JP5936038B2 - Action main identification method, action main identification system, action main identification program, and information storage medium - Google Patents

Description

  Some embodiments of the present invention relate to an operation main identification method, an operation main identification system, an operation main identification program, and an information storage medium.

  Human identification is important for many fields and is widely applied as shown in Non-Patent Document 1. In recent years, systems based on voice recognition and face recognition have been developed as human identification, but these alone are still insufficient to always recognize a person accurately. Therefore, as shown in Non-Patent Documents 2 and 4, human walking has also been used. As shown in Non-Patent Documents 5 to 7, information mainly used for authentication from walking is how to swing an arm, the length of one step, and the like.

A. Jain, J. Bolle, and S. Pankanti. Biometrics-Personal Identification in Networked Society. Springer US, 2002. Biometrics and Identity Management, volume 5372 of Lecture Notes in Computer Science.Springer Berlin / Heidelberg, 2008. M. Nixon. Gait biometrics. Biometric Technology Today, 16 (7-8): 8.9, 2008. J. Boyd and J. Little. Advanced Studies in Biometrics. Lecture Notes in Computer Science. Springer Berlin / Heidelberg, 2005. D. Jokisch, I. Daum, and N. Troje. Self recognition versus recognition of others by biological motion: Viewpoint-depedent effects. Perception, 35: 911.920, 2006. F. Loula, S. Prasad, K. Harber, and M. Shiffrar.Recognizing people from their movement.J. Of Experimental Psychology: Human Perception and Performance, 31 (1): 210.220, 2005. J. Cutting and L. Kozlowski.Recognizing friends by their walk: Gait perception without familiarity cues.Bulletin of Psychonomic Society, 9 (5): 353.356, 1977.

However, it is difficult to measure the length of one step and how to swing an arm when walking, and there is a problem that it changes, for example, when carrying luggage.
Also, when recognizing humans, instead of how to swing the arm and the length of one step, it is thought that the recognition accuracy will be improved by comprehensively analyzing the speed of the whole human, balance of the waist, posture of the torso, etc. However, when the amount of data used for analysis increases, there is a problem that the calculation load such as data processing increases and it is difficult to make a determination in real time.

  The present invention has been made in view of the above technical problems. According to some aspects of the present invention, it is possible to provide an operation main identification method, an operation main identification system, an operation main identification program, and an information storage medium capable of accurately identifying an operation main with a small amount of data and a calculation load. .

(1) The present invention
A method for identifying an actor,
As sample motion data, a sample motion data acquisition step for acquiring motion data for a given operation of a plurality of actors;
A first data conversion process is performed on the sampled motion data to generate a feature vector, and a second data conversion process is performed on the feature vector to generate a feature data of dimension m lower than the feature vector. And a database creation step of storing in the storage unit in association with the main actor as sample data,
A cluster information calculation step for clustering sample data for each operation main body to obtain cluster information including at least one of feature points and feature functions of the cluster in the m-dimensional space;
As a target motion data, a target motion data acquisition step for acquiring motion data of an operation subject to be identified;
A first data conversion process is performed on the target motion data to generate a feature vector, and a second data conversion process is performed on the feature vector to generate feature data of dimension m lower than the feature vector. The operation main identification step of calculating the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster and identifying the operation main based on the feature values And an operation main identification method.

The present invention also provides
An operator identification system for identifying an operator,
A first data conversion process is performed on the sampled motion data, a feature vector is generated, a second data conversion process is performed on the feature vector, and feature data of dimension m lower than the feature vector A database creation unit that generates and stores the sample data in the storage unit in association with the actor,
A cluster information calculation unit that clusters sample data for each actor and obtains cluster information including at least one of feature points and feature functions of the cluster in the m-dimensional space, and first data for the target motion data Performing a transformation process, generating a feature vector, performing a second data transformation process on the feature vector, generating feature data of a dimension m lower in dimension than the feature vector, and using it as target data; The present invention relates to an operation main identification system including an operation main identification unit that calculates a feature value of target data based on target data and at least one of feature points and feature functions of each cluster and identifies an operation main based on the feature value.

  The operation subject identification system may include a sample motion data acquisition unit that acquires motion data for a given operation of a plurality of operation subjects as sample motion data.

  The operating subject identification system may include a target motion data acquisition unit that acquires the motion data of the operating subject to be identified as target motion data.

  The present invention also relates to a program that causes a computer to function as each of the above-described units. The present invention also relates to an information storage medium storing a program that causes a computer to function as each of the above-described units.

The present invention also provides
An operator identification system for identifying an operator,
Feature data of a dimension m lower than the feature vector generated by performing the second data conversion processing on the feature vector generated by performing the first data conversion processing on the sample motion data, A sample data storage unit stored as sample data in association with the main actor,
A cluster information calculation unit for clustering sample data for each operation main body to obtain cluster information including at least one of feature points and feature functions of a cluster in an m-dimensional space;
A first data conversion process is performed on the target motion data, a feature vector is generated, a second data conversion process is performed on the feature vector, and feature data having a dimension m lower than the feature vector is obtained. Generated as target data, calculates the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster, and identifies the actor main based on the feature value And an operation main identification system.

The operation subject identification system may include a sample motion data acquisition unit that acquires motion data for a given operation of a plurality of operation subjects as sample motion data.
The operating subject identification system may include a target motion data acquisition unit that acquires the motion data of the operating subject to be identified as target motion data.

  The present invention also relates to a program that causes a computer to function as each of the above-described units. The present invention also relates to an information storage medium storing a program that causes a computer to function as each of the above-described units.

  The main actor may be a human or an animal, for example. A robot may also be used.

  The motion data may be data representing the movement of a predetermined part of the main operator. For example, an acceleration sensor or a gyro sensor (angular velocity sensor) is attached to a predetermined part of the operation main, and displacement acceleration (for example, triaxial displacement acceleration may be used), rotational speed or angular velocity (for example, triaxial rotational speed or angular velocity) may be used. ) To obtain motion data.

  Further, for example, a motion part or the like may be used to acquire a transition of a predetermined part of the operation main body, and a displacement acceleration, a rotation speed, or an angular speed of the predetermined part may be obtained to obtain motion data.

  The part for obtaining the motion data may be one place or a plurality of places. For example, in the case where a human is the main operator, motion data corresponding to the movement of the human waist may be acquired.

  The degree of freedom of motion data is a value determined by variables of data acquired as motion data. For example, when acquiring 3-axis acceleration and 3-axis rotation speed at a certain point as motion data, 3 variables each. Has 6 degrees of freedom.

  If the degree of freedom of motion data is n, the dimension of the feature data created based on the motion data is m (n> m). For example, principal component analysis (PCA) may be performed on the motion data to create feature data having a lower dimension than the motion data.

  It is preferable that a plurality of sample data of the operation main is held for one operation main. Then, the sample data is clustered for each operation subject. For example, when m = 2, the clustered sample data is arranged on a two-dimensional plane. Then, an average value of a plurality of sample data of each cluster may be obtained and used as the feature value of each cluster. Further, the centroid value of a plurality of sample data of each cluster may be obtained and used as the feature value of each cluster.

  The feature value of the target data may be obtained based on the distance (first feature value) between the plot points of the target data (feature data of motion data to be identified) in the m-dimensional space and the feature points of the cluster. Alternatively, it may be a vertical distance (second feature value) from a plot point of target data (feature data of motion data to be identified) in an m-dimensional space to an approximate straight line or the like indicated by a cluster feature function. For example, a linear approximate straight line may be obtained by the least square method using the sample points of each cluster and used as the approximate straight line.

  The feature value may be obtained based on at least one of the second feature value and the second feature value. Based on the feature value of each cluster, a cluster to which target data is likely to belong may be determined, and it may be determined that the operating data of the sample data of the cluster is the same person as the operating data of the target data.

  According to the present invention, it is possible to identify an operation main with high accuracy using feature data having a dimension lower than that of motion data based on the motion data. Therefore, it is possible to accurately identify the operation main with less data and a calculation load.

(2) In this operation main identification method, operation main identification system, operation main identification program and information storage medium,
The operation main identification step, or the operation main identification unit,
The feature value of the target data may be calculated by weighting the first feature value obtained based on the target data and the feature point of the cluster and the second feature value obtained based on the target data and the feature function. .

  The variation of the sample data constituting the cluster in the m-dimensional space may be a pattern that is uniformly or concentratedly distributed around the feature point, and a pattern that is distributed along the feature function. In this way, these distribution patterns represent the characteristics of the main walking pattern of the motion. Therefore, the features of various walking patterns can be covered by using both the first feature value and the second feature value.

(3) In this operation main identification method, operation main identification system, operation main identification program, and information storage medium,
The operation main identification step, or the operation main identification unit,
The feature value of the target data may be calculated by determining the weighting specific gravity for the first feature value and the second feature value based on the distribution state of the sample data of the cluster.

  The variation of the sample data constituting the cluster in the m-dimensional space may be a pattern that is uniformly or concentratedly distributed around the feature point, and a pattern that is distributed along the feature function. These distribution patterns represent the characteristics of the main walking pattern of the motion.

  When the elements (sample data) constituting the cluster are distributed along the feature function, the weight of the second feature value may be higher than the first feature value.

  When the elements (sample data) constituting the cluster are distributed around the feature points, the weight of the first feature value may be higher than the second feature value.

  The specific gravity of the weighting of the first feature value and the second feature value may be determined for each cluster according to the distribution of elements (sample data) constituting the cluster. Also, the weighting specific gravity may be changed when the sample data of the cluster increases or decreases. The case where any one of the weighting factors is 0 is also within the scope of the present invention.

(4) In this operation main identification method, operation main identification system, operation main identification program, and information storage medium,
As the first data conversion process, a feature vector may be generated using at least two motion data having a given time difference.

  For example, a feature vector may be generated based on an autocorrelation covariance matrix of motion data q (t) at time t and motion data q (t−1) at time t−1.

  In this way, since the temporal change of the walking pattern can be reflected in the feature vector, not only whether it is similar at a given time point, but also whether the change of the walking pattern is similar It can be.

(5) In this operation main identification method, operation main identification system, operation main identification program, and information storage medium,
As the second data conversion process, feature data may be generated by performing principal component analysis on the feature vector.

(6) In this operation main identification method, operation main identification system, operation main identification program, and information storage medium,
The cluster information calculation step, or the cluster information calculation unit,
An approximate straight line may be obtained by using the least square method for the sample data of each cluster and used as a feature function of each cluster.

  The approximate straight line may be a primary approximate line.

(7) In this operation main identification method, operation main identification system, operation main identification program, and information storage medium,
The main actor is a human and the given motion may be walking.

(8) In this operation main identification method, operation main identification system, operation main identification program, and information storage medium,
The sample motion data acquisition step, or the sample motion data acquisition unit,
As the motion data to be the sample, obtain the motion data of a part within a predetermined range from the waist or waist of the plurality of operation mains,
The target motion data acquisition step, the target motion data acquisition unit,
You may acquire the motion data of the site | part which exists in the predetermined range from the waist | hip | lumbar of the said operation | movement main body used as identification object, or a waist | hip | lumbar as a target.

(9) In this operation main identification method, operation main identification system, operation main identification program, and information storage medium,
The m may be 2 or 3.

The flowchart explaining the flow of database production of sample data. The figure for demonstrating an example of motion data. The figure for demonstrating an example of motion data. The figure which has arrange | positioned the plot point of sample data in m-dimensional space. The figure which arranged the plot point of the sample data, the feature point, and the feature function in m-dimensional space. The flowchart which shows the flow of a cluster information creation process. The flowchart which shows the flow of an operation main identification process. The figure for demonstrating the example of a sample data database. The figure which shows the example of acquisition of motion data. The figure which shows the example of acquisition of motion data. The figure which shows the example of acquisition of motion data. The arrangement | positioning figure of the sample point in case there are three main components. An example of the functional block diagram of the operation | movement main identification system of this Embodiment. FIGS. 14A to 14C are diagrams showing examples of generating feature vectors by the first conversion process. The figure which shows the example of the production | generation of the feature vector by 1st conversion processing. Diagram of network system.

  DESCRIPTION OF EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings. The embodiments described below do not limit the contents of the present invention described in the claims. Also, not all of the configurations described below are essential constituent requirements of the present invention.

1. Action Actor Identification Method 1.1 Overview A method for identifying an action hero using motion data according to the present embodiment will be described.

  In the present embodiment, motion data (displacement acceleration, rotation speed) about a walking pattern (an example of a given motion) of a plurality of humans (an example of an action main) is acquired as sample motion data. Then, a first data conversion process is performed on the sampled motion data, a feature vector is generated, a second data conversion process is performed based on the feature vector, and feature data of a dimension m lower than the feature vector is created. Then, a sample data database (an example of a storage unit) is created as sample data in association with the actor.

  Then, the sample data is clustered in association with the main actors, arranged in an m-dimensional space, and cluster information including at least one of the cluster feature points and feature functions (approximate straight lines) is obtained.

  A first data conversion process is performed on the motion data (target motion data) to be identified, a feature vector is generated, a second data conversion process is performed based on the feature vector, and a feature of dimension m lower in dimension than the feature vector Create data and use it as target data, calculate the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions (approximate straight line) of the cluster, and the target data belongs based on the feature values A cluster having a high possibility may be determined, and the main actor may be identified based on the determination result.

  The main actor may be, for example, a human, an animal such as a dog or a cat, or a machine such as a robot. However, in this embodiment, a description will be given by taking the case of a human as an example.

1.2 Creating Sample Data Database FIG. 1 is a flowchart for explaining the flow of creating a sample data database.

  First, motion data for a given operation of a plurality of operation owners is acquired as sample motion data (step S10). Next, a first data conversion process is performed on the sampled motion data to generate a feature vector (step S20). Next, a second data conversion process is performed on the generated feature vector to generate feature data having a dimension m lower than that of the feature vector (step S30). Next, the generated feature data is stored in the storage unit as sample data in association with the actor (step S40).

1.3 Acquisition of motion data The motion data may be data representing the movement of a predetermined part of the operation subject. For example, a displacement acceleration (for example, a triaxial displacement acceleration) or a rotation speed or an angular velocity (for example, a triaxial rotation speed or an angular speed) of a predetermined part of the motion may be acquired and used as motion data.

  2 and 3 are diagrams for explaining an example of motion data.

  Conventionally, when acquiring motion data of a subject, a plurality of markers are attached as shown in FIG. 2 and walking data is measured using motion capture. For example, using a 34-degree-of-freedom model in which 15 parts (rigid links) as shown in FIGS. 2 and 3 are used, the movement of the spine and the movement between the head and the pelvis are taken into account, so that walking is represented. Enough. Using such a model, motion data during walking is acquired, and human identification is attempted by various methods. However, because the amount of data to be acquired is large, analysis takes time and the main actor is identified in real time. There was a problem that it was difficult.

  In the present embodiment, it is possible to identify the main actor with the motion data of at least one or two parts. However, the case of using motion data of three or more parts is also within the scope of the present invention.

  Hereinafter, as an example of motion data to be sampled, the case of acquiring motion data of an operation main body P1 and a part within a predetermined range from the waist (for example, the groove P2) will be described as an example.

  9, FIG. 10 and FIG. 11 are diagrams showing examples of motion data acquisition.

For example, as shown in FIG. 9, an IMU including a motion sensor 410 (acceleration sensor and gyro sensor (angular velocity sensor)) is attached near the waist of the subject 400, and the displacement acceleration of the waist (for example, triaxial displacement acceleration may be used) Rotational speed or angular speed (for example, triaxial rotational speed or angular speed) may be acquired and used as motion data.
The IMU is a sensor that detects triaxial angular velocity and acceleration. A three-dimensional gyro and a three-direction accelerometer determine the three-dimensional angular velocity and acceleration. There are three main gyro principles: inertia, precession, Coriolis force, and Sagnac effect. The Coriolis force is “a force generated in a direction in which both the velocity and the angular velocity are orthogonal to each other when an angular velocity is applied to an object having a velocity”. For example, when an angular velocity is applied in a state where the mass m is vibrating in the X-axis direction, a Coriolis force is generated in the Y-axis direction, causing vibration. Since this vibration amplitude is proportional to the angular velocity, the input angular velocity can be measured by detecting the amplitude. Further, the acceleration is obtained from the weight displacement amount according to Hooke's law.

  For example, as shown in FIG. 10, the motion data of the subject 400 may be acquired by a general-purpose motion capture device 420 used as an input device for games or the like. The motion capture device 420 includes a camera, a depth sensor, a multi-array microphone, and a sensor with a built-in processor that operates dedicated software, and can recognize the position, movement, voice, and face of the subject 400. Unlike general motion capture, the motion capture device 420 does not require a special suit with a marker to be worn at the time of normal motion capture and a tracker used at the time of marker detection. The distance from the subject 400 to the motion capture device 420 is measured by projecting the subject on the camera included in the motion capture device 420, and various movements of the skeleton of the subject 400 are detected.

  For example, as shown in FIG. 11, the motion data may be acquired by a normal motion capture method in which a marker 432 is attached to the subject 400 and the camera 430 captures an image.

1.4 Generation of Feature Vector Next, a method for generating a feature vector by performing a first data conversion process on motion data will be described.

  The degree of freedom of motion data is determined by the number of data variables acquired as motion data. For example, when acquiring three-axis acceleration and three-axis rotation speed at a certain point as motion data, each has three variables, so the degree of freedom is six.

  Therefore, when acquiring motion data of two points (waist P1 and groove P2), the degree of freedom is 12.

  The structure of the feature vector of motion data will be described. The motion data includes the three-axis displacement acceleration (three degrees of freedom) of the waist (P1 in FIG. 2), the three-axis rotational speed (three degrees of freedom), and the three-axis displacement acceleration (three degrees of freedom) of the groove (P2 in FIG. 2). ) And 3-axis rotation speed (3 degrees of freedom) in total 12 degrees of freedom data. This model is called a 12 DOF (Degree Of Freedom) model.

  Then, the matrix shown in (Equation 1) is assumed to be (Equation 3), which is a covariance matrix based on the autocorrelation of the time series vector shown in (Equation 2).

Here, Ti is the length of the motion i, (Equation 4) is a motion vector at time k with N (N = 12) degrees of freedom for the i-th walk, and time k is a walk ring frequency of 33 Hz. It is a discrete time, and l is a time difference (here, l = 2).

  FIGS. 14A to 14C and FIG. 15 are diagrams illustrating examples of generating feature vectors by the first conversion process. FIG. 14A shows a motion vector at time t. FIG. 14B shows a vector obtained by transposing a motion vector at time t−1 (here, l = 2). FIG. 14C is a matrix that is the product of a vector obtained by transposing the motion vector at time t and the motion vector at time tl (here, l = 2).

  15 in FIG. 15 indicates the covariance matrix shown in (Equation 3), and 360 is a feature vector generated by rearranging the elements of the covariance matrix 350 in one line. The feature vector is expressed by the following equation.

The degree of freedom of the feature vector is N squared. Therefore, when N is 12, the degree of freedom of the feature vector is 144.

  This feature vector configuration has several features. By taking a time average of products defined locally in terms of time, it is possible to handle motions of any time width, and there is no need to normalize the length of motion data in advance. In addition, the speed can be reflected by taking the product by shifting the time. Furthermore, the long-range correlation can be reflected by the spatial globality of taking the product of combinations of all dimensions of the vector.

  By generating a feature vector using at least two motion data having a given time difference, such as time tl and time t, not only whether or not they are similar at a given time, but also changes in motion data It is possible to identify walking patterns including whether or not they are similar.

  In the above example, the case where a feature vector is generated using at least two motion data having a time difference has been described as an example. However, a feature vector is generated using one motion vector at a given time point and a transposed vector. May be.

1.5 Generation of Feature Data When the degree of freedom of motion data is n, the dimension of feature data generated by performing the second conversion process on motion data is m (n> m). For example, as the second conversion process, principal component analysis (PCA) or the like may be performed on the motion data to generate feature data having a lower dimension than the motion data.

  In the present embodiment, principal component analysis (an example of second data conversion processing) is performed on the feature vector for visualization and generation of a recognition algorithm. Here, a case where the second principal component is plotted by principal component analysis will be described as an example.

  Principal component analysis is a statistical analysis technique that summarizes information of multivariate data into a small number of total characteristic values. Extracting unrelated (non-correlated) components from the statistical data and explaining the observed value by a linear combination of these components is called principal component analysis, and the extracted components are called principal components.

  According to principal component analysis, the information of multidimensional data can be reduced to a low-dimensional space without losing as much as possible. If multidimensional data can be reduced to two-dimensional or three-dimensional data, the distribution of the entire data can be visualized, and differences between sample data can be clarified. For PCA procedures, either eigenvectors or singular value decomposition usually appear. This is because the unit vector of the principal component is immediately obtained as an eigenvector, and the size of the principal component is immediately obtained using singular value decomposition.

In the present embodiment, principal component analysis is performed in the following procedure.
(1) Finding a covariance matrix by autocorrelation of time series data (2) Finding a feature vector from the covariance matrix (3) Finding the left singular matrix of the matrix formed by the feature vector and its corresponding singular value (4) The order of principal components is determined in descending order of singular values. (5) The direction of each principal component is determined by the column vector of the left singular value matrix to which each corresponds.

1.6 Sample Data Database FIG. 8 is a diagram for explaining an example of the sample data database.

  The sample data database 500 stores sample data (feature data) in association with the operation main identification information 510. The operation main identification information 510 is information for identifying the subject. The sample data (feature data) is the values of the first principal component (x) and the second principal component (y) obtained by principal component analysis.

1.7 Cluster Information Creation Processing In this embodiment, an operation main is identified using a cluster analysis technique. Cluster analysis or clustering is to divide a set to be classified into subsets that achieve internal coupling and external separation.

  FIG. 6 is a flowchart showing the flow of the cluster information creation process.

  The sample data is clustered in association with the main operator and arranged in the m-dimensional space. (Step S110). Until all the clusters are processed (N in step S130), cluster information (i) including at least one of feature points and feature functions of cluster (i) is obtained based on the sample data of cluster (i) (step S130). S120)

  FIG. 4 shows plot points of sample data (feature data) of motion data acquired when a subject walks in an m-dimensional space (here, m = 2). A set of plot points for each subject is associated with each subject and clustered. C1 in FIG. 4 is a plot point of the sample data of the subject 1, C2 is a plot point of the sample data of the subject 2, C3 is a plot point of the sample data of the subject 3, and C4 is the sample data of the subject 4 Plot points. The position, shape, and size of each cluster (set), which is a set of plot points for each subject, have different characteristics for each subject. This is considered to be an individual difference caused by the walking pattern having different characteristics.

  In order to identify the walking pattern of the target motion data using this feature, the feature points and feature functions of each cluster may be obtained.

  Five plot points of each subject are regarded as clusters (sets), and four clusters (sets) are formed in an orthogonal coordinate system composed of a first principal axis (x axis) and a second principal axis (y axis) as shown in FIG. The feature points may be the average coordinate points of the five components of each cluster (set).

  Therefore, the feature points of the set j (1 to 4) are as follows.

Here, n is the number of feature vectors, that is, the number of points in the set.

  Further, a linear approximation line for each set may be obtained using the least square method for the components in the set, and may be used as a feature function.

  FIG. 5 is a diagram illustrating feature points and approximate lines.

  Reference numerals 312 and 310 denote feature points and feature functions of cluster 1 (sample data of subject 1), which is a set of plot points C1, and reference numerals 322 and 320 denote cluster 2 (sample data of subject 2), which is a set of plot points C2. ), 332 and 330 are the feature points and feature functions of cluster 3 (sample data of subject 3), which is a set of plot points C, and 342 and 340 are sets of plot points C4 Are feature points and feature functions of cluster 4 (sample data of subject 4).

1.8 Actuator Identification Process FIG. 7 is a flowchart showing the flow of the actor identification process.

  As the motion data to be the target, the motion data for a given operation of the operation subject to be identified is acquired (step S210). Next, a first data conversion process is performed on the target motion data to generate a feature vector (step S220). Next, a second data conversion process is performed on the generated feature vector to generate feature data having a dimension m lower than that of the feature vector, which is used as target data (step S230). Next, until the processing is completed for all clusters (N in step S250), based on the positional relationship in the target data, at least one of the feature points and feature functions of cluster (i) in the m-dimensional space, the feature value ( i) is calculated (step S240).

  Based on the feature values of all clusters, a cluster to which target data is highly likely to belong is determined (step S260). Then, the main operator associated with the determined cluster is determined as the main operator of the target data (step S270).

  The distance between the i-th point, which is the first feature value, and the feature point of the j group can be expressed by the following equation.

Further, the vertical distance from an arbitrary data point, which is the second feature value, to the approximate straight line of the set j is obtained by the following equation.

Here, the characteristic value for the group j of the i-th data point is defined by the following equation.

This feature value configuration has several features. As can be seen from FIG. 5, the variation in C3 is small, whereas the variation in C4 is large and is distributed almost linearly. Equation 9 shows the distance (first feature value) between the plot point of the sample data indicating the degree of concentration at the center of the cluster (set) and the feature point (first feature value) and the vertical distance (second feature value) between the approximate line. Since both can be reflected, both sets having different distribution patterns such as C2 and C4 can be handled. In addition, in Equation 9, which of the distance to the feature point (first feature value) and the vertical distance to the approximate line (second feature value) is prioritized according to the distribution state of the elements of the four sets? It can be adjusted by attaching a corresponding weight coefficient.

  When human identification is performed, the feature vector is analyzed by PCA, and when it is arranged on the two-dimensional plane as feature data composed of the first and second principal components, the features obtained using the target data plot points and cluster information There is a high possibility that the operator of the cluster having the smallest value is the operator of the target data.

  The feature value of the target data may be calculated by determining the weighting specific gravity for the first feature value and the second feature value based on the distribution state of the sample data of the cluster.

  The variation of the sample data constituting the cluster in the m-dimensional space is distributed along the feature function as in C4 while there is a pattern distributed uniformly or concentrated around the feature points as in C3. There are also patterns. These distribution patterns represent the characteristics of the main walking pattern of the motion.

  When the elements (sample data) constituting the cluster are distributed along the feature function as in C4, the specific gravity of the weight of the second feature value may be higher than the first feature value. .

  When elements (sample data) constituting the cluster are distributed around the feature points as in C3, the weight of the first feature value may be higher than the second feature value. .

  As described above, the weight of the first feature value and the second feature value may be determined according to the distribution of the elements (sample data) constituting the cluster for each cluster. Also, the weighting specific gravity may be changed when the sample data of the cluster increases or decreases. The case where any one of the weighting factors is 0 is also within the scope of the present invention.

  In the above embodiment, the case where the dimension of the feature data is 2 (the case where there are two main components) has been described as an example, but the present invention is not limited thereto. The dimension of feature data may be 3 or more (when there are 3 or more principal components).

  In this embodiment, since direct clustering is performed in a low-dimensional space, the amount of calculation is less than that of a nonlinear SVM (Supporting Vector Machine) that converts the original low-dimensional feature space into a multi-dimensional space that can be linearly separated. Therefore, it is considered that real-time recognition is excellent. Further, since the feature vector can be visualized, there is an advantage that the overall structure and correlation of the cluster can be intuitively grasped.

  Alternatively, the cluster having a clear structure may be recognized first, and if the recognition result is negative, the cluster may be excluded and the PCA may be performed only on the remaining data again. In this way, the distribution of the remaining data becomes larger and clustering becomes easier.

  FIG. 12 shows a case where plot points of sample data (feature data) of motion data acquired when a subject walks are arranged in a three-dimensional space (here, m = 3) when there are three main components. In this case, the feature point of each cluster is obtained as three-dimensional data, and the feature function is obtained as a three-dimensional function.

2. 2. Action Main Identification System 2.1 Configuration of Action Main Identification System FIG. 13 is an example of a functional block diagram of the action main identification system of the present embodiment.

  The operation main identification system 1 of this embodiment does not need to include all the components (respective parts) of FIG. 13 and may have a configuration in which some of them are omitted. In addition, the operation main identification system 1 of this Embodiment is realizable using a general purpose computer, for example.

  The operation main identification system 1 according to the present embodiment includes a storage unit 20. The storage unit 20 stores programs, data, and the like, and functions as a computer-readable medium (optical disk (CD, DVD, etc.), magneto-optical disk (MO), magnetic disk, hard disk, magnetic tape, or It can be realized by hardware such as a memory (ROM). The storage unit 20 may include hardware such as a RAM having a function as a temporary storage unit serving as a work area such as the processing unit 10.

  The storage unit 20 stores an operation main identification program 22 according to the present embodiment. The operation main identification program 22 is a unit (database creation processing unit 12, cluster information calculation processing unit 14, operation main identification processing unit 16, sample data storage unit 24) described below for the computer (processing unit 10, storage unit 20). It is a program to make it function as.

  The storage unit 20 functions as the sample data storage unit 24.

  The operation main identification system 1 of the present embodiment includes a processing unit 10. The processing unit 10 performs various processes based on the input data 16 and programs, and the functions are realized by hardware such as various processors (CPU, DSP, etc.), ASIC (gate array, etc.), and programs. it can. The processing unit 10 functions as a database creation processing unit 12, a cluster information calculation processing unit 14, and an operation main identification processing unit 16.

  The database creation processing unit 12 performs a first data conversion process on the sampled motion data, generates a feature vector, performs a second data conversion process on the feature vector, and calculates a dimension from the feature vector. Feature data of a low dimension m is generated and stored as sample data in the storage unit in association with the main actor.

  The sample data storage unit 24 performs a first data conversion process on the sample motion data and performs a second data conversion process on the generated feature vector. Feature data of a low dimension m is stored as sample data in association with the actor.

  The cluster information calculation processing unit 14 clusters the sample data for each operation main body and obtains cluster information including at least one of feature points and feature functions of the cluster in the m-dimensional space.

  The action main identification processing unit 16 performs a first data conversion process on the target motion data, generates a feature vector, performs a second data conversion process on the feature vector, and performs a dimension based on the feature vector. The feature data of the low dimension m is generated as target data, and the feature value of the target data is calculated based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster. Identify the actor.

  The action main identification processing unit 16 weights the first feature value obtained based on the target data and the feature points of the cluster and the second feature value obtained based on the target data and the feature function, and features the target data. The value may be calculated.

  In addition, the operation main identification processing unit 16 may calculate the feature value of the target data by determining the weight of the first feature value and the second feature value based on the distribution state of the sample data of the cluster.

  In addition, the cluster information calculation processing unit 14 may obtain an approximate straight line by using the least square method for the sample data of each cluster and may use it as a feature function of each cluster.

  The operation main identification system 1 may be configured by a single computer, or may be configured to execute processing in a distributed manner by a plurality of computers.

  The creation of the sample data database may be a system configuration that is performed by another system.

2.2 Other Configurations of Operator Identification System FIG. 16 shows an operator identification system implemented as a network system 100 including a server and a terminal (an operator identification device). The network system 100 includes a plurality of terminals 110A, 110B, 110C,. As shown in FIG. 16, the network system 100 of the present embodiment is configured such that a server 120 and terminals 110A, 110B, 110C,.

  The server 120 is a terminal (for example, an operation main identification device built in a robot, a personal computer, a portable information communication device such as a mobile phone or a smartphone, a game terminal, etc.) 110A, 110B,. This is an information processing apparatus capable of providing a service for identifying an operating person based on the motion data to be identified acquired by.

  The server 120 manages a database of sample data of a plurality of users.

  The terminals 110A, 110B, 110C,... Are, for example, an operation main identification device built in a robot, a portable terminal (mobile phone, PHS terminal, smartphone, PDA, portable game machine, etc.), personal computer (PC), game It is a device that can be connected to the server 120 via a network 130 such as the Internet (WAN) or LAN. The communication line between the terminals 110A, 110B, 110C,... And the server 120 may be wired or wireless.

  The terminals 110 </ b> A, 110 </ b> B, 110 </ b> C,... Have a communication control function for communicating with the server 120, a motion capture function for acquiring motion data to be identified, and the like.

  The terminals 110A, 110B, 110C,... May be configured as devices configured with dedicated hardware, or may be configured to install dedicated client software (applications) on a general-purpose computer (dedicated application network The configuration realized in (including the form downloaded and executed in the above) is also within the scope of the present invention.

  The server 120 of this embodiment may have a Web server function. Further, the terminals 110A, 110B, 110C,... May include a web browser capable of browsing a web page (HTML (Hyper Text Markup Language) format data).

  Also, a form that is realized only by the web browser function of the terminal without using dedicated software is within the scope of the present invention.

  The server 120 obtains motion data for a given operation of a plurality of actors as sample motion data, performs a first data conversion process on the sample motion data, and generates a feature vector. Then, a second data conversion process is performed on the feature vector to generate feature data having a dimension m lower than that of the feature vector, and a database is stored as sample data in association with the actor. The database may be created by the server 120, or a database created by another system may be used.

  The terminals 110 </ b> A, 110 </ b> B, 110 </ b> C,... Acquire the motion data of the operation subject to be identified as the target motion data. The process of identifying the main actor using the target motion data may be performed on the terminal side or on the server 120 side.

  And terminal 110A, 110B, 110C, ... may perform a given process using the identification result of the main actor of the target motion data. For example, when the terminals 110A, 110B, 110C,... Are motion main identification devices built in the robot, the motion capture function provided in the robot acquires human motion data existing around the robot. The acquired motion data may be transmitted to the server 120. Then, the server may perform an operation main identification process based on the received motion data, and transmit the identification result to the requesting terminals 110A, 110B, 110C,. The terminals 110A, 110B, 110C,... May recognize a person existing around the robot based on the identification result and perform a given process.

  The operation main analysis processing according to the present embodiment may be performed partially or entirely by the terminals 110A, 110B, 110C,... Or may be realized by cloud computing.

  In addition, this invention is not limited to this embodiment, A various deformation | transformation implementation is possible within the range of the summary of this invention.

The present invention includes configurations that are substantially the same as the configurations described in the embodiments (for example, configurations that have the same functions, methods, and results, or configurations that have the same objects and effects). In addition, the invention includes a configuration in which a non-essential part of the configuration described in the embodiment is replaced. In addition, the present invention includes a configuration that exhibits the same operational effects as the configuration described in the embodiment or a configuration that can achieve the same object. Further, the invention includes a configuration in which a known technique is added to the configuration described in the embodiment.

DESCRIPTION OF SYMBOLS 1 Action main identification system 10 Processing part 12 Database creation processing part 14 Cluster information calculation processing part 16 Operation main identification processing part 20 Storage part 22 Action main identification program 24 Sample data storage part

Claims (20)

A method for identifying an actor,
As sample motion data, a sample motion data acquisition step for acquiring motion data for a given operation of a plurality of actors;
First data conversion processing for the sample to become motion data to generate the feature vectors do, generates feature data of the lower dimensional m dimensioned than the feature vector performs a second data conversion processing on the feature vectors And a database creation step of storing in the storage unit in association with the main actor as sample data,
A cluster information calculation step for clustering sample data for each operation main body to obtain cluster information including at least one of feature points and feature functions of the cluster in the m-dimensional space;
As a target motion data, a target motion data acquisition step for acquiring motion data of an operation subject to be identified;
A first data conversion process is performed on the target motion data, a feature vector is generated, a second data conversion process is performed on the feature vector, and feature data having a dimension m lower than the feature vector is obtained. Generated as target data, calculates the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster, and identifies the actor main based on the feature value and the step, only including,
In the operation main identification step,
The first feature value obtained based on the target data and the feature point of the cluster and the second feature value obtained based on the target data and the feature function are weighted, and the feature value of the target data is calculated. The operation main identification method. In claim 1 ,
In the operation main identification step,
An operation main identification method comprising: calculating a feature value of the target data by determining a weighting specific gravity for the first feature value and the second feature value based on a distribution state of the sample data of the cluster. In claim 1 or 2 ,
As the first data conversion process, a feature vector is generated using at least two motion data having a given time difference. In any one of claims 1 to 3,
An operation main identification method characterized in that, as the second data conversion process, a principal component analysis is performed on a feature vector to generate feature data. In any one of claims 1 to 4,
In the cluster information calculation step,
An operation main identification method characterized in that an approximate straight line is obtained by using the least square method for sample data of each cluster and used as a feature function of each cluster. In any one of claims 1 to 5,
A method for identifying an action main person, characterized in that the action main person is a human and the given action is walking. In claim 6 ,
In the sample motion data acquisition step,
As the motion data to be the sample, obtain the motion data of a part within a predetermined range from the waist or waist of the plurality of operation mains,
In the target motion data acquisition step,
A motion main identification method characterized in that motion data of a part within a predetermined range from the waist of the motion main subject to be identified or a waist is acquired as the target motion data. In any one of claims 1 to 7,
The operation main identification method, wherein m is 2 or 3. An operator identification system for identifying an operator,
For the sample to become motion data, generates a feature vector performs a first data conversion process, the feature data of the feature vectors for the second performs a data conversion process the feature vectors from the lower dimension dimension m A database creation unit that generates and stores in the storage unit as sample data in association with the actor;
A cluster information calculation unit for clustering sample data for each operation main body to obtain cluster information including at least one of feature points and feature functions of a cluster in an m-dimensional space;
A first data conversion process is performed on the target motion data, a feature vector is generated, a second data conversion process is performed on the feature vector, and feature data having a dimension m lower than the feature vector is obtained. Generated as target data, calculates the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster, and identifies the actor main based on the feature value and parts, only including,
The operation main identifier is
The first feature value obtained based on the target data and the feature point of the cluster and the second feature value obtained based on the target data and the feature function are weighted, and the feature value of the target data is calculated. The operation main identification system. An operator identification system for identifying an operator,
Feature data of a dimension m lower than the feature vector generated by performing the second data conversion processing on the feature vector generated by performing the first data conversion processing on the sample motion data, A sample data storage unit stored as sample data in association with the main actor,
A cluster information calculation unit for clustering sample data for each operation main body to obtain cluster information including at least one of feature points and feature functions of a cluster in an m-dimensional space;
A first data conversion process is performed on the target motion data to generate a feature vector, and a second data conversion process is performed on the feature vector to generate feature data of dimension m lower than the feature vector. An action main identification unit that calculates the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster, and identifies the action owner based on the feature value When,
Only including,
The operation main identifier is
The first feature value obtained based on the target data and the feature point of the cluster and the second feature value obtained based on the target data and the feature function are weighted, and the feature value of the target data is calculated. The operation main identification system. In claim 9 or 10 ,
The operation main identifier is
An operating main identification system characterized in that the feature value of the target data is calculated by determining the weighting specific gravity for the first feature value and the second feature value based on the distribution state of the sample data of the cluster. In any one of claims 9 to 11,
As the first data conversion process, a feature vector is generated using at least two motion data having a given time difference. In any one of claims 9 to 12,
An operation main identification system characterized in that, as the second data conversion process, a principal component analysis is performed on a feature vector to generate feature data. In any one of claims 9 to 13,
The cluster information calculation unit includes:
An operation main identification system characterized in that an approximate straight line is obtained by using the least square method for sample data of each cluster and used as a feature function of each cluster. In any one of claims 9 to 14,
An action main identification system characterized in that an action main person is a human and a given action is walking. In claim 15 ,
As sample motion data, a sample motion data acquisition unit that acquires motion data for a given operation of a plurality of operation owners;
A target motion data acquisition unit that acquires motion data of an operation subject to be identified as target motion data;
The sample motion data acquisition unit
As the motion data to be the sample, obtain the motion data of a part within a predetermined range from the waist or waist of the plurality of operation mains,
The target motion data acquisition unit
The motion main identification system characterized by acquiring motion data of a part within a predetermined range from the waist or waist of the motion main subject to be identified as the target motion data. In any one of claims 9 to 16,
The operating main identification system, wherein m is 2 or 3. A program for identifying an operating person using a computer,
A first data conversion process is performed on the sample motion data to generate a feature vector, a second data conversion process is performed on the feature vector, and feature data having a dimension m lower than the feature vector is obtained. A database creation unit that generates and stores in the storage unit as sample data in association with the actor;
A cluster information calculation unit for clustering sample data for each operation main body to obtain cluster information including at least one of feature points and feature functions of a cluster in an m-dimensional space;
A first data conversion process is performed on the target motion data to generate a feature vector, and a second data conversion process is performed on the feature vector to generate feature data of dimension m lower than the feature vector. An action main identification unit that calculates the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster, and identifies the action owner based on the feature value When,
Make your computer work ,
The operation main identifier is
The first feature value obtained based on the target data and the feature point of the cluster and the second feature value obtained based on the target data and the feature function are weighted, and the feature value of the target data is calculated. programs that. A program for identifying an operating person using a computer,
Feature data of a dimension m lower than the feature vector generated by performing the second data conversion processing on the feature vector generated by performing the first data conversion processing on the sample motion data, A sample data storage unit stored as sample data in association with the main actor,
A cluster information calculation unit for clustering sample data for each operation main body to obtain cluster information including at least one of feature points and feature functions of a cluster in an m-dimensional space;
A first data conversion process is performed on the target motion data to generate a feature vector, and a second data conversion process is performed on the feature vector to generate feature data of dimension m lower than the feature vector. An action main identification unit that calculates the feature value of the target data based on the target data in the m-dimensional space and at least one of the feature points and feature functions of each cluster, and identifies the action owner based on the feature value When,
Make your computer work ,
The operation main identifier is
The first feature value obtained based on the target data and the feature point of the cluster and the second feature value obtained based on the target data and the feature function are weighted, and the feature value of the target data is calculated. programs that. A computer-readable information storage medium, wherein the program according to claim 18 or 19 is stored.