KR20100124532A - Data processing apparatus and method - Google Patents

Abstract

PURPOSE: A data processing apparatus and a method thereof are provided to calculate the whole operation by sensing the movements of points when capturing motions. CONSTITUTION: A processing unit(120) randomly assigns a plurality of input data to clusters. A calculation unit(130) uses the regression analysis to calculate the transformation matrix coping with clusters. The calculation unit calculates the return value applying transformation matrixes to the first data allocated to first cluster. The processing unit assigns the first data based on the return value to the second cluster.

Description

DATA PROCESSING APPARATUS AND METHOD}

Some embodiments of the present invention relate to a data processing apparatus and method for analyzing training data accumulated by motion sensing. More specifically, the present invention relates to a data processing apparatus and method for estimating and generating an operation of a 3D model such as an avatar by analyzing the training data when there is motion sensing.

Sensing the movement of target objects (such as human bodies) in real space, such as video games, virtual worlds, and computer graphics (CGs), and 3-Dimensional (3D). There is a growing interest in technology that can be implemented in space.

In this technique, which can be called various names such as motion estimation or motion capture, a long calculation time from sensing of motion to deriving motion capture results is disadvantageous.

In the case of sensing and capturing a human motion, the larger the number of sensors attached to the human body, the more accurate the result is expected, but this means that the process for sensing is cumbersome and limited. That is, the movement of the human body to which many sensors are attached is limited, and it is difficult to attach or remove the sensors.

Therefore, there is a need to propose a method of attaching a relatively small number of sensors to sense motion and perform motion capture.

Some embodiments of the present invention are directed to providing a data processing apparatus and method for calculating the overall motion by sensing only a relatively small number of points of motion in motion capture.

In addition, some embodiments of the present invention provide more efficient and accurate clustering when a plurality of highly correlated data are inputted to establish a computational model for motion capture, thereby increasing the accuracy of the result value. It is to carry out.

According to an embodiment of the present invention, a processing unit for randomly allocating a plurality of input data to k clusters (where k is a natural number) and corresponding to each of the k clusters using regression analysis A data processing device including a calculation unit for calculating k transform matrices and calculating a result of applying each of the k transform matrices to first data allocated to a first cluster among the k clusters is provided.

In this case, the processor allocates the first data to a second cluster among the k clusters based on the calculated result value.

Here, each of the plurality of input data is physical information of n points (n is a natural number) detected in one frame of motion sensing, and physical information of n1 points (n1 is a natural number) among the n points. It may include.

The regression analysis may be a nonlinear analysis based on KCCA (Kernel Canonical Correlation Analysis) or a linear analysis based on CCA (Canonical Correlation Analysis).

According to an embodiment of the present invention, the processor determines, based on the calculated result value, a second cluster corresponding to a transform matrix that provides a result value most similar to the first data among the k transform matrices. The first data is allocated to the second cluster.

After the first data is allocated to the second cluster, the calculator calculates and updates a transform matrix corresponding to the first cluster and a transform matrix corresponding to the second cluster by using regression analysis. can do.

In this case, the calculation unit calculates and updates a transformation matrix corresponding to the first cluster except for the first data, and recalculates and updates the transformation matrix including the first data and corresponding to the second cluster. Can be.

According to an embodiment of the present invention, the data processing apparatus further includes an interface for receiving second data including physical information of n1 feature points detected by motion sensing.

In this case, the processing unit selects at least one cluster having a high correlation with the second data among the k clusters, and the calculating unit selects a transform matrix corresponding to each of the selected at least one cluster from the second cluster. The physical information of the n points corresponding to the second data is calculated based on the result of applying the data.

According to an embodiment of the present invention, the calculation unit may apply weights proportional to the correlation with the second data to the result values of applying each transformation matrix of the at least one selected cluster to the second data. The physical information of the n points corresponding to the second data is calculated as the linear sum of the weighted result values.

According to another embodiment of the present invention, randomly allocating a plurality of input data to k clusters (where k is a natural number) and k corresponding to each of the k clusters using regression analysis Calculating a transformation matrix, and applying the first data to a second cluster of the k clusters based on a result of applying each of the k transformation matrices to first data allocated to a first cluster of the k clusters. A data processing method is provided that includes assigning to.

Here, the regression analysis may be a nonlinear analysis based on KCCA, or the regression analysis may be a linear analysis based on CCA.

According to some embodiments of the present invention, in motion capture, the entire motion can be calculated by sensing only a relatively small number of points of motion, which is easy to implement.

Further, according to some embodiments of the present invention, in clustering a plurality of highly correlated data among a plurality of input training data input to establish a calculation model for motion capture, the clustering is performed more efficiently and accurately.

Hereinafter, some embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention is not limited or limited by the embodiments. Like reference numerals in the drawings denote like elements.

1 illustrates a data processing apparatus 100 according to an embodiment of the present invention.

According to an embodiment of the present invention, the data processing apparatus 100 includes an I / O interface 110. The interface 110 is a path for exchanging sensed learning data, data to be captured, and result values with the outside.

The training data for calculating the transformation matrix for the motion estimation is input through the interface 110 and first stored in the first storage 140.

According to an embodiment of the present invention, the processor 120 first randomly allocates a plurality of learning data stored in the first storage unit! 40 to k clusters (k is a natural number) and randomly allocates the second training unit. Save to 150. In this case, it may not be guaranteed that similar learning data are allocated to the same cluster.

The calculator 130 calculates a transformation matrix of each of the k clusters. A detailed process of obtaining the transformation matrix will be described later in detail with reference to FIG. 4.

The processor 120 selects any one of the k clusters of learning data (first learning data) in the first cluster. In addition, when the calculator 130 applies the first training data to the transformation matrix for each of the k clusters and provides the result, the processor 120 compares the result.

If there is no transform matrix that provides a result value more similar to the transform matrix for the first cluster, the processor 120 maintains the first training data in the first cluster.

However, if there is a transform matrix that provides a result value that is more similar than the transform matrix for the first cluster, the processor 120 selects a second cluster having the result value most similar to the first training data, and selects the first cluster. Assign training data to the second cluster.

According to an embodiment of the present invention, in this case, the transformation matrix of the first cluster and the second cluster is updated. Specifically, in the case of the first cluster, the transform matrix is obtained again only with the remaining learning data belonging to the first cluster except the first learning data. The newly obtained transform matrix replaces the existing transform matrix.

In addition, in the case of the second cluster, a transformation matrix is obtained again by using all the training data belonging to the second cluster by newly including the first training data. Again, the newly obtained transformation matrix replaces the existing transformation matrix.

The data processing apparatus 100 performs this process until it satisfies a predetermined termination condition (for example, performing the above process once for all learning data, or changing the cluster no longer). Repeat

Then, each of the clusters has a transformation matrix generated by binding training data having high similarity.

Hereinafter, more detailed embodiments will be described in more detail with reference to FIG. 4 and below.

2 is a conceptual diagram illustrating a conventional inverse kinematics method.

In the past, an inverse kinematics method was used to estimate the motion. To briefly explain this, a process of estimating the movement of the plurality of points 210, 220, 221, and 222 in a two-dimensional space will be described.

The position between two neighboring points of the root point 210, the leaf point 220, and the points 221 and 222 therebetween are related to each other.

For example, the position of the point 210 and the point 221 is a distance from each other and only the relative position (angle) only changes. The same is true of point 221 and point 222, or point 222 and point 220.

Thus, assuming only two-dimensional space, if the location of the root point 210 and the distance between neighboring points among the points are known in advance, then the location of the points 221, 222, and 220 is Only three angles are given.

On the other hand, given the location of the point 220, the point 221 and the point 222 may be obtained. More details are described below with reference to FIG. 3.

FIG. 3 shows motion estimation by the Inverse Kinematics method of FIG. 2.

The distance between the root point 310 and the point 321 is fixed as l ₁ , the distance between the point 321 and the point 322 is l ₂ , and the distance between the point 322 and the point 320 is fixed as 320.

In this case, l ₁ , l ₂ , and l ₃ are constants, and assuming the root point 310 as the origin, the coordinates (x, y) of the point 320 are represented by the following equation.

[Equation 1]

Where additional information (such as x> 0, y> 0, θ ₁ > 0, θ ₂ <0, θ ₃ > 0) is present, θ ₁ , θ ₂ , and The position of θ ₃ can be obtained as the only solution, and thus the positions of the points 321 and 322 can also be specified.

That is, even if only the relative coordinates of the point 320 with respect to the root point 310 is known, the positions of the remaining points 321 and 322 can be calculated. This process is called Inverse Kinematics. In the above example, the physical information of the point 321, point 322, using the physical information (relative coordinates (x, y)) of one point 320 (except the root point 310) by the Inverse Kinematics method Information (coordinates) was available.

However, in the case of simple analysis in two dimensions, such an inverse kinematics method may be applied, but as the dimension increases and the amount of input data increases, the inverse kinematics method encounters limitations.

Accordingly, according to one embodiment of the present invention, regression analysis based on Canonical Correlation Analysis (CCA) on linear data, or Kernel Canonical Correlation Analysis (KCCA) on non-linear data By regression analysis based on, a transformation matrix for estimating a large number of points of physical information from a small number of points of physical information is obtained using training data. More details will be described later with reference to FIG. 4.

4 illustrates a plurality of points included in data of one frame input by motion sensing processed according to an embodiment of the present invention.

According to an embodiment of the present invention, in one frame, physical information at 15 points located in a main joint of a person is sensed and input. However, the present invention is not limited thereto, and the number of points per frame may be modified as much as the exemplary embodiment.

In this embodiment, points included in one frame are related to the root point 400.

Point 411, point 441 and point 451 are each associated with root point 400. Also, points 410, points 421, and 431 are associated with point 411, point 442 is associated with point 441, and point 452 is associated with point 451, respectively. .

Further, point 420 is associated with point 422, point 430 with point 432, point 440 with point 442, and point 450 with point 452.

The relationship between the associated points may be considered as a hierarchical relationship in which the root point 400 is a parent node. In this case, some of the points 410, 420, 430, 440, and 450 may be designated as feature points.

When the correlation of the physical information between the feature point points and all points is learned, when only the physical information of the feature point points is sensed and input, it is possible to calculate the physical information of all points.

According to some embodiments of the present invention, through the correlation learning, a transformation matrix is generated between physical information of feature point points and physical information of all points.

Meanwhile, the physical information of each point may include a position, a speed, an acceleration, a direction (angle), an angular velocity, an angular acceleration, and the like.

According to an embodiment, it is possible to include only some of the physical information, and the physical information may include only a direction (angle), an angular velocity, and an angular acceleration with respect to a higher point.

According to one embodiment of the invention, the relationship between each point is learned. When one frame data shown in FIG. 4 is input to the data processing apparatus 100 of FIG. 1, the processor 120 may convert the one frame data (hereinafter, simply referred to as “data”) into a first storage unit. Save to 140.

When a plurality of such frame data is input, the processor 120 stores the input frame data in the first storage unit 140.

In addition, the processor 120 divides a plurality of input data (also referred to as training data, or the same below) stored in the first storage unit 140 into a plurality of groups (hereinafter, referred to as "clusters").

According to an embodiment of the present invention, the processor 120 randomly distributes the plurality of learning data stored in the first storage unit to k clusters (where k is a natural number). In this case, the correlation between each learning data may not be considered.

The calculation unit 130 calculates k transformation matrices corresponding to each cluster by regression analysis of the training data belonging to each of the k clusters.

As described above, the transformation matrix may be an operator or a function for estimating the total point using only some feature points among all the points required for capturing motion of one frame. )to be.

Hereinafter, a process of obtaining any one of the k transform matrices corresponding to each cluster by using regression analysis, particularly regression analysis based on CCA or KCCA, will be described with reference to equations.

(1) CCA Based Regression

Canonical Correlation Analysis (CCA) is one of the machine learning algorithms that maximizes the correlation between training data consisting of input and output pairs. To do this, the dimensions of the input and output values are reduced to maximize the correlation between the transformed input and output.

The reason for using the CCA is because the correlation between the reduced input and output through the CCA has been maximized, and when the regression analysis is performed on the reduced input and output through the CCA, This is because overfitting can be avoided rather than regression analysis.

According to an embodiment of the present invention, since the input value (some feature point information) and the output value (all point information) of the training data have a nonlinear relationship, Kernel Trick may be applied to CCA to reconstruct a problem in nonlinear space. Can be.

(2) Kernel Method

In the learning step of storing a determinant of a relationship between input data (some feature point information) inputted by 3D motion sensing and motion capture data (full point information), in an embodiment in which these two data have a nonlinear relationship, The application of the CCA method of learning linear relationships is limited. Therefore, we use Kernel CCA (KCCA) for learning nonlinear relationships.

를 입력 데이터 보다 더 높은 차원인 feature space로 변환하는 매핑(mapping)을 나타내는 식이다.Here is the input data

This expression represents the mapping to transform the feature into a feature space that is higher than the input data.

[Equation 2]

The difference between the CCA and the Kernel CCA is that the CCA thinks of problems in the original dimension of the learning data and the Kernel CCA thinks of problems in a higher dimension, feature space.

To avoid converting the input data to a higher dimension, we make the original problem the product of the two vector dot products. In this case, use a kernel function.

는 feature space에 존재하는 두 벡터의 내적,

이다. 예를 들면 입력데이터

가 주어져 있을 때, 입력 데이터를 2차 monomial 형태로 변환하면

이 된다. 이때 커널 함수는 다음과 같이 나타낼 수 있다.Kernel functions

Is the dot product of two vectors in feature space,

to be. For example input data

Given, converting the input data to the second order monomial form

Becomes In this case, the kernel function can be expressed as:

&Quot; (3) &quot;

(3) KCCA (Kernel CCA)

개의 입력 데이터 쌍을

라 하고, 여기서

,

이며,

는 i 번째 학습의 입력 값,

는 i 번째 학습의 출력 값이다. 학습의 입력, 출력 값을 다음과 같은 행렬의 형태로 나타내자.

Input data pairs

Where,

,

,

Is the input value of the i th training,

Is the output of the i th training. Let's express the input and output values for learning in the form of a matrix:

&Quot; (4) &quot;

를 최대로 하는 식은 다음과 같이 나타낼 수 있다.In CCA learning linear relationships, the correlation of two data

The equation for maximizing can be expressed as

[Equation 5]

,

이고,

이다.In Equation 5,

,

ego,

to be.

,

,

를 다음과 같이 나타낼 수 있다.However, in Kernel CCA, which has raised the dimension,

,

,

Can be expressed as

&Quot; (6) &quot;

,

가 비선형 변환이 된 더 높은 차원에서 기저(basis) 쌍인

,

는,

와

가 이루는 공간에 존재하기 때문에,

,

로 나타낼 수 있다.Also

,

Is a base pair in the higher dimension with nonlinear transformation

,

Quot;

Wow

Because it exists in the space that

,

It can be represented as.

를 최대로 하는 식은 다음과 같이 나타낼 수 있다.In this case, the correlation of the two data in the Kernel CCA

The equation for maximizing can be expressed as

[Equation 7]

이고,

이다.here,

ego,

to be.

은, SVD(singular value decomposition)를 통해서 다음과 같이 구할 수가 있다.In this case, the i th

Can be obtained through SVD (singular value decomposition) as follows.

[Equation 8]

[Equation 9]

,

는 각각

번째의

,

열 벡터(column vector)이다. 위와 같은 방법으로

,

의 쌍을 여러 개 얻을 수가 있는데, 최대

개를 얻을 수 있다.here

,

Respectively

Th

,

It is a column vector. In the same way as above

,

You can get multiple pairs of up to

You can get a dog.

개의

,

를 추출한다. 추출한

,

들을 행렬 형태로 나타내면 아래와 같다.In order to reduce the dimension of the input data,

doggy

,

Extract Extracted

,

These are expressed in matrix form as follows.

[Equation 10]

를 basis

에 사영(projection) 한다.And input value

Basis

Project on.

[Equation 11]

는 다음과 같이 계산된다.Therefore, the input data with reduced dimensions

Is calculated as follows.

[Equation 12]

는 커널 함수들로 이루어진 벡터로 이 벡터의

번째 원소는

이다.This tillage,

Is a vector of kernel functions.

The first element

to be.

(4) Regression

According to an embodiment of the present invention, regression may include three steps.

개의 입력 쌍

가 있을 때, Kernel CCA 방법으로

개의 기저들을 얻어낸다.first,

Input pairs

When there is, Kernel CCA way

Get the bases of the dogs.

,

에서

,

로 mapping 하는 행렬

를 상기 최소자승법을 통해서 구하게 된다.And,

,

in

,

Matrices mapped by

Is obtained by the least square method.

[Equation 13]

를 구한다. 이 경우는 전 단계와 마찬가지로 최소자승법으로 구한다.If so far we have obtained the transformation matrix in the reduced dimension, then

. In this case, the least square method is obtained as in the previous step.

[Equation 14]

,

,

는 연속적인 선형 mapping이므로, 하나의 연산자

로 만들 수 있다.So obtained

,

,

Is a continuous linear mapping, so one operator

Can be made with

가 있다면 원하는 출력 값

는 아래 수학식에 의해 계산될 수 있다.So new input value

Output value you want

Can be calculated by the following equation.

[Equation 15]

연산자가 상기한 변환 행렬이다.Obtained here

The operator is the transformation matrix described above.

According to an embodiment of the present invention, the calculation unit 130 of FIG. 1 obtains the transformation matrix for each of the k clusters. Hereinafter, the process will be described with reference to FIG. 5.

5 is a conceptual diagram illustrating input data allocated to a plurality of clusters according to an embodiment of the present invention.

The training data 510 to 560 input by motion sensing sense various motions of a person.

The training data 510 and the training data 520 are obtained during a similar first action (eg, a boxing match). In addition, the training data 530 and the training data 540 are obtained during a similar second action (eg, snowboarding). In addition, the training data 550 and the training data 560 are acquired during similar third behaviors (eg, swimming).

In such a case, it may be required to obtain each transformation matrix by dividing it into a plurality of clusters rather than obtaining one transformation matrix through the entire training data such as the training data 510 to the training data 560.

For example, the learning data estimated as the first behavior such as the learning data 510 and the learning data 520 as a first cluster, and the learning estimated as the second behavior such as the learning data 530 and the learning data 540. The data is allocated to the second cluster, and the training data estimated by the third behavior such as the training data 550 and the training data 560 is allocated to the third cluster.

According to an embodiment of the present invention, similarities are clustered by comparing the inputted plurality of training data, and a transform matrix for each cluster is calculated.

However, according to another exemplary embodiment of the present invention, the processor 120 of FIG. 1 first randomly allocates a plurality of learning data stored in the first storage 140 to k clusters, and thereby the second storage 150. ).

In this case, it is not guaranteed that the training data 510 and the training data 520 are allocated to the same cluster. For example, the training data 510 and the training data 530 may be allocated to the same first cluster, and the training data 520 and the training data 540 may be allocated to the same second cluster.

According to an embodiment of the present invention, the calculation unit 130 obtains each transformation matrix for the first cluster, the second cluster, and the like in this state.

Then, the processor 120 selects any learning data in the first cluster. For example, the processor 120 selects the training data 530. In this case, the selection of the training data may be sequential or random.

In addition, the calculator 130 calculates a result value of applying the selected training data 530 to the transform matrix for the first cluster, a result value applied to the transform matrix for the second cluster, and the like.

In this case, the processor 120 compares the calculated result values, and if there is no transform matrix providing a result value more similar to the transform matrix for the first cluster, the training data 530 remains in the first cluster. Keep it.

However, if there is a transform matrix that provides a result value that is more similar than the transform matrix for the first cluster, then select a second cluster having the result value most similar to the learning data 530 and select the learning data 530. Assign to the second cluster.

In the case of the first cluster, except for the training data 530, the transformation matrix is obtained again only with the remaining training data belonging to the first cluster. The newly obtained transform matrix replaces the existing transform matrix.

In addition, in the case of the second cluster, a transformation matrix is obtained again by using the new learning data 530 and all learning data belonging to the second cluster. Again, the newly obtained transformation matrix replaces the existing transformation matrix.

This process is iteratively repeated until a predetermined termination condition (eg, the process has been performed once for all learning data, or no more cluster changes occur).

Then, each of the clusters has a transformation matrix generated by binding training data having high similarity.

6 illustrates a process of acquiring data including physical information of a plurality of points by motion sensing, according to an embodiment of the present invention.

When k clustering and transformation matrices are obtained as described above, sensing of some characteristic points may be performed in the exemplary configuration as shown in FIG. 6.

Sensor 601 analyzes the physical information of characteristic points 610-650 of person 602. For example, direction, angular velocity, angular acceleration.

Then, the processor 120 of FIG. 1 selects a transformation matrix of one cluster having a high similarity among the k clusters, for data including the five points 610 to 650. The calculation unit 130 applies the selected transformation matrix to the data, and calculates a value for estimating the operation of all points (eg, 15).

For example, when data including the five points 610 to 650 is input, all point information of FIG. 4 may be calculated.

In some embodiments of the present invention, one cluster having the highest similarity among the k clusters is selected to apply only a transformation matrix for the cluster.

However, the present invention is not limited thereto, and by selecting two or more clusters of the k clusters, applying a transformation matrix of each selected cluster to the data, and multiplying the weights according to the similarities to obtain a linear sum, thereby obtaining all points. You can also calculate the information.

7 illustrates a data processing method according to an embodiment of the present invention.

In step S710, the input plurality of learning data is first randomly assigned to k clusters.

In this case, it may not be guaranteed that similar learning data are allocated to the same cluster.

In step S720, a transformation matrix of each of the k clusters is calculated. The detailed process of obtaining the transformation matrix is as described above with reference to FIG. 4.

In step S730, one of the k clusters, one learning data (first learning data) in the first cluster is selected.

Then, in step S740, the first training data is applied to the transformation matrix for each of the k clusters and the results are compared.

In step S750, if there is no transform matrix that provides a result value that is more similar than the transform matrix for the first cluster, the first training data remains in the first cluster. In operation S780, it is determined whether a predetermined termination condition is satisfied.

However, if there is a transform matrix in step S750 that provides a result value that is more similar than the transform matrix for the first cluster, then a second cluster having the result value most similar to the first training data is selected, and the first The training data is allocated to the second cluster (step S760).

In operation S770, the related transformation matrices are updated. Specifically, in the case of the first cluster, the transform matrix is obtained again only with the remaining learning data belonging to the first cluster except the first learning data. The newly obtained transform matrix replaces the existing transform matrix.

In addition, in the case of the second cluster, a transformation matrix is obtained again by using all the training data belonging to the second cluster by newly including the first training data. Again, the newly obtained transformation matrix replaces the existing transformation matrix.

This process is iteratively repeated until a predetermined termination condition (eg, the process has been performed once for all learning data, or no more cluster changes occur).

Then, each of the clusters has a transformation matrix generated by binding training data having high similarity.

The above process and the following detailed embodiments of the present invention are as described above with reference to FIGS. 4 to 6.

Method according to an embodiment of the present invention is implemented in the form of program instructions that can be executed by various computer means may be recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the media may be those specially designed and constructed for the purposes of the present invention, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

As described above, the present invention has been described by way of limited embodiments and drawings, but the present invention is not limited to the above embodiments, and those skilled in the art to which the present invention pertains various modifications and variations from such descriptions. This is possible.

Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims below but also by the equivalents of the claims.

1 shows a data processing apparatus according to an embodiment of the present invention.

2 is a conceptual diagram illustrating a conventional inverse kinematics method.

FIG. 3 shows motion estimation by the Inverse Kinematics method of FIG. 2.

FIG. 4 illustrates physical information of a plurality of points included in input data by motion sensing processed according to an embodiment of the present invention.

5 is a conceptual diagram illustrating input data allocated to a plurality of clusters according to an embodiment of the present invention.

6 illustrates a process of acquiring data including physical information of a plurality of points by motion sensing, according to an embodiment of the present invention.

7 illustrates a data processing method according to an embodiment of the present invention.

Claims (17)

A processing unit for randomly allocating a plurality of input data to k clusters (where k is a natural number); And Regression analysis is used to calculate k transform matrices corresponding to each of the k clusters, and apply each of the k transform matrices to first data allocated to a first cluster of the k clusters. Calculation unit for calculating the result value Including, And the processing unit allocates the first data to a second cluster of the k clusters based on the calculated result value. The method of claim 1, Each of the plurality of input data includes physical information of n points (n is a natural number) detected in one frame of motion sensing, and n1 (n1 is a natural number) of the n points. A data processing device comprising the physical information of the. The method of claim 1, The regression analysis is a data processing apparatus that is a nonlinear analysis based on Kernel Canonical Correlation Analysis (KCCA). The method of claim 1, The regression analysis is a linear processing based on Canonical Correlation Analysis (CCA). The method of claim 1, The processor determines, based on the calculated result value, a second cluster corresponding to a transform matrix that provides a result value most similar to the first data among the k transform matrices, and converts the first data into the first data. Data processing unit to allocate to 2 clusters. The method of claim 1, After the first data is allocated to the second cluster, the calculation unit calculates and updates a transform matrix corresponding to the first cluster and a transform matrix corresponding to the second cluster by using regression analysis. Data processing unit. The method of claim 6, The calculation unit, Calculating and updating a transformation matrix corresponding to the first cluster except for the first data, and recalculating and updating the transformation matrix including the first data and corresponding to the second cluster. Data processing unit. The method of claim 2, The apparatus may further include an interface configured to receive second data including physical information of n1 feature points detected by motion sensing. The processor selects at least one cluster having a high correlation with the second data among the k clusters, And the calculation unit calculates physical information of n points corresponding to the second data based on a result of applying a transformation matrix corresponding to each of the selected at least one cluster to the second data. The method of claim 8, The calculation unit may apply a weight proportional to a correlation with the second data to the result values of applying each transform matrix of the selected at least one cluster to the second data, and apply the weights of the result values to which the weight is applied. And calculate physical information of n points corresponding to the second data as a linear sum. Randomly assigning a plurality of input data to k clusters, where k is a natural number; Calculating k transformation matrices corresponding to each of the k clusters using regression analysis; And Allocating the first data to a second cluster of the k clusters based on a result of applying each of the k transformation matrices to first data allocated to a first cluster among the k clusters; Data processing method comprising a. The method of claim 10, Each of the plurality of input data includes physical information of n points (n is a natural number) detected in one frame of motion sensing, and n1 (n1 is a natural number) of the n points. Data processing method that includes the physical information of. The method of claim 10, The regression analysis is a non-linear analysis based on Kernel Canonical Correlation Analysis (KCCA). The method of claim 10, The regression analysis is a linear analysis based on Canonical Correlation Analysis (CCA). The method of claim 10, Allocating the first data to a second cluster, Calculating result values of applying the first data allocated to the first cluster among the k clusters to each of the k transformation matrices; Determining a second cluster of the k transform matrices corresponding to the transform matrix that provides a result value most similar to the first data; And Allocating the first data to the second cluster Data processing method comprising a. The method of claim 10, Regression analysis is performed to calculate and update a transformation matrix corresponding to the first cluster except for the first data, and to recalculate and update the transformation matrix including the first data and corresponding to the second cluster. The data processing method further comprises a step. The method of claim 11, When second data including physical information of n1 feature points detected by motion sensing is input, Selecting at least one cluster having a high correlation with the second data among the k clusters; And Calculating physical information of n points corresponding to the second data based on a result of applying a transformation matrix corresponding to each of the selected at least one cluster to the second data; Data processing method comprising a. The method according to any one of claims 10 to 16, A computer-readable recording medium containing instructions for performing the data processing method.

Images