JP2015111332A - Attitude detection device, attitude detection method, and attitude detection program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To compare measurement data with teacher data to quickly determine whether or not both categories are the same.SOLUTION: There is disclosed an attitude detection device. The attitude detection device determines a first rotation moment matrix being a variance-covariance matrix from a matrix obtained from observation data showing a first attitude; determines a left singular vector and a right singular vector obtained by performing singular value decomposition of a second rotation moment matrix, which is obtained by adding the variance-covariance matrices to each of a plurality of pieces of teacher data over the plurality of pieces of teacher data, and first similarity between the plurality of pieces of teacher data and observation data from a matrix obtained by using the first rotation moment matrix; and compares the first similarity with second similarity between the plurality of pieces of teacher data obtained by using matrix elements of a specific value matrix obtained by performing specific value decomposition of the second rotation moment matrix, so as to determine whether or not the first attitude belongs to a category of attitude shown in the plurality of pieces of teacher data.

Description

  The present invention relates to an attitude detection device, an attitude detection method, and an attitude detection program.

  In providing services to users in various fields, it is required to select and provide appropriate service contents according to the user or the user situation. Along with this, the importance of sensing technology for knowing the state of a user is increasing more than ever.

  On the other hand, devices (for example, Kinect (registered trademark)) that can sense the coordinates of representative positions of people have appeared as a device for detecting the posture of a user, instead of motion capture.

  By using such a device, it has become possible to easily observe the three-dimensional coordinates (hereinafter sometimes simply referred to as posture) of the user's skeleton model. By using these devices, posture data can be easily obtained as effective information for obtaining the user's state.

  For the method of identifying the posture data acquired by the device (sometimes called measurement data or observation data), the observation data is compared with the teacher data acquired in advance. Is generally a method for determining which posture corresponds to the teacher data. As such a method, for example, a pattern recognition method including posture determination by competitive learning using a mutual subspace method is known (for example, Non-Patent Document 1 and Patent Document 1). In this method, the degree of similarity is determined based on the canonical angle of the subspace spanned by the basis vector of the observation data and the subspace spanned by the base vector of the teacher data.

  Also, in the subspace method, learning the quadratic function from the feature pattern group for dictionary generation using a polynomial neural network, and selecting the subspace that stores the main component of the quadratic function, thereby reducing the dimension of the feature space A method is known (for example, Patent Document 2).

JP 2012-181568 A JP 2010-39778 A

Noriyoshi Kurosawa and Kenichi Kanaya, “Separation of Moving Objects by Subspace Separation Method and Model Selection” IPSJ Research Report CVIM [Computer Vision and Image Media] 124-4 (November 21, 2000) 2000 (106), pp . 25-32

  However, in general, the method of determining whether the category indicated by the observation data (for example, the posture indicated by the data) and the category indicated by the teacher data are the same by comparing the observation data with the teacher data is an iterative calculation. Therefore, there is a problem that the processing time is long.

  Therefore, as one aspect, the present invention provides a posture detection apparatus and posture detection that quickly determines whether the category indicated by the observation data is the same as the category indicated by the teacher data by comparing the observation data with the teacher data. It is an object to provide a method and an attitude detection program.

  An attitude detection device is disclosed. The posture detection device obtains a first rotational moment matrix that is a variance-covariance matrix from a matrix obtained from observation data representing the posture of an object, and a variance-covariance matrix for each of a plurality of teacher data is distributed over the plurality of teacher data. From the matrix obtained using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix obtained by addition, and the first rotational moment matrix, the plurality of teacher data, A plurality of similarity calculation units for obtaining a first similarity between the observation data and a matrix element of a singular value matrix obtained by singular value decomposition of the second rotational moment matrix; Whether the posture belongs to the posture category represented by the plurality of teacher data by comparing the second similarity between the teacher data and the first similarity Characterized in that it comprises a determining similarity posture determining unit.

  By comparing the observation data with the teacher data, it can be determined at high speed whether the category indicated by the observation data is the same as the category indicated by the teacher data.

It is a figure which shows the example of the functional block of the attitude | position detection apparatus of embodiment. It is a figure which shows the example of a structure of the attitude | position detection apparatus of embodiment. It is a figure which shows the example of the flow of a pre-processing. It is a figure which shows the example of the flow of a process of the attitude | position detection method of embodiment. It is a figure explaining the example of the formulation in the example of application to action pattern analysis. It is a figure explaining the example of the formulation in the example of application to action pattern analysis.

  In the following, a posture detection device, a posture detection method, and a posture detection program for observing the coordinates of each part of a person and determining what kind of posture (pose) the person is based on the result of the observation will be described. However, it can also be applied to determinations other than the detection of the posture of a person. For example, the present invention can be applied to any object other than animals other than humans, such as still life, robots, rockets, automobiles, airplanes, ships, and other machines. Here, the posture of the person may be classified into a plurality of categories. As an embodiment, an example in which an algorithm used in the posture detection apparatus can be applied to human behavior prediction will be described. Also about an Example, action prediction is not limited to a human thing. It may be an animal other than a person.

  First, a posture detection apparatus that detects the posture of a user by observing the coordinates of a plurality of parts of the user with a sensor such as a camera and comparing the observation results with teacher data will be described.

<Comparison example of posture detection>
First, a comparative example of posture detection will be described.

The coordinates (x _i , y _i , z _i ) (i = 1 to n) of the n parts of the user are observed. n may be an integer of about 10, more or less. Each part is represented as coordinates in a three-dimensional space. The following parameter matrix X is generated from n coordinates (x _i , y _i , z _i ).

That is, the parameter matrix X is a matrix obtained by expressing the coordinates (x _i , y _i , z _i ) of the i-th part as vertical vectors and horizontally arranging n vertical vectors. The parameter matrix X is a rank 3 3 × n matrix.

  However, this parameter matrix X is affected by rotation, translation, enlargement, etc. depending on the position of the sensor. Therefore, a variance covariance matrix M is obtained from the parameter matrix X. The variance-covariance matrix M is invariant under rotation and translation.

によって定義すると、分散共分散行列Ｍは、

である。分散共分散行列Ｍは、ランク３のｎ×ｎ行列である。以下でＴは行列の転置を表すことがある。 The coordinates x _av , y _av , and z _av are averages of the coordinates x _i , y _i , and z _i (i = 1 to n), respectively. That is, (x _av , y _av , z _av ) is the center of gravity of n coordinates (x _i , y _i , z _i ). Matrix X ′

The variance-covariance matrix M is defined as

It is. The variance-covariance matrix M is a rank-3 n × n matrix. In the following, T may represent the transpose of a matrix.

The variance-covariance matrix M is subjected to eigenvalue decomposition (or singular value decomposition) to obtain a subspace spanned by eigenvectors (singular vectors). In the above example, since the rank of the parameter matrix X is 3, the rank of the variance-covariance matrix M is also 3. Then, by performing eigenvalue decomposition (or singular value decomposition) on the variance-covariance matrix M, a subspace spanned by three eigenvectors (singular vectors) is obtained. That is, the partial space is uniquely determined with respect to the coordinates (x _i , y _i , z _i ) (i = 1 to n) of the n parts of the user.

This subspace includes information regarding the coordinates (x _i , y _i , z _i ) of n parts of the user. The relationship between the subspace spanned by the three eigenvectors and the user's posture will be described.

とする。定義より、行列Λを特異値行列（固有値行列）として、

である。さらに、

The three eigenvectors of the variance-covariance matrix M are a ₁ = (a ₁₁ , a ₁₂ ,..., A _1n ), a ₂ = (a ₂₁ , a ₂₂ ,..., A _2n ), a ₃ = ( a ₃₁ , a ₃₂ ,..., a _3n ). Let A be a matrix obtained by arranging these three eigenvectors. That is,

And By definition, the matrix Λ is a singular value matrix (eigenvalue matrix),

It is. further,

と書くことができる。つまり、行列Ａ’は回転、並進の自由度を除いて行列Ｘ’に相当する。実際には、行列Ａ’は行列Ｘ’に３次元の回転を作用させることによって得られることがある。このように、行列Ａ’は、回転の自由度を除き、ユーザのｎ個の部位の座標（ｘ_ｉ、ｙ_ｉ、ｚ_ｉ）を一意に表現する行列である。 Introducing the matrix A ′ by

Can be written. That is, the matrix A ′ corresponds to the matrix X ′ except for the degrees of freedom of rotation and translation. In practice, the matrix A ′ may be obtained by applying a three-dimensional rotation to the matrix X ′. Thus, the matrix A ′ is a matrix that uniquely expresses the coordinates (x _i , y _i , z _i ) of the n parts of the user, excluding the degree of freedom of rotation.

  Therefore, for example, quantifying how similar two subspaces obtained with respect to the coordinates of two n parts is two postures represented by the coordinates of the two n parts. Is synonymous with obtaining the degree of similarity that expresses how similar.

Some definitions of similarity have been proposed so far.
For example, the coordinates of two n parts are called observation data and teacher data. Then, it has been proposed that the observation data and the teacher data are added and the “rank 3 likelihood” of the obtained matrix is used as the similarity.

  Let P be a variance-covariance matrix obtained from observation data, and P ′ be a variance-covariance matrix obtained from teacher data. Both are added to obtain a matrix Q = P + P ′. The rank of this matrix Q is obtained. Both the matrix P and the matrix P ′ are rank 3 matrices. If there is no relationship between the matrix P and the matrix P ′, the rank of the matrix Q is 6. If the matrix P and the matrix P ′ are the same except for the degrees of freedom of rotation and expansion, the rank of the matrix Q is 3. Therefore, if it is possible to quantify how close the rank of the matrix R is to 3, it can be a similarity.

  However, in reality, because of the quantization error and the calculation error, if the rank of the matrix R is strictly determined, it becomes 6.

  In addition, it has been proposed to use the value of the eigenvalue (singular value) of the matrix Q to make it “rank 3 likeness”.

と定義する。 Also, let P be the variance-covariance matrix obtained from the observation data, and P ′ be the variance-covariance matrix obtained from the teacher data, and use the first canonical angles of the two subspaces A and A ′ obtained from them as the degree of similarity. It has also been proposed. The first canonical angle is defined as follows. Among the unit vectors on one partial space A, a vector v1 having the maximum projection length onto the other partial space A ′ is selected, and a vector obtained by projecting it onto the other partial space A ′ is defined as v2. The angle formed by the vector v1 and the vector v2 is the first canonical angle. The first canonical angle is an angle at which the two partial spaces A and A ′ are closest. In this case, assuming that the first canonical angle is θ, the similarity R is

It is defined as

  The calculation of the similarity R defined as described above involves an iterative calculation such as eigenvalue decomposition (singular value decomposition), and there is a problem that the processing time is long.

<Attitude detection device>
Hereinafter, an attitude detection device according to an embodiment will be described with reference to the drawings.

  The posture detection system 10 includes a sensor 21 that detects the coordinates of a plurality of parts of the user, a posture detection device 20, and an output device 30 such as a display and a printer.

In addition, teacher data 22 prepared in advance is input to the posture detection device 20.
The posture detection device 20 described below obtains a first rotational moment matrix that is a variance-covariance matrix from a matrix obtained from observation data representing the posture of an object, and a plurality of variance-covariance matrices for each of a plurality of teacher data. From the matrix obtained by using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix obtained by adding over the teacher data, and a plurality of teachers A second similarity between a plurality of teacher data obtained using a matrix element of a singular value matrix obtained by obtaining a first similarity between the data and the observed data and performing a singular value decomposition of the second rotational moment matrix; It is possible to determine whether or not the first posture belongs to the posture category represented by the plurality of teacher data by comparing the first similarity and the first similarity.

  The posture detection apparatus 20 includes a teacher data input unit 202, a teacher data feature amount calculation unit 204, a teacher data feature amount storage unit 206, a similarity threshold determination unit 208, a posture data input unit 210, a difference similarity calculation unit 212, and a similarity. A calculation unit 214, a similarity storage unit 216, a similar posture determination unit 218, and an output unit 210 are included.

  However, an apparatus including the teacher data input unit 202, the teacher data feature amount calculation unit 204, and the teacher data feature amount storage unit 206 that perform preprocessing may be provided as an external device. In this case, the posture detection apparatus 20 includes a similarity threshold determination unit 208, a posture data input unit 210, a difference similarity calculation unit 212, a similarity calculation unit 214, a similarity storage unit 216, a similar posture determination unit 218, and an output unit. 210.

  The teacher data feature amount calculation unit 204 calculates a teacher data feature amount, and the teacher data feature amount can be stored in the teacher data feature amount storage unit 206. The similarity calculation unit 214 calculates the similarity, and the similarity can be stored in the similarity storage unit 216.

  The output unit 220 of the posture detection device 20 outputs the determination result from the similar posture determination unit 218 to the output device 30.

  Teacher data 22 relating to a category with a posture is input to a teacher data input unit 202 of the posture detection device 20. The teacher data input unit 202, the teacher data feature amount calculation unit 204, and the similarity threshold determination unit 208 have the teacher data before the data regarding the coordinates of the plurality of parts of the user is input from the sensor 21 to the posture data input unit 210. Preferably 22 is processed.

  The teacher data 22 is input to the posture detection device 20 in order to generate data (similarity) to be referred to when determining the posture of the user from the coordinates of the plurality of parts of the user observed by the sensor 21. Is done.

The teacher data 22 may be coordinates x _pk (k = 1, 2, 3,...) Corresponding to a predetermined category p of posture. For example, the coordinate xpk may include ten coordinates (x _pi , y _pi , z _pi ) (i = 1 to 10). The predetermined category p of posture includes, for example, a posture in which the right hand is raised, a sitting posture, a standing posture, and the like.

となる。 The teacher data feature amount calculation unit 204 of the posture detection device 20 performs the following processing.
First, the teacher data feature amount calculation unit 204 of the posture detection device 20 includes coordinates xpk (k = 1, 2, 3,...) Included in the teacher data belonging to a predetermined category p of posture and a plurality of coordinates ( x _pi , y _pi , z _pi ) (i = 1 to n) is the center of gravity (x _pm , y _pm , z _pm ) and each coordinate (x _pi , y _pi , z _pi ) (i = 1 to n) A matrix Xp in which relative vectors from centroids (x _pm , y _pm , z _pm ) are arranged horizontally is defined. For example, if the number of coordinates (x _pi , y _pi , z _pi ) is 10, Xp is

It becomes.

で定義される。行列Ｘｐが１０個の座標（ｘ_ｐｉ、ｙ_ｐｉ、ｚ_ｐｉ）（ｉ＝１〜１０）を含むとき、行列Ｍｐは、１０×１０行列である。 Next, the teacher data feature amount calculation unit 204 of the posture detection apparatus 20 obtains a variance-covariance matrix Mp using the matrix Xp. The matrix Mp is sometimes called a rotational moment matrix. The matrix Mp is

Defined by When the matrix Xp includes 10 coordinates (x _pi , y _pi , z _pi ) (i = 1 to 10), the matrix Mp is a 10 × 10 matrix.

  In general, posture teacher data categorized into a predetermined category p of posture includes a plurality of samples. Therefore, an average of these samples is set as a rotational moment matrix Msum (p) of posture p.

ここで、｜＃ｓａｍｐｌｅ｜はサンプル数であり、和は全てのサンプルについて取るものとする。

Here, | #sample | is the number of samples, and the sum is taken for all samples.

のように分解し、左特異ベクトルＵｐと、右特異ベクトルＶｐ、特異値行列Ｓを得る。これら左特異ベクトルＵｐと、右特異ベクトルＶｐ、特異値行列Ｓの全てまたは一部は、教師データ特徴量２０６に含まれ得る。 The teacher data feature amount calculation unit 204 of the posture detection device 20 performs singular value decomposition on the rotational moment matrix Msum (p) of the predetermined category p of this posture. That is, the rotational moment matrix Msum (p) is

The left singular vector Up, the right singular vector Vp, and the singular value matrix S are obtained. All or part of the left singular vector Up, the right singular vector Vp, and the singular value matrix S may be included in the teacher data feature amount 206.

  When the rotational moment matrix Msum (p) for the predetermined category p of the posture is calculated from a plurality of samples as described above, the difference between samples for the predetermined category p of the posture, coordinate observation error, intermediate calculation and communication Errors such as quantization errors at the time of input may occur and be accumulated.

  Therefore, the similarity threshold determination unit 208 of the posture detection device 20 calculates the similarity Rp for the rotational moment matrix Mp of each teacher data of a plurality of samples.

と類似度Ｒｐを定義する。 The magnitude of the eigenvalue (singular value) is the contribution ratio of the eigenvector related to the observation data and the teacher data to the matrix Q before decomposition. Therefore, the ratio of the sum of the top three values in which eigenvalues (singular values) are arranged in descending order and the sum of the other eigenvalues may be “rank 3 likelihood”. That is, if the i-th eigenvalue is s _ii ,

And the similarity Rp is defined.

によって定義しても良い。この類似度Ｒｐは、姿勢検出の際、観測データから得られた類似度Ｒと比較される。たとえば類似度Ｒが、複数のサンプルの各教師データの回転モーメント行列Ｍｐの類似度Ｒｐの最小値以下の場合、観測データに対応する姿勢は、教師データに対応する姿勢のカテゴリに属すると判定しても良い。 The similarity Rp is defined as a singular value matrix S = {s _ij } (1 ≦ l <≦ k).

May be defined by The similarity Rp is compared with the similarity R obtained from the observation data at the time of posture detection. For example, when the similarity R is equal to or less than the minimum value of the similarity Rp of the rotational moment matrix Mp of each teacher data of a plurality of samples, it is determined that the attitude corresponding to the observation data belongs to the attitude category corresponding to the teacher data. May be.

  The sensor 21 observes the coordinates of a plurality of parts of the user. If the coordinates of 10 parts are observed, the sensor 21 obtains 10 three-dimensional coordinates (xi, yi, zi) (i = 1 to 10) as posture data. Then, the sensor 21 sends information related to the observation data to the attitude data input unit 210 of the attitude detection device 20. The attitude data is observation data (measurement data) obtained by the sensor 21.

  The posture data input unit 210 of the posture detection device 20 sends information related to posture data sent from the sensor 21 to the difference similarity calculation unit 212.

  The difference similarity calculation unit 212 performs a speed-up process by performing sequential processing when the posture data from the sensor 21 changes from moment to moment or when the difference between the posture data from the sensor 21 is a sparse matrix. . Therefore, if such processing is not performed, the difference similarity calculation unit 212 may simply send the posture data to the similarity calculation unit 214.

とする。 In the similarity calculation unit 214 of the posture detection device 20, the center of gravity of the ten coordinates (x _i , y _i , z _i ) (i = 1 to 10) is calculated as (x _m , y _m , z _m ) and matrix Y

And

で定義される。行列Ｙが１０個の座標（ｘ_ｉ、ｙ_ｉ、ｚ_ｉ）（ｉ＝１〜１０）を含むとき、行列Ｍは、１０×１０行列である。 Next, the similarity calculation unit 214 of the posture detection device 20 obtains a variance-covariance matrix M ′ using the matrix Y. The matrix M ′ may also be called a rotational moment matrix. The matrix M ′ is

Defined by When the matrix Y includes 10 coordinates (x _i , y _i , z _i ) (i = 1 to 10), the matrix M is a 10 × 10 matrix.

と定義される。これは、姿勢のカテゴリｐに対する左特異ベクトルＵｐと、右特異ベクトルＶｐを用いて計算された特異値行列である。もし、姿勢がｎ種類、つまり姿勢にはｎ個のカテゴリがあるとすると、特異値行列もｎ個計算される。上記の計算には、特異値分解のように、計算時間およびコストが掛かる反復計算は含まれていない。 The similarity calculation unit 214 of the posture detection apparatus 20 uses the left singular vector Up and the right singular vector Vp included in the teacher data feature amount corresponding to the posture category p, from the rotational moment matrix M ′ to the singular value matrix S. Calculate The singular value matrix Sp is

Is defined. This is a singular value matrix calculated using the left singular vector Up and the right singular vector Vp for the posture category p. If there are n types of postures, that is, there are n categories of postures, n singular value matrices are also calculated. The above calculation does not include an iterative calculation that requires calculation time and cost like singular value decomposition.

によって計算する。ここで｜ｓ_ｉｊ｜は、行列Ｓｐのｉ行ｊ列要素のノルムである。類似度計算部２１４は、カテゴリｐに対応する右特異ベクトルＵｐと左特異ベクトルＶｐを用いて特異値行列Ｓｐを得て、その特異値行列Ｓｐから類似度Ｒを求める。つまり、カテゴリｐの種類の数だけ類似度が求められる。これらの類似度のうち最大のものを類似度Ｒとして出力しても良い。類似度は類似度記憶部２１６に格納され得る。 In the similarity calculation unit 214, the similarity R of the teacher data with respect to the posture data and the posture category p is calculated from the elements of the matrix Sp described above.

Calculate by Here, | s _ij | is the norm of the i-th row and j-th column element of the matrix Sp. The similarity calculation unit 214 obtains the singular value matrix Sp using the right singular vector Up and the left singular vector Vp corresponding to the category p, and obtains the similarity R from the singular value matrix Sp. That is, the degree of similarity is obtained by the number of types of category p. The maximum of these similarities may be output as the similarity R. The similarity can be stored in the similarity storage unit 216.

  The similarity posture determination unit 218 compares the similarity R obtained from the observation data with the similarity Rp obtained by the similarity threshold determination unit 208. For example, when the similarity R obtained from the observation data is less than or equal to the minimum value of the similarity Rp of the rotational moment matrix Mp of each teacher data of a plurality of samples belonging to the category p, the attitude to the observation data is the attitude category p. You may determine that it belongs to.

The determination result in the similar posture determination unit 218 is sent to the output unit 220.
The output unit 220 sends the determination result to the output device 30.

  The output device 30 displays the determination result to the user or the like by a method such as displaying on a display or printing on paper.

The difference similarity calculation unit 212 increases the processing speed by performing sequential processing.
Here, a case where posture data changes every moment will be described.

となる。ここで、ｄＭ＝Ｍ（ｔ＋１）−Ｍ（ｔ）である。このように差分類似度算出部２１２は、類似度記憶部２１６に格納されている前回観測された姿勢データとの差分のみを計算することによって、処理の逐次化を行う。そして、処理を逐次化することによって、観測ごとの差分が小さい（たとえば、ｄＭが疎行列である）場合、逐次化を行うことによって、姿勢データが時々刻々と変化する場合に、計算の高速化を図ることができる。 Let M (t) be the rotational moment matrix obtained from the attitude data at the t-th observation time. Then

It becomes. Here, dM = M (t + 1) −M (t). As described above, the difference similarity calculation unit 212 performs the process serialization by calculating only the difference from the previously observed posture data stored in the similarity storage unit 216. Then, by serializing the processing, if the difference for each observation is small (for example, dM is a sparse matrix), the serialization speeds up the calculation when the posture data changes from moment to moment Can be achieved.

  As described above, the posture detection device 20 determines whether or not the posture of the object represented by the observation data including a plurality of coordinates belongs to the category of posture represented by a plurality of teacher data each including a plurality of coordinates. A variance-covariance matrix is obtained for each of the plurality of teacher data, and the second rotational moment matrix obtained by adding the variance-covariance matrix over the plurality of teacher data is subjected to singular value decomposition to obtain a left singular vector and a right singular vector. A vector and a singular value matrix are obtained, and a second similarity representing the similarity between the plurality of teacher data obtained by using a matrix element of the singular value matrix is represented as a similarity between the plurality of teacher data and the observation data. The determination is made by comparing with the first similarity.

For example, the teacher data feature amount calculation unit 204 of the posture detection device 20 obtains a variance covariance matrix M = X ′ ^T X ′ for each of a plurality of teacher data belonging to the category p with posture, and the variance covariance matrix A singular value decomposition Mn = USV ^T is applied to the second rotational moment matrix Mn obtained by adding M = X ′ ^T X ′ over a plurality of teacher data, and the left singular vector V, the right singular vector U, and the singular value matrix S = {s _ij } is obtained, and a second similarity Rp between a plurality of teacher data is obtained using the matrix elements of the singular value matrix S.

から複数の教師データと観測データの類似を表す第１の類似度を求める。 Similarity calculation unit 214 of the posture detection device 20 'first rotational moment matrix M is a covariance matrix from' matrix Y obtained from the observed data = seeking Y ^{'T Y',} the first rotational moment matrix Matrix obtained using M ′ = ^X′T X ′, left singular vector V, right singular vector U

A first similarity representing the similarity between the plurality of teacher data and the observation data is obtained.

  The similar posture determination unit 218 of the posture detection device 20 determines whether the posture of the object belongs to the category defined by the teacher data by comparing the first similarity and the second similarity.

  For example, the similar posture determination unit 218 uses the second similarity obtained from the teacher data for the posture category p and the left singular vector and the right singular vector obtained from the teacher data for the posture category p as posture data (observation). The first similarity calculated from the matrix obtained by acting on the variance-covariance matrix (rotation moment matrix) obtained from the data) is compared with the second similarity obtained from the teacher data for the posture category p. . If the first similarity is smaller than the second similarity obtained from the teacher data for the posture category p, the posture corresponding to the posture data may be determined to belong to the posture category p.

The posture detection apparatus 20 includes a difference similarity calculation unit 212.
When the observation data includes the first observation data and the second observation data, the first rotation moment matrix obtained from the first observation data is the first rotation matrix obtained from the first observation matrix and the second observation data. If the moment matrix is the second observation matrix, the difference similarity calculation unit 212 calculates the difference between the second observation matrix and the first observation matrix. Also, the similarity calculation unit 214 can obtain the first similarity from a matrix obtained using the difference, the left singular vector, and the right singular vector.

  In this way, the posture detection device 20 obtains the first rotational moment matrix, which is a variance-covariance matrix, from the matrix obtained from the observation data representing the posture of the object in the similarity calculation unit 214, and applies to each of the plurality of teacher data. Obtained by using the left singular vector and right singular vector obtained by singular value decomposition of the second rotational moment matrix obtained by adding the variance-covariance matrix over a plurality of teacher data, and the first rotational moment matrix A first similarity between a plurality of teacher data and observation data is obtained from the matrix, and the similarity posture determination unit 218 obtains matrix elements of the singular value matrix obtained by singular value decomposition of the second rotational moment matrix. By comparing the second similarity between the plurality of teacher data obtained by using the first similarity, the posture is represented by the plurality of teacher data. It determines whether or not belong to Li.

  The posture detection device 20 does not include the iterative calculation that requires calculation time and cost unlike the singular value decomposition when comparing the observation data and the teacher data. Therefore, the category indicated by the observation data is indicated by the teacher data. Whether it is the same as the category can be determined at high speed.

FIG. 2 is a diagram illustrating an example of the configuration of the posture detection apparatus according to the embodiment.
The computer 100 includes a central processing unit (CPU) 102, a read only memory (ROM) 104, and a random access memory (RAM) 106. The computer 500 further includes a hard disk device 108, an input device 110, a display device 112, an interface device 114, and a recording medium driving device 116. Note that these components are connected via a bus line 120, and various data can be exchanged under the control of the CPU.

  A central processing unit (CPU) 102 is an arithmetic processing unit that controls the operation of the entire computer 100, and functions as a control processing unit of the computer 100.

  The Read Only Memory (ROM) 104 is a read-only semiconductor memory in which a predetermined basic control program is recorded in advance. The CPU 102 can control the operation of each component of the computer 100 by reading out and executing the basic control program when the computer 100 is started up.

  A random access memory (RAM) 106 is a semiconductor memory that can be written and read at any time and used as a working storage area as necessary when the CPU 102 executes various control programs.

  The hard disk device 108 is a storage device that stores various control programs executed by the CPU 102 and various data. The MPU 502 can perform various control processes, which will be described later, by reading and executing a predetermined control program stored in the hard disk device 108.

  The input device 110 is, for example, a mouse device or a keyboard device. When operated by a user of the information processing device, the input device 110 acquires input of various information associated with the operation content and sends the acquired input information to the CPU 102. To do.

  The display device 512 is, for example, a liquid crystal display, and displays various texts and images according to display data sent from the MPU 502.

  The interface device 114 manages the exchange of various information with various devices connected to the computer 100.

  The recording medium driving device 116 is a device that reads various control programs and data recorded on the portable recording medium 118. The CPU 102 can read out and execute a predetermined control program recorded on the portable recording medium 118 via the recording medium driving device 116 to perform various control processes described later. As the portable recording medium 118, for example, a flash memory equipped with a USB (Universal Serial Bus) standard connector, a CD-ROM (Compact Disc Read Only Memory), a DVD-ROM (Digital Versatile Disc Read Only Memory). and so on.

  In order to configure the posture detection apparatus using such a computer 100, for example, a control program for causing the CPU 102 to perform the processing in each processing unit described above is created. The created control program is stored in advance in the hard disk device 108 or the portable recording medium 118. Then, a predetermined instruction is given to the CPU 102 to read and execute the control program. By doing so, the CPU 102 provides the functions of the posture detection device.

<Attitude detection processing>
Processing in the posture detection apparatus 10 will be described with reference to FIGS.

  When the posture detection device 10 is the general-purpose computer 100 as shown in FIG. 2, the following description defines a control program for performing such processing. That is, hereinafter, it is also a description of a control program that causes a general-purpose computer to perform the processing described below.

  FIG. 3 is a diagram illustrating an example of the flow of pre-processing. The pre-processing is preferably performed before data relating to the coordinates of a plurality of parts of the user is input from the sensor 21 to the posture data input unit 210.

のように定義する。 When the processing is started, in S100, the teacher data feature amount calculation unit 204 of the posture detection device 20 sets the coordinates of the user's part corresponding to the predetermined category p of the posture to x _pk (k = 1, 2, 3,...・). For example, there are 10 pieces of teacher data corresponding to a predetermined category p of posture, and the center of gravity (x _pm , y _pm , z _pm ) of the coordinates (x _pi , y _pi , z _pi ) (i = 1 to 10) is obtained. Then, the teacher data feature amount calculation unit 204 converts the matrix Xp into

Define as follows.

で定義される。たとえば行列Ｘｐが１０個の座標（ｘ_ｐｉ、ｙ_ｐｉ、ｚ_ｐｉ）（ｉ＝１〜１０）を含むとき、行列Ｍｐは、１０×１０行列である。 In S100, the teacher data feature amount calculation unit 204 of the posture detection device 20 obtains the rotational moment matrix Mp using the matrix Xp. The matrix Mp is

Defined by For example, when the matrix Xp includes 10 coordinates (x _pi , y _pi , z _pi ) (i = 1 to 10), the matrix Mp is a 10 × 10 matrix.

  In general, the learning data of the posture that is categorized with the category p of the predetermined posture includes a plurality of samples. Therefore, the average of these samples is set as the rotational moment matrix Msum (p) of the posture category p.

ここで、｜＃ｓａｍｐｌｅ｜はサンプル数であり、和は全てのサンプルについて取るものとする。

Here, | #sample | is the number of samples, and the sum is taken for all samples.

のように左特異ベクトルＵｐと、右特異ベクトルＶｐ、特異値行列Ｓを用いて分解する。Ｓ１０２で姿勢検出装置２０の教師データ特徴量算出部２０４は、算出された左特異ベクトルＵｐと、右特異ベクトルＶｐ、特異値行列Ｓを教師データ特徴量記憶部２０６に格納しても良い。またＳ１０２で姿勢検出装置２０の教師データ特徴量算出部２０４は、算出された左特異ベクトルＵｐと、右特異ベクトルＶｐのみを教師データ特徴量記憶部２０６に格納しても良い。 In next step S102, the teacher data feature amount calculation unit 204 of the posture detection apparatus 20 decomposes the rotational moment matrix Msum (p) of the posture category p in a singular value manner. That is, the rotational moment matrix Msum (p)

As shown, the left singular vector Up, the right singular vector Vp, and the singular value matrix S are used for decomposition. In S102, the teacher data feature amount calculation unit 204 of the posture detection device 20 may store the calculated left singular vector Up, right singular vector Vp, and singular value matrix S in the teacher data feature amount storage unit 206. In S102, the teacher data feature amount calculation unit 204 of the posture detection apparatus 20 may store only the calculated left singular vector Up and right singular vector Vp in the teacher data feature amount storage unit 206.

  In next step S104, the similarity threshold determination unit 208 of the posture detection apparatus 20 calculates the similarity Rp for the posture category p from the rotational moment matrix Mp of each teacher data of a plurality of samples.

  In S104, the similarity Rp for the calculated posture category p may be stored in the teacher data feature storage unit 206.

The above is an example of the flow of pre-processing.
FIG. 4 is a diagram illustrating an example of a processing flow of the posture detection method of the embodiment.

  The processing shown in FIG. 4 relates to processing when the user's posture changes from moment to moment, but the processing when the user's posture does not change may be omitted sequentially. Further, it is assumed that the pre-process shown in FIG. 3 has been completed as a premise of the process shown in FIG.

とする。 In S200, the similarity calculation unit 214 of the posture detection apparatus 20 determines the center of gravity of the ten coordinates (x _i , y _i , z _i ) (i = 1 to 10) for the posture data as well as the teacher data (x _m , y _m , z _m ) and matrix Y

And

で定義される。行列Ｘが１０個の座標（ｘ_ｉ、ｙ_ｉ、ｚ_ｉ）（ｉ＝１〜１０）を含むとき、行列Ｍは、１０×１０行列である。 In S200, the similarity calculation unit 214 of the posture detection device 20 obtains the rotational moment matrix M using the matrix Y. The matrix M is

Defined by When the matrix X includes 10 coordinates (x _i , y _i , z _i ) (i = 1 to 10), the matrix M is a 10 × 10 matrix.

と定義される。これは、姿勢のカテゴリｐに対する左特異ベクトルＵｐと、右特異ベクトルＶｐを用いて計算された特異値行列である。もし、姿勢がｎ種類あるとすると、特異値行列もｎ個計算される。上記の計算には、特異値分解のように、計算時間およびコストが掛かる反復計算は含まれていない。 In step S202, the similarity calculation unit 214 of the posture detection device 20 uses the left singular vector Up and the right singular vector Vp included in the teacher data feature amount for the posture category p, and uses the singular value from the rotational moment matrix M ′. The matrix S is calculated. The singular value matrix Sp is

Is defined. This is a singular value matrix calculated using the left singular vector Up and the right singular vector Vp for the posture category p. If there are n types of postures, n singular value matrices are also calculated. The above calculation does not include an iterative calculation that requires calculation time and cost like singular value decomposition.

によって計算する。ここで｜ｓ_ｉｊ｜は、行列Ｓｐのｉ行ｊ列要素のノルムである。類似度Ｒは、姿勢のカテゴリがｋ種類あるとすると、ｋ個算出され得る。類似度計算部２１４は、このようなｋ個の類似度Ｒの最小のものを類似度として出力しても良い。得られた類似度は類似度記憶部２１６に格納され得る。 In S202, the similarity calculation unit 214 calculates the similarity R between the posture data (observation data) and the teacher data corresponding to the posture category p from the elements of the matrix Sp described above.

Calculate by Here, | s _ij | is the norm of the i-th row and j-th column element of the matrix Sp. If there are k types of posture categories, k similarities R can be calculated. The similarity calculation unit 214 may output the smallest of the k similarities R as the similarity. The obtained similarity degree can be stored in the similarity degree storage unit 216.

  In S204, the difference similarity calculation unit 212 calculates dM = M (t + 1) −M (t), where M (t) is a rotational moment matrix obtained from the posture data at the t-th observation time.

を計算する。 In the next S206, the difference similarity calculation unit 212

Calculate

In S206, the similarity calculation unit 214 calculates the similarity R between the posture data (observation data) and the teacher data corresponding to the posture category p from the elements of the matrix Sp (t) + Up ^T dMVp.

  In S206, the similar posture determination unit 218 compares the similarity R with the similarity Rp obtained by the similarity threshold determination unit 208 in S104, and the posture represented by the posture data corresponds to which category of posture. Determine whether to do. For example, when the similarity is equal to or smaller than the minimum value of the similarity Rp of the rotational moment matrix Mp of each teacher data of a plurality of samples, it may be determined that they belong to the same posture category. When the process of this step is completed, the process proceeds to S208.

  In S208, the similar posture determination unit 218 determines whether or not to end the posture detection process. This determination may be made based on, for example, whether or not the number of observations has reached a predetermined value, that is, whether or not a predetermined time has elapsed since the start of measurement. Alternatively, the determination may be made based on whether dM = M (t + 1) −M (t) is smaller than a predetermined value. If the result of this determination is “Yes”, that is, if it is determined that the posture detection process is to be terminated, the process is terminated. If the result of this determination is “No”, that is, if it is determined not to end the posture detection process, the process returns to S204.

  By performing such processing, it is possible to quickly determine whether or not the categories indicated by the data are the same by comparing the observation data and the teacher data.

<Example>
The posture detection apparatus and the posture detection method as described above can also be applied to user behavior prediction.

  Systems that interact with users are known. In such a system, there is a desire to improve usability by eliminating extra user effort. Recommendations for products at shopping sites on the Internet can also be included in this example. In the shopping site on the Internet, we recommend what the user wants from the purchase history, but here there are many functions by predicting what to do next from the past user behavior pattern The access procedure can be simplified (for example, a shortcut to the predicted function can be dynamically presented).

  In the following, an example in which a user's state, for example, what he is trying to do is estimated from input information (also called an event) obtained from various sensors anywhere.

  An “event” may refer to information about a user (or its peripheral device) obtained from a sensor. Examples include human face detection, gaze and front face detection, gaze and gesture detection, posture detection (such as a pose), touch sensor reaction, and voice reaction. Many of these events are accumulated in time series, and the importance of information and the meaning (recognition by human beings) are often not classified.

  The “user action” may refer to event information (accessed information, logged out, made a request, etc.) that is a goal of interaction, not an action actually performed by the user. An event in which information from any sensor or interaction target device is input may be referred to as “user behavior”.

  The data to be processed as the behavior pattern indicates time series information of the event having the user's behavior as the last element. For example, consider a system in which there are A1, A2, and A3 as user actions, and B1, B2, B3, and B4 exist as other events. Examples of behavior patterns include “B1-B3-B4-B2-B3-A1”, “B1-B2-B3-B2-B4-A2”, “B1-B2-B3-B2-B3-A3”, “ B1-B3-B2-B3-A3 "etc. can be considered. In the above example, only one user action is included in the action pattern, but this is not the case. Further, since it is not necessary to record a certain period of human behavior, it is assumed that the behavior pattern itself takes a variable length.

  A recurrence plot is an example of stylizing a variable-length behavior pattern. This is a visualization technique used for statistical analysis of chaotic time series and plots on the uv plane, defining the variable u as the current state and the variable v as the previous state.

  Here, using the recurrence plot technique, the action pattern is plotted on a two-dimensional plane having 7 states of A1, A2, A3, B1, B2, B3, and B4 on both the u and v axes. Even if it exists, it can be converted into 7 × 7 matrix data. When one standardization is performed as an example, the behavior pattern “B1-B2-B3-B4-B2-A2” becomes a matrix X as shown in FIG.

  In order to determine which pattern a user's action is classified into, it is necessary to categorize the action pattern beforehand based on the teacher data. For example, even when behavior patterns ending with A2 are collected as described above, whimsical behavior is mixed in human behavior, and the same data is not necessarily obtained. For example, when 100 pieces of data are taken, the total of X is assumed to be Xtotal shown in FIG.

  The behavior pattern of the user at a certain time and the description format of the teacher data (the size of the matrix, the order of the elements) are the same, and the behavior pattern is analyzed by a method essentially the same as the “calculation of matrix similarity” be able to.

  In the above, human behavior prediction has been described, but any object other than a human may be used.

である付記１または２に記載の姿勢検出装置。
（付記４）
前記第１の類似度は、前記第１の回転モーメント行列をＭ’、前記左特異ベクトルをＶ、右特異ベクトルをＵとして、

で表される行列Ｔの要素を用いて定義される付記１乃至３のいずれか一項に記載の姿勢検出装置。
（付記５）
前記第１の類似度Ｒは、前記行列Ｔの要素ｔ_ｉｊ（１≦ｌ＜≦ｋ）を用いて、

である付記４に記載の姿勢検出装置。
（付記６）
コンピュータによって実行される姿勢検出方法であって、
物体の姿勢を表す観測データから得られる行列から分散共分散行列である第１の回転モーメント行列を求め、複数の教師データの各々に対する分散共分散行列を前記複数の教師データにわたって加算して得られる第２の回転モーメント行列を特異値分解して得られる左特異ベクトルと右特異ベクトルと、前記第１の回転モーメント行列とを用いて得られる行列から、前記複数の教師データと前記観測データとの間の第１の類似度を求めるとと、
前記第２の回転モーメント行列を特異値分解して得られる特異値行列の行列要素を用いて得られる、前記複数の教師データ間の第２の類似度と、前記第１の類似度とを比較することによって、前記姿勢が前記複数の教師データで表される姿勢のカテゴリに属するかどうかを判定することと
を含む姿勢検出方法。
（付記７）
前記観測データが第１の観測データと第２の観測データを含むとき、前記第１の観測データから得られる前記第１の回転モーメント行列を第１観測行列、前記第２の観測データから得られる前記第１の回転モーメント行列を第２観測行列とすると、
さらに、
前記第２観測行列と前記第１観測行列との差分を計算することと、
前記差分、前記左特異ベクトル、前記右特異ベクトルを用いて得られる行列から第１の類似度を求めることと
を含む、付記６に記載の姿勢検出方法。
（付記８）
前記第２の類似度Ｒｐは、前記特異値行列をＳ＝｛ｓ_ｉｊ｝（１≦ｌ＜≦ｋ）として、

である付記６または７に記載の姿勢検出方法。
（付記９）
前記第１の類似度は、前記第１の回転モーメント行列をＭ’、前記左特異ベクトルをＶ、右特異ベクトルをＵとして、

で表される行列Ｔの要素を用いて定義される付記６乃至８のいずれか一項に記載の姿勢検出方法。
（付記１０）
前記第１の類似度Ｒは、前記行列Ｔの要素ｔ_ｉｊ（１≦ｌ＜≦ｋ）を用いて、

である付記９に記載の姿勢検出方法。
（付記１１）
物体の姿勢を表す観測データから得られる行列から分散共分散行列である第１の回転モーメント行列を求め、複数の教師データの各々に対する分散共分散行列を前記複数の教師データにわたって加算して得られる第２の回転モーメント行列を特異値分解して得られる左特異ベクトルと右特異ベクトルと、前記第１の回転モーメント行列とを用いて得られる行列から、前記複数の教師データと前記観測データとの間の第１の類似度を求め、
前記第２の回転モーメント行列を特異値分解して得られる特異値行列の行列要素を用いて得られる、前記複数の教師データ間の第２の類似度と、前記第１の類似度とを比較することによって、前記姿勢が前記複数の教師データで表される姿勢のカテゴリに属するかどうかを判定する
処理をコンピュータに実行させることを特徴とする姿勢検出プログラム。
（付記１２）
前記観測データが第１の観測データと第２の観測データを含むとき、前記第１の観測データから得られる前記第１の回転モーメント行列を第１観測行列、前記第２の観測データから得られる前記第１の回転モーメント行列を第２観測行列とすると、
さらに、
前記第２観測行列と前記第１観測行列との差分を計算することと、
前記差分、前記左特異ベクトル、前記右特異ベクトルを用いて得られる行列から第１の類似度を求める
処理をコンピュータに実行させることを特徴とする付記１１に記載の姿勢検出プログラム。
（付記１３）
前記第２の類似度Ｒｐは、前記特異値行列をＳ＝｛ｓ_ｉｊ｝（１≦ｌ＜≦ｋ）として、

である付記１１または１２に記載の姿勢検出プログラム。
（付記１４）
前記第１の類似度は、前記第１の回転モーメント行列をＭ’、前記左特異ベクトルをＶ、右特異ベクトルをＵとして、

で表される行列Ｔの要素を用いて定義される付記１１乃至１３のいずれか一項に記載の姿勢検出プログラム。
（付記１５）
前記第１の類似度Ｒは、前記行列Ｔの要素ｔ_ｉｊ（１≦ｌ＜≦ｋ）を用いて、

である付記１４に記載の姿勢検出プログラム。 Regarding the above embodiment and examples, the following additional notes are disclosed.
(Appendix 1)
A first rotational moment matrix that is a variance-covariance matrix is obtained from a matrix obtained from observation data representing the posture of an object, and obtained by adding a variance-covariance matrix for each of a plurality of teacher data over the plurality of teacher data. From the matrix obtained by using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix and the first rotational moment matrix, the plurality of teacher data and the observation data A similarity calculation unit for obtaining a first similarity between,
A second similarity between the plurality of teacher data obtained by using a matrix element of a singular value matrix obtained by singular value decomposition of the second rotational moment matrix is compared with the first similarity. And a similar posture determination unit that determines whether or not the posture belongs to a posture category represented by the plurality of teacher data.
(Appendix 2)
When the observation data includes first observation data and second observation data, the first rotational moment matrix obtained from the first observation data is obtained from the first observation matrix and the second observation data. When the first rotational moment matrix is a second observation matrix,
further,
A difference similarity calculation unit for calculating a difference between the second observation matrix and the first observation matrix;
The posture detection apparatus according to appendix 1, wherein the similarity calculation unit calculates the first similarity from a matrix obtained using the difference, the left singular vector, and the right singular vector.
(Appendix 3)
The second similarity Rp is obtained by setting the singular value matrix as S = {s _ij } (1 ≦ l <≦ k).

The attitude detection apparatus according to appendix 1 or 2, wherein
(Appendix 4)
The first similarity is represented by M ′ as the first rotational moment matrix, V as the left singular vector, and U as the right singular vector.

4. The posture detection apparatus according to any one of supplementary notes 1 to 3, which is defined using an element of a matrix T represented by:
(Appendix 5)
The first similarity R is calculated using an element t _ij (1 ≦ l ≦≦ k) of the matrix T,

The attitude detection apparatus according to appendix 4, wherein
(Appendix 6)
A posture detection method executed by a computer,
A first rotational moment matrix that is a variance-covariance matrix is obtained from a matrix obtained from observation data representing the posture of an object, and obtained by adding a variance-covariance matrix for each of a plurality of teacher data over the plurality of teacher data. From the matrix obtained by using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix and the first rotational moment matrix, the plurality of teacher data and the observation data When the first similarity between them is obtained,
A second similarity between the plurality of teacher data obtained by using a matrix element of a singular value matrix obtained by singular value decomposition of the second rotational moment matrix is compared with the first similarity. And determining whether or not the posture belongs to a posture category represented by the plurality of teacher data.
(Appendix 7)
When the observation data includes first observation data and second observation data, the first rotational moment matrix obtained from the first observation data is obtained from the first observation matrix and the second observation data. When the first rotational moment matrix is a second observation matrix,
further,
Calculating a difference between the second observation matrix and the first observation matrix;
The posture detection method according to claim 6, further comprising: determining a first similarity from a matrix obtained using the difference, the left singular vector, and the right singular vector.
(Appendix 8)
The second similarity Rp is obtained by setting the singular value matrix as S = {s _ij } (1 ≦ l <≦ k).

The attitude detection method according to appendix 6 or 7, wherein
(Appendix 9)
The first similarity is represented by M ′ as the first rotational moment matrix, V as the left singular vector, and U as the right singular vector.

The posture detection method according to any one of appendices 6 to 8, which is defined using an element of a matrix T represented by:
(Appendix 10)
The first similarity R is calculated using an element t _ij (1 ≦ l ≦≦ k) of the matrix T,

The attitude detection method according to appendix 9, wherein
(Appendix 11)
A first rotational moment matrix that is a variance-covariance matrix is obtained from a matrix obtained from observation data representing the posture of an object, and obtained by adding a variance-covariance matrix for each of a plurality of teacher data over the plurality of teacher data. From the matrix obtained by using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix and the first rotational moment matrix, the plurality of teacher data and the observation data Find the first similarity between
A second similarity between the plurality of teacher data obtained by using a matrix element of a singular value matrix obtained by singular value decomposition of the second rotational moment matrix is compared with the first similarity. By doing so, the computer executes a process of determining whether or not the posture belongs to the posture category represented by the plurality of teacher data.
(Appendix 12)
When the observation data includes first observation data and second observation data, the first rotational moment matrix obtained from the first observation data is obtained from the first observation matrix and the second observation data. When the first rotational moment matrix is a second observation matrix,
further,
Calculating a difference between the second observation matrix and the first observation matrix;
The attitude detection program according to appendix 11, wherein the computer executes a process of obtaining a first similarity from a matrix obtained using the difference, the left singular vector, and the right singular vector.
(Appendix 13)
The second similarity Rp is obtained by setting the singular value matrix as S = {s _ij } (1 ≦ l <≦ k).

The attitude detection program according to appendix 11 or 12,
(Appendix 14)
The first similarity is represented by M ′ as the first rotational moment matrix, V as the left singular vector, and U as the right singular vector.

The attitude detection program according to any one of supplementary notes 11 to 13, which is defined using an element of a matrix T represented by:
(Appendix 15)
The first similarity R is calculated using an element t _ij (1 ≦ l ≦≦ k) of the matrix T,

The attitude detection program according to supplementary note 14, wherein

DESCRIPTION OF SYMBOLS 20 Posture detection apparatus 21 Sensor 22 Teacher data 202 Teacher data input part 204 Teacher data feature-value calculation part 206 Teacher-data feature-value storage part 208 Similarity threshold determination part 210 Posture data input part 212 Difference similarity calculation part 214 Similarity degree calculation part 216 Similarity storage unit 218 Similar posture determination unit 220 Output unit 30 Output device

Claims (7)

A first rotational moment matrix that is a variance-covariance matrix is obtained from a matrix obtained from observation data representing the posture of an object, and obtained by adding a variance-covariance matrix for each of a plurality of teacher data over the plurality of teacher data. From the matrix obtained by using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix and the first rotational moment matrix, the plurality of teacher data and the observation data A similarity calculation unit for obtaining a first similarity between,
A second similarity between the plurality of teacher data obtained by using a matrix element of a singular value matrix obtained by singular value decomposition of the second rotational moment matrix is compared with the first similarity. And a similar posture determination unit that determines whether or not the posture belongs to a posture category represented by the plurality of teacher data. When the observation data includes first observation data and second observation data, the first rotational moment matrix obtained from the first observation data is obtained from the first observation matrix and the second observation data. When the first rotational moment matrix is the second observation matrix,
A difference similarity calculation unit for calculating a difference between the second observation matrix and the first observation matrix;
The posture detection apparatus according to claim 1, wherein the similarity calculation unit calculates a second similarity from a matrix obtained using the difference, the left singular vector, and the right singular vector. The second similarity Rp is obtained by setting the singular value matrix as S = {s _ij } (1 ≦ l <≦ k).

The posture detection apparatus according to claim 1 or 2. The first similarity is represented by M ′ as the first rotational moment matrix, V as the left singular vector, and U as the right singular vector.

The attitude detection apparatus according to claim 1, wherein the attitude detection apparatus is defined using an element of a matrix T represented by: The first similarity R is calculated using an element t _ij (1 ≦ l ≦≦ k) of the matrix T,

The posture detection apparatus according to claim 4. A posture detection method executed by a computer,
A first rotational moment matrix that is a variance-covariance matrix is obtained from a matrix obtained from observation data representing the posture of an object, and obtained by adding a variance-covariance matrix for each of a plurality of teacher data over the plurality of teacher data. From the matrix obtained by using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix and the first rotational moment matrix, the plurality of teacher data and the observation data When the first similarity between them is obtained,
A second similarity between the plurality of teacher data obtained by using a matrix element of a singular value matrix obtained by singular value decomposition of the second rotational moment matrix is compared with the first similarity. And determining whether or not the posture belongs to a posture category represented by the plurality of teacher data. A first rotational moment matrix that is a variance-covariance matrix is obtained from a matrix obtained from observation data representing the posture of an object, and obtained by adding a variance-covariance matrix for each of a plurality of teacher data over the plurality of teacher data. From the matrix obtained by using the left singular vector and the right singular vector obtained by singular value decomposition of the second rotational moment matrix and the first rotational moment matrix, the plurality of teacher data and the observation data Find the first similarity between
A second similarity between the plurality of teacher data obtained by using a matrix element of a singular value matrix obtained by singular value decomposition of the second rotational moment matrix is compared with the first similarity. By doing so, the computer executes a process of determining whether or not the posture belongs to the posture category represented by the plurality of teacher data.

Images