwert is to (v _k, λ _i) conduct ${\tilde{K}}_p{f}_k={\mathrm{\&lambda;}}_k^{\&prime;}{f}_k$ Separate.We allow like this $e_k\left(x\right)=y_k\left(x\right)\sqrt{{\mathrm{\&lambda;}}_k}$ Be used for representing the mapping of new data x, following equation so just arranged to lower dimensional space:

${\mathrm{\&lambda;}}_k^{\&prime;}=\frac{1}{n}{\mathrm{\&lambda;}}_k,f_k\left(x\right)=\frac{\sqrt{n}}{{\mathrm{\&lambda;}}_k}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{v}_{\mathrm{ik}}\tilde{K}(x,x_i),$

$f_k\left(x_i\right)=\sqrt{n}{v}_{\mathrm{ik}},y_k\left(x\right)=\frac{f_k\left(x\right)}{\sqrt{n}}=\frac{1}{{\mathrm{\&lambda;}}_k}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{v}_{\mathrm{ik}}\tilde{K}(x,x_i)$

$y_k\left(x_i\right)=y_{\mathrm{ik}},e_k\left(x\right)=\sqrt{{\mathrm{\&lambda;}}_k}{y}_k\left(x\right),e_k\left(x_i\right)=e_{\mathrm{ik}}.$

Below we utilize above equation that the non-linearity manifold study algorithm is expanded.D _G(a b) is bee-line between known training data concentrates at 2.It is as follows that we obtain a standard kernel function:

$\tilde{K}(a,b)=-\frac{1}{2}\{D_G(a,b)-E_x[D_G^2(x,b)]-E_{x^{\&prime;}}[D_G^2(a,x^{\&prime;})]+E_{x,x^{\&prime;}}[D_G^2(x,x^{\&prime;})\left]\right\}$

Following like this equation just can be used for expanding the non-linearity manifold study algorithm, makes it can apply on the new data:

$e_k\left(x\right)=\frac{1}{2\sqrt{{\mathrm{\&lambda;}}_k}}\underset{i}{\mathrm{\&Sigma;}}{v}_{\mathrm{ik}}\left\{E_{x^{\&prime;}}\right[{\tilde{D}}^2(x^{\&prime;},x_i)]-{\tilde{D}}^2(x_i,x)\}$

If we remove estimation space with the exercise data of sufficient amount Last K _pFundamental function, so just can obtain the non-training data mapping an of the best,

From the above mentioned, just be extended to can be with new data map to low n-dimensional subspace n for the classical non-linearity manifold study method that can only handle the given data of training set inside originally.

Step 4: adopt hidden Markov model to learn for the exercise data after the dimensionality reduction, obtain hidden Markov model parameter:, be hidden Markov model λ of each type of sports study for j two-dimensional geometry feature based on the common type of sports of two-dimensional geometry feature _i(i=1,2 ..., M); A given observation sequence O is for each hidden Markov model utilizes forward-backward algorithm algorithm computation P (O| λ); Motion Recognition based on j two-dimensional geometry feature is just as follows by the expression formula that searching has the type of sports i of maximum P (O| λ) value to realize:

$\mathrm{action}\left(O\right)=\underset{i:i=1,\&CenterDot;\&CenterDot;\&CenterDot;,M}{\arg \max }(P\left(O\right|{\mathrm{\&lambda;}}_i)$

By above expression formula, M type of sports and N two-dimensional geometry feature have been formed the latent Markov models matrix of a M * N, and the hidden Markov model of j articulation point of i kind type of sports is expressed as HMM _{I, j}, corresponding parameter is λ _{I, j}The pairing Hidden Markov Model (HMM) of row j is gathered the sorter of corresponding j two-dimensional geometry feature,

Need to calculate given state s at t moment observed events O _tProbability P (O _t| s _t=s); Hidden Markov model calculates P (O with probability density function continuously _t| s _t=s), expression formula is as follows:

$\underset{m=1}{\overset{3}{\mathrm{\&Sigma;}}}\left(w_{\mathrm{sm}}\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{\frac{d}{2}}{\left|{\mathrm{\&Sigma;}}_{\mathrm{sm}}\right|}^{\frac{1}{2}}}{e}^{-\frac{1}{2}(O_t-{\mathrm{\&mu;}}_{\mathrm{sm}}){\mathrm{\&Sigma;}}_{\mathrm{sm}}^{-1}{(O_t-{\mathrm{\&mu;}}_{\mathrm{sm}})}^T}\right).$

Step 5: adopt adaptive propulsion method that weak hidden Markov model is strengthened, obtain a reinforced integrated classifier that can reflect the human body integral movable information, realize correctly Motion Recognition efficiently: given n training sample (x ₁, y ₁) ..., (x _n, y _n) training set, y wherein _i=| 0,1|, (i=1,2 ..., n) mistake of corresponding sample identification and correct has u error sample in the sample, v correct sample; The space characteristics for the treatment of each articulation point of branch type games is expressed as f _j(), wherein 1≤j≤16; For i training sample x _i, it be characterized as f _j(x _i); The Weak Classifier h of the space characteristics of j articulation point _j(x) by a feature f _j, a threshold values θ _jBias p with an indication inequality direction _jConstitute:

$h_j\left(x\right)=\left\{\begin{array}{c}1,p_j{f}_j<p_j{\mathrm{\&theta;}}_{\mathrm{<figure-callout\; id="0"\; label="j"\; filenames="A20071006768400231.png"\; state="\{\{state\}\}">j</figure-callout>}}\\ 0,\mathrm{otherwise}\end{array}\right.$

By aligning counter-example analysis, select T minimum Weak Classifier of error rate, optimal combination becomes a strong classifier,

Because being emphasis, the thought of the method for integrated study trains those difficult samples that divides, like this, every take turns loop ends after, increased by the weight of the sample of mis-classification, when entering the next round round-robin, thereby these are wrong to divide samples to be trained by emphasis to make these samples become correctly to be classified easily, but we are when using the integrated study method at physical features, though raising has contribution to discrimination, but the space of improving is also not very big, analyze reason, accuracy main and Weak Classifier itself has very big relation, because there is a large amount of noise sample in the data of physical features, these all are difficult to correctly be classified, therefore can be trained by emphasis, before these noise sample were by correct classification, their weight can be added to a very big value, by contrast, the weight of the sample that those are correctly classified is set to very little number on the contrary, and this runs counter to true criterion in fact, has too despised the correctly work of classification.So we use the two-dimensional geometry feature, thereby eliminate the much noise that physical features exists, made sorter to discern clearly.

Embodiment 1

Training sample comprises 150 various types of motions of walking, and accompanying drawing 4 has provided and utilized the comparison to the Motion Recognition precision of walking of weak hidden Markov model and integrated two kinds of learning methods of hidden Markov model.The concrete steps that following geometry method of the present invention describes this example enforcement in detail are as follows:

(1) extracts all two-dimensional geometry features of the motion of walking with the described method of step 1: extracts a two-dimensional geometry feature and be used for expressing and walk that a right crus of diaphragm toe is positioned at left ankle, the anchor in the place ahead on the plane of left buttocks and root node composition in the motion.We define

1≤i≤4 are four three-dimensional point, wherein＜and p ₁, p ₂, p ₃Represent the determined reference plane of first three point, towards the order that then depends on three points.We are defined as follows then:

By above definition, we are as follows to fundamental function of any four adjacent segment points definition:

$F_{\mathrm{plane}}^{(j_1,j_2,j_3;j_4)}\left(P\right):=B(P^{j_1},P^{j_2},P^{j_3};P^{j_4})$

Wherein $F_{\mathrm{plane}}^{(j_1,j_2,j_3;j_4)}:\mathrm{\&Pi;}\&RightArrow;\left\{\mathrm{0,1}\right\}.$

(2) adopt step 2 to describe the method that non-linear flow pattern learns to the two-dimensional geometry characteristic of human motion and carry out dimensionality reduction, obtain the low n-dimensional subspace n of high dimensional feature data by the spectral factorization to matrix, step is as follows:

1. set up adjacency matrix G, mainly contain the limit that two kinds of ways are determined adjacency matrix, epsilon neighborhood and K neighborhood.When a j at an i, or some i is when dropping on the ε of a j or K neighborhood, setting up a weight between i and j is d _x(i, limit j);

2. utilize Dijkstra or F1yod algorithm, calculation level between bee-line, obtain matrix D _G, all elements wherein all is the shortest path of every pair of point among the G

D _G＝{d _G(i，j)}；

3. setting up d n-dimensional subspace n embeds: allow λ _pBe matrix τ (D _G) preceding p eigenwert (eigenwert by descending sort), v _p ⁱBe i component of p proper vector; The coordinate vector y of d n-dimensional subspace n _iP component just equal

(3) adopt the described method of step 3 that the non-linearity manifold study method is expanded to non-training dataset:

Utilize study major component feature kernel function that non-linearity manifold study is expanded.D _G(a b) is bee-line between known training data concentrates at 2.It is as follows that we obtain a standard kernel function:

$\tilde{K}(a,b)=-\frac{1}{2}\{D_G(a,b)-E_x[D_G^2(x,b)]-E_{x^{\&prime;}}[D_G^2(a,x^{\&prime;})]+E_{x,x^{\&prime;}}[D_G^2(x,x^{\&prime;})\left]\right\}$

Following like this equation just can be used for expanding the algorithm of non-linearity manifold study, makes it can apply on the new data:

$e_k\left(x\right)=\frac{1}{2\sqrt{{\mathrm{\&lambda;}}_k}}\underset{i}{\mathrm{\&Sigma;}}{v}_{\mathrm{ik}}\left\{E_{x^{\&prime;}}\right[{\tilde{D}}^2(x^{\&prime;},x_i)]-{\tilde{D}}^2(x_i,x)\}$

If we remove estimation space with the exercise data of sufficient amount

Last K _pFundamental function, so just can obtain the non-training data mapping an of the best.

(4) adopt the described method of step 4 that human body movement data is learnt with hidden Markov model:

For j two-dimensional geometry feature, we are hidden Markov model λ of each type of sports study _i(i=1,2 ..., M).A given observation sequence O, we calculate P (O| λ) for each hidden Markov model utilizes forward-backward algorithm algorithm (Forward-Backward).Motion Recognition based on j two-dimensional geometry feature just can have the type of sports i of maximum P (O| λ) value to realize by searching, and is as follows:

$\mathrm{action}\left(O\right)=\underset{i:i=1,\&CenterDot;\&CenterDot;\&CenterDot;,M}{\arg \max }(P\left(O\right|{\mathrm{\&lambda;}}_i)$

By above expression formula, M type of sports and N two-dimensional geometry feature have been formed the latent Markov models matrix of a M * N, and as shown in Figure 3, we are expressed as HMM to the hidden Markov model of j articulation point of i kind type of sports _{I, j}, corresponding parameter is λ _{I, j}The pairing Hidden Markov Model (HMM) of row j is gathered the sorter of corresponding j two-dimensional geometry feature.

(5) it is integrated to adopt the described adaptive propulsion method of step 5 that a series of weak hidden Markov learners are carried out, and obtains strong learner, and step is as follows:

Step 1: establish ω _{T, i}It is the error weight of i sample in the t time circulation.Error weight in the training sample is pressed following formula initialization: for y _i=0 sample, ${\mathrm{\&omega;}}_{1,i}=\frac{1}{2u};$ For y _i=1 sample,

${\mathrm{\&omega;}}_{1,i}=\frac{1}{2v};$

The normalization of step 2:for t=1 to T weight makes ${\mathrm{\&omega;}}_{t,i}=\frac{{\mathrm{\&omega;}}_{t,i}}{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{t,j}};$

Step 3: for each articulation point, train the sorter of its weak hidden Markov model, determine threshold values θ _j, bias p _j, make objective function ${\mathrm{\&epsiv;}}_j=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{t,j}|h_j\left(x_i\right)-y_i|$ Reach minimum;

Step 4: from top Weak Classifier, find one to have minimal error ε _tSorter h _t, the weight of all samples is upgraded: ${\mathrm{\&omega;}}_{t+1,i}={\mathrm{\&omega;}}_{t,i}{\mathrm{\&beta;}}_t^{1-{\mathrm{\&epsiv;}}_i};$

Step 5: determine e _iIf, x just now _iBy h _iCorrect identification, e _i=0; Otherwise e _i=1; $\mathrm{\&beta;}=\frac{{\mathrm{\&epsiv;}}_t}{1-{\mathrm{\&epsiv;}}_t}$ The Weak Classifier that can guarantee back training extraction can be strengthened these identification error sample trainings more;

Step 6: obtain at last advancing calculation to send out the strong classifier that the Weak Classifier of all articulation points is integrated based on self-adaptation:

$h\left(x\right)=\left\{\begin{array}{cc}1,& \underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t{h}_t\left(x\right)\&GreaterEqual;0.5\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{\mathrm{<figure-callout\; id="0"\; label="t"\; filenames="A20071006768400231.png"\; state="\{\{state\}\}">t</figure-callout>}}\\ 0,& \mathrm{otherwise}\end{array}\right.$

Wherein ${\mathrm{\&alpha;}}_t=\log \frac{1}{{\mathrm{\&beta;}}_t}.$

(6) adopt the hidden Markov model that trains that new motion is discerned, step is as follows:

Step 1: a given new motion sequence O=O ₁O ₂... O _T

Step 2: search for all type of sports, calculate P (O| λ _{I, j});

Step 3: return the type of sports at maximal value place, finish the identification of the motion of walking.

Embodiment 2

Training sample comprises 110 various types of boxings, and accompanying drawing 5 has provided and utilized the comparison to the boxing accuracy of identification of weak hidden Markov model and integrated hidden Markov model two kinds of learning methods.The concrete steps that following geometry method of the present invention describes this example enforcement in detail are as follows:

(1) extracts a two-dimensional geometry feature with the described method of step 1 and be used for expressing that a right hand is positioned at left shoulder in the boxing, the anchor in the place ahead on the plane that left buttocks and root node are formed.We define

1≤i≤4 are four three-dimensional point, wherein＜and p ₁, p ₂, p ₃, represent the determined reference plane of first three point, towards the order that then depends on three points.We are defined as follows then:

By above definition, we are as follows to fundamental function of any four adjacent segment points definition:

$F_{\mathrm{plane}}^{(j_1,j_2,j_3;j_4)}\left(P\right):=B(P^{j_1},P^{j_2},P^{j_3};P^{j_4})$

Wherein $F_{\mathrm{plane}}^{(j_1,j_2,j_3,j_4)}:\mathrm{\&Pi;}\&RightArrow;\left\{\mathrm{0,1}\right\}.$

(2) adopt step 2 to describe the method that non-linear flow pattern learns to the two-dimensional geometry characteristic of human motion and carry out dimensionality reduction, obtain the low n-dimensional subspace n of high dimensional feature data by the spectral factorization to matrix, step is as follows:

1. set up adjacency matrix G, mainly contain the limit that two kinds of ways are determined adjacency matrix, epsilon neighborhood and K neighborhood.When a j at an i, or some i is when dropping on the ε of a j or K neighborhood, setting up a weight between i and j is d _x(i, limit j);

2. utilize Dijkstra or Flyod algorithm, calculation level between bee-line, obtain matrix D _G, all elements wherein all is the shortest path of every pair of point among the G

D _G＝d _G(i，j)}；

3. setting up d n-dimensional subspace n embeds: allow λ _pBe matrix τ (D _G) preceding p eigenwert (eigenwert by descending sort), v _p ⁱBe i component of p proper vector; The coordinate vector y of d n-dimensional subspace n _iP component just equal

(3) adopt the described method of step 3 that the non-linearity manifold study method is expanded to non-training dataset:

Utilize study major component feature kernel function that non-linearity manifold study is expanded.D _G(a b) is bee-line between known training data concentrates at 2, and it is as follows that we obtain a standard kernel function:

$\tilde{K}(a,b)=-\frac{1}{2}\{D_G(a,b)-E_x[D_G^2(x,b)]-E_{x^{\&prime;}}[D_G^2(a,x^{\&prime;})]+E_{x,x^{\&prime;}}[D_G^2(x,x^{\&prime;})\left]\right\}$

Following like this equation just can be used for expanding the algorithm of non-linearity manifold study, makes it can apply on the new data:

$e_k\left(x\right)=\frac{1}{2\sqrt{{\mathrm{\&lambda;}}_k}}\underset{i}{\mathrm{\&Sigma;}}{v}_{\mathrm{ik}}\left\{E_{x^{\&prime;}}\right[{\tilde{D}}^2(x^{\&prime;},x_i)]-{\tilde{D}}^2(x_i,x)\}$

If we remove estimation space with the exercise data of sufficient amount

Last K _pFundamental function, so just can obtain the non-training data mapping an of the best.

(4) adopt the described method of step 4 that human body movement data is learnt with hidden Markov model:

For j two-dimensional geometry feature, we are hidden Markov model λ of each type of sports study _i(i=1,2 ..., M).A given observation sequence O, we utilize forward-backward algorithm algorithm computation P (O| λ) for each hidden Markov model.Motion Recognition based on j two-dimensional geometry feature just can have the type of sports i of maximum P (O| λ) value to realize by searching, and is as follows:

$\mathrm{action}\left(O\right)=\underset{i:i=1,\&CenterDot;\&CenterDot;\&CenterDot;,M}{\arg \max }(P\left(O\right|{\mathrm{\&lambda;}}_i)$

By above expression formula, M type of sports and N two-dimensional geometry feature have been formed the latent Markov models matrix of a M * N, and as shown in Figure 3, we are expressed as HMM to the hidden Markov model of j articulation point of i kind type of sports _{I, j}, corresponding parameter is λ _{I, j}The pairing Hidden Markov Model (HMM) of row j is gathered the sorter of corresponding j two-dimensional geometry feature.

(5) it is integrated to adopt the described adaptive propulsion method of step 5 that a series of weak hidden Markov learners are carried out, and obtains strong learner, and step is as follows:

Step 1: establish ω _{T, i}It is the error weight of i sample in the t time circulation.Error weight in the training sample is pressed following formula initialization: for y _i=0 sample, ${\mathrm{\&omega;}}_{1,i}=\frac{1}{2u};$ For y _i=1 sample,

${\mathrm{\&omega;}}_{1,i}=\frac{1}{2v};$

The normalization of step 2:for t=1 to T weight makes ${\mathrm{\&omega;}}_{t,i}=\frac{{\mathrm{\&omega;}}_{t,i}}{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{t,j}};$

Step 3: for each articulation point, train the sorter of its weak hidden Markov model, determine threshold values θ _j, bias p _j, make objective function ${\mathrm{\&epsiv;}}_j=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{t,j}|h_j\left(x_i\right)-y_i|$ Reach minimum;

Step 4: from top Weak Classifier, find one to have minimal error ε _tSorter h _t, the weight of all samples is upgraded: ${\mathrm{\&omega;}}_{t+1,i}={\mathrm{\&omega;}}_{t,i}{\mathrm{\&beta;}}_t^{1-{\mathrm{\&epsiv;}}_i};$

Step 5: determine e _iIf, x just now _iBy h _iCorrect identification, e _i=0; Otherwise e _i=1; $\mathrm{\&beta;}=\frac{{\mathrm{\&epsiv;}}_t}{1-{\mathrm{\&epsiv;}}_t}$ The Weak Classifier that can guarantee back training extraction can be strengthened these identification error sample trainings more;

Step 6: obtain at last advancing algorithm, the strong classifier that the Weak Classifier of all articulation points is integrated based on self-adaptation:

$h\left(x\right)=\left\{\begin{array}{cc}1,& \underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t{h}_t\left(x\right)\&GreaterEqual;0.5\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{\mathrm{<figure-callout\; id="0"\; label="t"\; filenames="A20071006768400231.png"\; state="\{\{state\}\}">t</figure-callout>}}\\ 0,& \mathrm{otherwise}\end{array}\right.$

Wherein ${\mathrm{\&alpha;}}_t=\log \frac{1}{{\mathrm{\&beta;}}_t}.$

(6) adopt the hidden Markov model that trains that new motion is discerned, step is as follows:

Step 1: a given new motion sequence O=O ₁O ₂... O _T

Step 2: search for all type of sports, calculate P (O| λ _{I, j});

Step 3: return the type of sports at maximal value place, finish the identification of boxing.

Claims (7)

1. recognition methods based on the human body movement data of integrated hidden Markov model learning method is characterized in that comprising the steps:

(1) from 3 d human motion data, extracts a kind of two-dimensional geometry feature that can reflect the local geometric relation;

(2) the dimensionality reduction algorithm of employing non-linearity manifold study projects to the higher-dimension raw data in the subspace of a low-dimensional, discloses the immanent structure of human motion by this lower dimensional space, realizes the data dimensionality reduction;

(3) adopt study major component feature kernel function to come dimensionality reduction is approached, realize the expansion of non-linearity manifold study algorithm, enable to handle the new exercise data outside the training set;

(4) adopt hidden Markov model to learn for the exercise data after the dimensionality reduction, obtain hidden Markov model parameter based on the common type of sports of two-dimensional geometry feature;

(5) adopt adaptive propelling algorithm, weak hidden Markov model has been set up a reinforced integrated study device, finish identification motion.

2. the recognition methods of a kind of human body movement data based on integrated hidden Markov model learning method according to claim 1, it is characterized in that, described from 3 d human motion data, extract a kind of two-dimensional geometry feature that can reflect the local geometric relation: represent with Boolean function: F: ∏ → 0,1}; For the vector that f Boolean function formed, obtain mixed function F a: ∏ → 0,1} ^fF is regarded as a fundamental function, vector F (p) then is proper vector or briefly is the feature of posture p ∈ ∏, be applied to motion capture data D:[1:T by combination F о D] → ∏ on, wherein, F о D has expressed two peak values of motion cycle, can obtain the speed of moving very easily by peak value, feature F is for global position with towards, bone size, and the different local space distortion and the vertical moving of shank have unchangeability.

3. the recognition methods of a kind of human body movement data based on integrated hidden Markov model learning method according to claim 1, it is characterized in that, the dimensionality reduction algorithm of described employing non-linearity manifold study, the higher-dimension raw data is projected in the subspace of a low-dimensional, disclose the immanent structure of human motion by this lower dimensional space, realize the data dimensionality reduction: adopt non-linear manifold learning arithmetic to make up an adjacent map earlier, each point only links to each other with own adjacent point, 2 manifold distance is approached by their Euclidean distance in the neighborhood, 2 manifold distance in same neighborhood is not approached by their the shortest Dijkstra distance on adjacent map, non-subsequently linear manifold learning is based on this adjacency matrix, directly calculate the embedding of low-dimensional stream shape, make following objective function minimize:

$E={\left|\right|D-\tilde{D}\left|\right|}_{L_2}$

D wherein _IjCertain distance measure of point on the expression stream shape,

The statement embedded space point between distance, L ₂The Frobenius norm of representing matrix, ${\left|\right|A\left|\right|}_{L2}=\sqrt{{\mathrm{\&Sigma;}}_{i,j}{A}_{\mathrm{ij}}^2}.$

4. the recognition methods of a kind of human body movement data based on integrated hidden Markov model learning method according to claim 3 is characterized in that the step of described non-linearity manifold study algorithm is as follows:

1) set up adjacency matrix G., mainly contain the limit that two kinds of ways are determined adjacency matrix, epsilon neighborhood and K neighborhood. when a j at an i, or some i is when dropping on the ε of a j or K neighborhood, setting up a weight between i and j is d _x(i, limit j).

2) utilize Dijkstra or Flyod algorithm, calculation level between bee-line, obtain matrix D _G, all elements wherein all is the shortest path of every pair of point among the G

D _G＝{d _G(i，j)}

3) setting up d n-dimensional subspace n embeds: allow λ _pBe matrix τ (D _G) preceding p eigenwert (eigenwert by descending sort), v _p ⁱBe i component of p proper vector; The coordinate vector y of d n-dimensional subspace n _iP component just equal

5. the recognition methods of a kind of human body movement data based on integrated hidden Markov model learning method according to claim 1, it is characterized in that, described employing study major component feature kernel function comes dimensionality reduction is approached, realize the expansion of non-linearity manifold study algorithm, enable to handle the new exercise data outside the training set: allow D={x ₁, x ₂... x _nBe a sampled data set of unknown distribution, its continuous density is p, allows the P be that its corresponding experience distributes.Consider a Hilbert space The inner product function be expressed as follows:

<f，g> _p＝∫f(x)g(x)p(x)dx

Wherein, p (x) is a weighting function, and kernel function K is in the space like this

A linear operator K neutralizes _pGet in touch as follows:

(K _pf)(x)＝∫K(x，y)f(y)p(y)ddy

Replace unknown distribution p with experience distribution P then and redefine " experience " Hilbert space

Kernel function

Produce a symmetric matrix based on D

Wherein ${\tilde{M}}_{\mathrm{ij}}=\tilde{K}(x_i,x_j);$ Allow proper vector, eigenwert is to (v _k, λ _i) conduct ${\tilde{K}}_p{f}_k={\mathrm{\&lambda;}}_k^{\&prime;}{f}_k$ Separate; Allow $e_k\left(x\right)=y_k\left(x\right)\sqrt{{\mathrm{\&lambda;}}_k}$ Be used for representing the mapping of new data x to lower dimensional space, the expansion non-linearity manifold study makes it can apply on the new data:

$e_k\left(x\right)=\frac{1}{2\sqrt{{\mathrm{\&lambda;}}_k}}\underset{i}{\mathrm{\&Sigma;}}{v}_{\mathrm{ik}}\left\{E_{x^{\&prime;}}\right[{\tilde{D}}^2(x^{\&prime;},x_i)]-{\tilde{D}}^2(x_i,x)\}$

Exercise data with sufficient amount removes estimation space Last K _pFundamental function, just obtain a non-training data mapping.

6. the recognition methods of a kind of human body movement data based on integrated hidden Markov model learning method according to claim 1, it is characterized in that, describedly adopt hidden Markov model to learn for the exercise data after the dimensionality reduction, obtain hidden Markov model parameter:, be hidden Markov model λ of each type of sports study for j two-dimensional geometry feature based on the common type of sports of two-dimensional geometry feature _i(i=1,2 ..., M); A given observation sequence O is for each hidden Markov model utilizes forward-backward algorithm algorithm computation P (O| λ); Motion Recognition based on j two-dimensional geometry feature is just as follows by the expression formula that searching has the type of sports i of maximum P (O| λ) value to realize:

$\mathrm{action}\left(O\right)=\underset{i:i=1,...,M}{\arg \max }\left(P\left(O\right|{\mathrm{\&lambda;}}_i\right)$

By above expression formula, M type of sports and N two-dimensional geometry feature have been formed the latent Markov models matrix of a M * N, and the hidden Markov model of j articulation point of i kind type of sports is expressed as HMM _{I, j}, corresponding parameter is λ _{I, j}The pairing Hidden Markov Model (HMM) of row j is gathered the sorter of corresponding j two-dimensional geometry feature;

Need to calculate given state s at t moment observed events O _tProbability P (O _t| s _t=s); Hidden Markov model calculates P (O with probability density function continuously _t| s _t=s), expression formula is as follows:

$\underset{m=1}{\overset{3}{\mathrm{\&Sigma;}}}\left(w_{\mathrm{sm}}\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{\frac{d}{2}}{\left({\mathrm{\&Sigma;}}_{\mathrm{sm}}\right)}^{\frac{1}{2}}}{e}^{-\frac{1}{2}(O_t-{\mathrm{\&mu;}}_{\mathrm{sm}}){\mathrm{\&Sigma;}}_{\mathrm{sm}}^{-1}{(O_t-{\mathrm{\&mu;}}_{\mathrm{sm}})}^T}\right).$

7. the recognition methods of a kind of human body movement data based on integrated hidden Markov model learning method according to claim 1, it is characterized in that, the adaptive propulsion method of described employing is strengthened weak hidden Markov model, obtain a reinforced integrated classifier that can reflect the human body integral movable information, realize correctly Motion Recognition efficiently: given n training sample (x ₁, y ₁) ..., (x _n, y _n) training set, y wherein _i=| 0,1|, (i=1,2 ..., n) mistake of corresponding sample identification and correct has u error sample in the sample, v correct sample; The space characteristics for the treatment of each articulation point of branch type games is expressed as f _j(), wherein 1≤j≤16; For i training sample x _i, it be characterized as f _j(x _i); The Weak Classifier h of the space characteristics of j articulation point _j(x) by a feature f _j, a threshold values θ _jBias p with an indication inequality direction _jConstitute:

$h_j\left(x\right)=\left[\begin{array}{c}1,p_j{f}_j<p_j{\mathrm{\&theta;}}_j\\ 0,\mathrm{otherwise}\end{array}\right]$

By aligning counter-example analysis, select T minimum Weak Classifier of error rate, optimal combination becomes a strong classifier.

Images