JP2018068516A - Exercise motion evaluation system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To objectively evaluate play and performance of a player who is taking exercise.SOLUTION: A series of motions of a player who is taking exercise are acquired as pieces of motion data. From the pieces of motion data, parts which are evaluation objects (actions) are extracted on the basis of a condition such as a physical relationship of a player. Then, by a classification type machine learning, to the extracted actions as objects, an action corresponding to play or performance to be the evaluation object is identified. Then, to the identified action, a regression type machine learning is used for evaluating completion degree of play or performance. In the step, the classification type machine learning and regression type machine learning use a common feature vector. Therefore, the play or performance can be objectively and accurately evaluated.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a technique for evaluating the degree of perfection as a specific play or performance in relation to the movement of a player during exercise.

In various sports, in order to improve the player's ability, it is required to evaluate the player's play in practice and games and to instruct him to make better movements. In some sports, such as figure skating, the movement of the player itself may be scored. Conventionally, the evaluation of the movements of players has been performed by coaches and referees based on experience, and it cannot be said that the objectivity is sufficiently secured. For example, when a recorded player's movement is played back and the movement is evaluated, the evaluation result usually varies with each evaluation even though the same person is evaluating the same movement.
Attempts have been made to use computers to evaluate the movement of these players. Non-Patent Document 1 discloses a technique for determining a screen play used by an attacker in a basketball by a computer. Specifically, it is a technology that converts the position and movement of each player in basketball into digital data and loads it into a computer, and cuts out a scene of screen play from a series of movements. In this technique, segmentation-type machine learning is used to perform segmentation.

K. Fujii, 5 other authors, “Resilient help to switch and overlap hierarchical subsystems in a small human group”, Scientific Reports 6, 23911 Armand MacQueen and 2 other authors, “Automatically Recognizing On-Ball Screens”, 2014 MIT Sloan Sports Analytics Conference, 2014

Attempts to use computers, especially machine learning, for the movement of athletes during exercise have just begun, and even the limited play of basketball screen play is the prior art disclosed in Non-Patent Document 1. It cannot be said that it has been completed. For example, the prior art covers only screen play related to a player who holds the ball during play, but screen play is also performed between players who do not hold the ball, so the play is cut out. Is not sufficient in that it cannot be performed. Further, the prior art is merely a technique for extracting a screen play from a series of movements, and has not yet been evaluated by a computer. Such a problem is not unique to basketball screen play, but is a common problem when evaluating the movement of players in various sports.
In view of these problems, an object of the present invention is to provide a technique for evaluating the movement of a player during exercise using a computer.

The present invention
An exercise motion evaluation system for evaluating the completeness of a specific play or performance with respect to the movement of a player during exercise,
An operation data input unit for reading operation data representing the movement of the player;
Based on the motion data, it can be configured as an exercise motion evaluation system including an evaluation unit that performs the evaluation by regression type machine learning using a feature vector set by classification type machine learning for the play or performance. .

According to the present invention, since a player's movement can be evaluated by using regression machine learning, an objective evaluation is possible. In general, regression type machine learning requires a feature vector for evaluating a motion to be evaluated. In the present invention, a feature vector set by classification type machine learning is used. In other words, whether or not the various movements of the players correspond to the movements to be evaluated is trained by supervised classification machine learning, and in this process, feature vectors that are suitable for this classification are found. Then, using this feature vector, the training of the regression machine learning, that is, teaching the evaluation score for the movements of various players is repeatedly executed. With sufficient training in this way, new movements can be evaluated by regression machine learning.
The inventor researched suitable feature vectors by performing classification-type machine learning and regression-type machine learning using various feature vectors, and found that feature vectors with good results in classification-type machine learning were regressed. By using it for machine learning, we found the fact that regression-type machine learning can also provide stable evaluation results. The principle of obtaining such a result is not necessarily clear, but it is considered that one of the reasons is that there is no need to change the parameters used for the analysis in the classification and regression of the movement to be evaluated.
In general, regression type machine learning can take various evaluation points, so it is difficult to judge whether the feature vector is good or bad. On the other hand, classification type machine learning trains according to whether it corresponds to the motion to be evaluated. It is considered easy to judge good or bad. In the present invention, since the feature vector is set by classification-type machine learning in which it is easy to determine whether the quality is good or bad, there is an advantage that the accuracy of regression type machine learning can be easily improved.

The evaluation object in the present invention is the completeness of a specific play or performance. Examples of the former include a standard play in which a plurality of players cooperate in team play such as a basketball screen play, a technique in judo or kendo, and the like. Examples of the latter include acting in figure skating.
The movement to be read as operation data may be determined according to the evaluation target. For example, in the case of basketball screen play, the position and movement of each player and ball are read. In the case where a technique such as judo or performance in figure skating is targeted, in addition to the position of the player, the movement of the limbs and the like are also read. Various known techniques can be applied to the method of converting motion data into digital data.
The present invention is an aspect that is used for instruction by evaluating the player's play, an aspect that improves the objectivity of scoring by evaluating the performance of the player in the game, the evaluation by the referee for the movement of the player, and the evaluation according to the present invention. Can be used in various modes, such as a mode used for judging a referee's ability.

In the present invention,
The motion data is data representing a series of movements including play or performance to be evaluated,
Prior to the evaluation by the evaluation unit, a segmentation unit may be provided that extracts, from the motion data, a portion corresponding to the play or performance to be evaluated by classification machine learning.

According to this aspect, the play or performance to be evaluated can be easily extracted by the present invention, and the burden on the evaluator can be reduced. For example, in the case of basketball, since many screen plays are performed in one game, the effort to cut out the movement of the part where such play is performed is not light. However, according to the said aspect, this operation | work can be performed by this invention.
In the present invention, classification-type machine learning is used for the above-described segmentation, that is, segmentation. Since the feature vectors used for this machine learning are the same as those used for motion evaluation, the present invention has the advantage that classification machine learning can be applied without preparing a new feature vector. is there.

  In addition, prior to the classification type machine learning, the segmentation unit extracts a section where the movement of the player satisfies a predetermined condition set according to the play or performance from the motion data, and the extracted The classification-type machine learning may be applied to a section.

In the above-described aspect, before the classification type machine learning is applied, the section corresponding to the play or performance to be evaluated is easily extracted. Since the machine learning is not used for the extraction in advance, it is possible to perform processing with a relatively light load. As described above, the combination of the extraction process that does not use machine learning and the classification machine learning makes it possible to perform segmentation more effectively than the classification machine learning is applied to the entire series of operations.
The predetermined conditions used for extraction can be set according to the play or performance to be evaluated. For example, in the case of basketball screen play, a condition may be set such that the distance between players approaches within a predetermined distance. For jumps in figure skating, conditions can be set such that both feet are away from the ice. The conditions for extraction need not be limited to one, and two or more conditions can be used.

The segmentation unit further applies the classification type machine learning after the plurality of sections are merged into one section when the plurality of sections extracted are determined as a series of sections. Also good.
When extracting from a player's movement data under certain conditions, there are usually a series of movements such as noise included in the movement data and when the player's movement is near the boundary of whether or not the condition is met. There is a possibility that the situation will be divided. In such a case, if the classification-type machine learning is applied to the divided sections, the play or performance to be evaluated may not be accurately determined.
In the above aspect, the efficiency and accuracy of segmentation by classification-type machine learning can be improved by performing a process of fusing the divided sections into one section. The conditions for merging can be arbitrarily set based on, for example, the time interval between divided sections.

Regardless of whether or not a segmented portion is provided, in the present invention,
The feature vector may be set based on evaluation of a result obtained by performing the classification machine learning on a plurality of types of feature vectors and combinations thereof.
By doing this, it is possible to select one having a good evaluation result from a plurality of types of feature vectors and combinations thereof. Therefore, it is possible to improve the accuracy of classification-type machine learning, and hence the evaluation accuracy of regression-type machine learning.

The present invention is applicable to various exercises, for example
The play is a screen play in basketball,
The feature vector may be set based on the positional relationship of the four players related to the screen play without considering the positional relationship with the basketball goal.
The inventor has found that by using such a feature vector, it is possible to extract and evaluate screen play performed everywhere regardless of the position of the ball or goal in basketball. Therefore, the use value of the present invention in the extraction and evaluation of basketball screen play can be greatly improved by using such feature vectors.

The present invention does not necessarily include all of the various features described above, and may be configured by omitting or combining some of them as appropriate.
The present invention is not limited to the above-described aspect of the exercise motion evaluation system, but may be configured as an exercise motion evaluation method for evaluating exercise motion by a computer. Moreover, it can also be comprised as a computer program for making a computer perform this evaluation, and a computer-readable recording medium which recorded such a program. Here, as a computer-readable recording medium, various media such as a flash memory, a hard disk, and an optical disk can be used.

It is explanatory drawing which shows the structure of an exercise | movement motion evaluation system. It is a flowchart of an action extraction process. It is explanatory drawing which shows the process example of action extraction. It is a flowchart of a play identification process. It is a flowchart of a play evaluation process. It is a flowchart of a machine learning training process. It is explanatory drawing which shows the example of evaluation of a restriction | limiting parameter. It is explanatory drawing which shows the example of evaluation of a regression result. It is explanatory drawing which shows the structure of the referee ability evaluation system in 2nd Example. It is a flowchart of a referee ability evaluation process.

Hereinafter, an example of an athletic motion evaluation system will be described taking as an example a system configured to evaluate a basketball screen play. In other words, the movement evaluation system of the embodiment inputs the movement of a player playing basketball as digital data, extracts a portion corresponding to a play called screen play from a series of movements, and completes the screen play. It is a system that evaluates.
Here, the screen play will be briefly described. In basketball, defensive players often respond one-on-one to attacking players. In such a case, one of the attacking players (this player is referred to as “screener”) approaches another attacking player (this player is referred to as “user”), and the defensive player (this player is referred to as “player”). "User defense") is blocked. By doing so, the user can use the screener to dodge the user defense and play for free. Such a play may be performed as an attacking player who possesses the ball as a user or a player who does not possess the ball as a user so that the user can easily receive a pass. The actual movement of each player varies depending on the positional relationship in the play, but screen play is a standard in basketball in the sense that the screener is close to the user and blocks the movement of the user defense. One of the plays.
In the embodiment, the object of evaluation is limited to screen play, but the present invention can be applied to various other games, various plays and performances.

A. System configuration:
FIG. 1 is an explanatory diagram showing a configuration of an exercise motion evaluation system. The exercise motion evaluation system according to the embodiment is configured by software by installing a computer program for realizing each function shown in FIG. 1 in a computer having a CPU, a RAM, a ROM, and the like. It is also possible to configure a part or all of the illustrated functions as hardware.
In the embodiment, an example in which the system is realized by a stand-alone computer is shown, but a system including a plurality of computers and servers connected via a network may be used.
Hereinafter, each function in the figure will be described.

The movement data input unit 11 has a function of inputting the movement of a player who plays basketball or the ball as digital data (hereinafter referred to as “movement data”). Examples of the motion data to be input include a player position, a moving direction, a moving speed, a ball position, a goal position, and the like. Such data can be generated by, for example, a method in which a player who is playing is photographed with a plurality of cameras around the court and the position of each player or ball is specified. A camera that looks down on the court from directly above may be installed so as to grasp the position of the player two-dimensionally. Since there is a well-known technique already established as a method for acquiring a player's movement or the like as digital data, a detailed description thereof will be omitted.
In this embodiment, the screen play is evaluated, but the input motion data represents not only the scene where the screen play is being performed but also a series of movements during the basketball game.

  In this embodiment, the position of the player is input as movement data. However, when evaluating judo or kendo skills, skating performance, etc., the movement data of the athlete's limbs in addition to the position of the athlete is used. You may make it input as. The motion data to be input in this way can be set as appropriate according to the play or performance to be evaluated.

The motion data storage unit 12 stores the input motion data. It can be constituted by a hard disk or a large-capacity memory.
The action extraction unit 13 extracts from the motion data stored in the motion data storage unit 12 only the scene where the screen play that is the object of motion evaluation is being performed. Each extracted scene is sometimes called an action.
The action data storage unit 14 stores the actions extracted by the action extraction unit 13.

The identification processing unit 15 further narrows down the scenes corresponding to the screen play using the classification-type machine learning for the extracted actions, and updates the contents of the action data storage unit 14 based on the result. In this embodiment, a support vector machine which is one of supervised machine learning is used.
The feature vector storage unit 16 stores a feature vector used in the classification machine learning of the identification processing unit 15. Although the feature vector storage unit 16 is illustrated as a separate configuration from the identification processing unit 15, it is not always necessary to construct a database separate from the identification processing unit 15, and the feature vector storage unit 16 is prepared by being incorporated in the identification processing unit 15. May be.
The evaluation processing unit 17 evaluates the screen play action stored in the action data storage unit 14 by regression machine learning. In this embodiment, support vector regression which is one of supervised machine learning is used. In this embodiment, the feature vectors used in the regression machine learning are the same as those used by the identification processing unit 15, and both are stored in the feature vector storage unit 16. Similar to the identification processing unit 15, the feature vector storage unit 16 may be prepared by being incorporated in the evaluation processing unit 17.

  As described above, in the present embodiment, machine learning is applied in the identification processing unit 15 and the evaluation processing unit 17. In machine learning, in order to achieve sufficient accuracy, it is necessary to have the system learn in advance about screen play identification and evaluation. In this regard, this embodiment uses supervised machine learning, particularly support vector machine and support vector regression.

The training unit 20 performs the above-described learning, that is, the function of performing training. In the training unit 20, the teacher data setting unit 23 sets teacher data for the training action stored in the action data storage unit 14, and stores this in the teacher data storage unit 22.
For example, in the case of categorized machine learning for identifying screen play, the teacher data is data representing a “correct answer” indicating whether the action is screen play. The teacher data can be generated by reproducing and displaying the action to the operator and inputting the “correct answer” for each action by the operator.
In the case of regression machine learning, the teacher data is data representing the degree of completion as a screen play for an action. This data can also be generated by reproducing and displaying the action and inputting an evaluation value by the operator.

  The machine learning training processing unit 21 trains the identification processing unit 15 and the evaluation processing unit 17 using the teacher data. That is, the action stored in the action data storage unit 14 is identified by the identification processing unit 15 as to whether or not it is screen play, and the evaluation processing unit 17 performs the evaluation, and then the teacher data storage unit The process of providing the “correct answer” is repeatedly executed with reference to the teacher data prepared in 22.

  In the example of FIG. 1, the system configuration including the training unit 20 is shown. However, after the training of the identification processing unit 15 and the evaluation processing unit 17 is completed, a configuration in which the training unit 20 is omitted can be employed. That is, after the identification processing unit 15 and the evaluation processing unit 17 that can evaluate new motion data with sufficient accuracy by training are obtained, a system in which the training unit 20 is omitted is manufactured by a method of transplanting the identification processing unit 15 and the evaluation processing unit 17. It may be.

B. Action extraction process:
FIG. 2 is a flowchart of the action extraction process. This is a process of extracting a play to be evaluated from the motion data storage unit 12 stored in the exercise motion evaluation system 10, that is, a screen play in this embodiment, and is a process executed by the action extraction unit 13 in FIG. Applicable. The extraction is performed not by machine learning but by whether or not the player's movement satisfies a predetermined condition (hereinafter also referred to as “criteria”). That is, the action extraction process performed here corresponds to a pre-process for executing a process of cutting out screen play by machine learning later.
This process is performed on the entire motion data by repeatedly executing the process for each scene obtained by dividing all the motion data at a predetermined time interval. The time interval for dividing the scene can be arbitrarily set. In this embodiment, the interval is set to 1 second or less, and the processing is performed at a frequency of 25 to 100 Hz, that is, 0.01 to 0.04 second. did.

  When the process is started, the motion motion evaluation system 10 inputs motion data from the motion data storage unit 12 (step S10). And the location which has shot is extracted (step S11). Whether or not the player is shooting can be identified relatively easily by paying attention to the movement of the ball, such as whether the ball is moving toward the goal after leaving the player.

  Next, the exercise motion evaluation system 10 selects a screener (may be abbreviated as “S” in the figure) (step S12). Then, each candidate of the user (sometimes abbreviated as “U” in the figure) and the user defense (sometimes abbreviated as “UD” in the figure) is selected (step S13). In this embodiment, two attacking players close to the screener are selected as user candidates, and three defense players close to the screener are selected as user defense candidates. The user and user defense candidates are not limited to such conditions, and can be selected based on various conditions. All players other than the screener may be user candidates, and all defense players may be user defense candidates.

The exercise motion evaluation system 10 determines a signal based on criteria using the positional relationship of these three players (step S14). If the scene being determined satisfies all three conditions, it is determined that the scene corresponds to screen play, and the signal is turned on. Otherwise, the signal is turned off. The three conditions for determining are as follows.
Condition (1) “Distance between screener (S) and user defense (UD) ≦ 1.2 m”. This indicates that the screener is approaching the user defense, which is a characteristic positional relationship of screen play.
Condition (2) “User defense (UD) is closest to user (U)”. If a defender near the screener is selected as a user defense candidate, the conditioner (1) may be satisfied regardless of screen play, so the screener approaches the user defense instead of his defense. The characteristic positional relationship of screen play is defined as condition (2).
Condition (3) “Not shot”. This is because screen play is used in scenes that are not shot.

Here, in order to help understanding, an example of signal determination processing is shown.
FIG. 3 is an explanatory diagram showing an example of action extraction processing. In the upper row, the position of each scene is shown schematically with the attacking player as a circle and the defensive player as a triangle. The upper left side represents a scene in which the screener (S) moves in the direction of the arrow s1 and approaches the user defense (UD). In the center, the user (U) moves in the direction of the arrow u1, but the screener (S) blocks the movement of the user defense (UD), so the user defense (UD) turns around as indicated by the arrow ud1. Represents a scene in which the user (U) has to be dealt with. The right side shows a scene where the user (U) and the user defense (UD) move away from the screener (S) and move as indicated by arrows u2 and ud2, respectively. The locus of the user defense (UD) is a locus detoured as indicated by a broken line d, and as a result, represents a state in which the user (U) is chased from behind, that is, a screen play has been successful.

The signal determination results are shown in the second row from the top. For each scene shown in the upper part, it is possible to determine whether the signal is on or off by judging the three conditions described in step S14 of FIG. In FIG. 3, “X” is shown for a signal-on scene, that is, a scene that satisfies all three conditions, and “X” for a scene that does not.
In the scene before the start of the screen play on the left side of the top row, the signal off (×) generally continues, in the center scene during the screen play, the signal on (○) generally continues, and in the right scene after the screen play, the signal is approximately The result is an off (x) scene. However, in the scene before the start of screen play in the figure, the signal ON (O) appears only at one place, although the signal OFF (X) continues as "XX OX ...". As described above, since the positional relationship of the players included in the motion data includes errors and noise, it may occur that a signal determination different from the preceding and following scenes is made in each scene for a very short period of time.

Returning to FIG. 2, the next process will be described. When the signal determination is finished, the exercise motion evaluation system 10 executes a noise removal process (step S15). This is a process of excluding those that satisfy any of the following conditions from the signal-on state as inappropriate.
Condition (1) “Signal length ≦ 0.1 seconds”. This is because an extremely short signal is judged as noise.
Condition (2) “Circumstance distance = 0 m”. The circuitous distance refers to a moving distance that the user defense (UD) is detoured by the movement of the screener (S). If the circuitous distance is 0 m, it cannot be determined that the screen play has been performed, and thus those that meet such a condition are excluded.

The noise removal process will be described with reference to FIG. The result of noise removal processing is shown in the third row from the top. Assuming that the interval between scenes is about 0.1 seconds, in the first signal determination (second stage from the top), the portion where two or more signal on (O) does not continue is the condition (1 ). As a result, the signal ON “◯” in the “XX ○ ×” part at the beginning of the second stage and the “XXX XX” part at the end is judged as noise. , Signal off “x” is replaced. The time interval that is the criterion for the condition (1) can be arbitrarily set.
The condition (2) can be determined based on the locus d of the user defense (UD) shown on the upper right side. How far the user defense (UD) makes a detour with the screener (S) present against the trajectory of the user defense (UD) moving to follow the user (U) in the absence of the screener (S) You just have to judge whether The determination of the condition (2) can also be made by an arbitrary method. For example, it may be determined whether it is a detour based on the degree of deformation of the movement trajectory based on the presence or absence of the screener (S), not the movement distance of the user defense (UD). Further, the condition (2) may be excluded from the noise removal process.

Returning to FIG. 2, the next process will be described. When the noise removal process is finished, the exercise motion evaluation system 10 executes the fusion process (step S16). The fusion process is a process of fusing the divided scenes of the signal-on scenes as a series of plays. In this embodiment, those satisfying the following three conditions are fused.
Condition (1) “screener (S), user (U), and user defense (UD) are the same”. This is because if the players involved in the screen play are different, it is determined that they are different.
Condition (2) “interval between segments ≦ 3 seconds”. A segment refers to a group of segments in which signal-on is continued for a certain period. When the interval between the segments is large, it is determined that the play is different. The interval of “3 seconds” that serves as a determination criterion can be arbitrarily set.
Condition (3) “A chute is not sandwiched”. This is because the play performed before and after the shot is considered as another play.

  The fusion process will be described with reference to FIG. As a result of noise removal, segment 1 and segment 2 in which signal-on continues are found as shown in the figure. Between the segments 1 and 2, “XX” is shown in the figure, and it can be seen that the period is very short. Therefore, the condition (2) is satisfied. Therefore, if the screener (S), the user (U), and the user defense (UD) of the segments 1 and 2 satisfy the same condition (condition (1)) and do not sandwich the chute (condition (3)), the segment 1 2 are merged into one segment 3.

Returning to FIG. 2, the next process will be described. When the fusion process is finished, the exercise motion evaluation system 10 determines the action start and end times (step S17). The segment obtained by the processing up to the fusion processing (step S16) is obtained by extracting only the scenes that satisfy the signal ON, that is, the conditions described in step S14. However, in order to determine whether or not it corresponds to screen play and to evaluate the completeness of play, it is necessary to determine not only the scene satisfying the positional relationship shown in step S14 but also the movement before and after that. There is. Therefore, in the process of step S17, a fixed period before and after the start of screen play is added before and after the segment, and this series of sections is set as a fixed period to be evaluated for screen play. This fixed period is called an action.
That is, in this embodiment, the action start time is set 0.5 seconds before the start of the signal, and the action end time is set 3.0 seconds after the end of the signal (step S17).

  The action start / end time determination process will be described with reference to FIG. Assume that segment 3 is specified by the fusion processing. The motion motion evaluation system 10 identifies the earliest scene in the segment 3 and sets the action start time 0.5 seconds before that. Also, the latest scene of segment 3 is specified, and the action end time is 3.0 seconds later. As a result, before and after the segment 3, an action including a signal off (×) period is set.

Returning to FIG. 2, the next process will be described. The motion motion evaluation system 10 repeatedly executes the above processing until all the screener candidates are completed (step S18). Through the above processing, an action that is considered to be a screen play is specified from all the plays.
Since the action extraction process described in FIG. 2 is a process for extracting a scene that satisfies a predetermined criterion, there is an advantage that it can be determined more easily than an identification process using machine learning described later. This embodiment has an advantage that screen play can be efficiently identified by executing such action extraction processing before processing by machine learning.
In this example, 140 out of 144 plays judged as screen play by human eyes were extracted as actions. That is, it was confirmed that the processing of this example had an accuracy of 97.2% (140/144). The conditions used in this embodiment can be arbitrarily set, and conditions other than those described in FIG. 2 may be used.

C. Play identification process:
Next, the play identification process will be described. This process is a process for identifying whether or not the screen play is performed using classification machine learning, and is a process executed by the identification processing unit 15 in FIG. In this embodiment, a support vector machine is used as machine learning.
FIG. 4 is a flowchart of the play identification process. When the process is started, the exercise motion evaluation system 10 reads the result of the action extraction process (step S30).
Then, an action to be identified is selected from the read information (step S31), and a feature amount to be used for identification by machine learning is extracted (step S32). Then, play identification by machine learning is executed (step S33). The exercise motion evaluation system 10 executes these processes for all actions (step S34), and outputs an action corresponding to screen play (step S35).

The feature vector used in the above process can be arbitrarily set, but in the present embodiment, the following five types of variables are used as candidates.
(1) The minimum distance from the action start time to the action end time (referred to as “minimum distance”);
(2) The difference between the distance of the action start time and the minimum distance;
(3) Average value of distance from the distance of the action start time to the minimum distance;
(4) Difference between the minimum distance and the distance at the action end time;
(5) Average value of the distance to reach the minimum distance and the distance of the action end time;
Each distance is calculated among the screener (S), the user (U) candidates (two people), and the user defense (UD). Since the screen play may be performed at a place unrelated to the ball and the goal, the distance from the ball and the goal is not considered in the feature vector.

In this example, machine learning was performed with all combinations of the above five variables, that is, (2 ⁵ −1) combinations, and the results were evaluated. A method for evaluating the grade will be described later. Then, the combination of variables (1), (3), and (4) that had the best results was selected.
And (6) the distance traveled from the start time to the end time of the action;
Was added as a variable.
As a result, in this embodiment, the variables (1), (3), (4), and (6) are used as feature vectors.
In the present embodiment, it is possible to identify the screen play with high accuracy by using the machine learning based on the feature vector described above.

D. Play evaluation process:
Next, the play evaluation process will be described. This process is a process for evaluating the degree of completion as a screen play for the action identified as the screen play by the screen identification process up to FIG. 4, and is a process executed by the evaluation processing unit 17 in FIG. For the evaluation of this embodiment, regression machine learning is used, and in this embodiment, support vector regression, which is one of them, is used. That is, by giving the evaluation results for some screen plays in advance as teacher data, the exercise motion evaluation system 10 is trained, and based on the training results, the degree of completion for the new screen play is indicated by a score or the like. Evaluate.

FIG. 5 is a flowchart of the play evaluation process. When the processing is started, the exercise motion evaluation system 10 reads an action identified as screen play (step S40), and calculates a feature amount (step S41). The feature amount used here can be arbitrarily set, but in this embodiment, it is the same as that used in the screen identification process (FIG. 4). The motor motion evaluation system 10 performs play evaluation using machine learning based on the feature amount (step S42), and outputs the result (step S43).
As described above, according to the play evaluation process of the present embodiment, the screen play can be objectively evaluated by using the support vector regression. In the present embodiment, as described above, a feature vector suitable for screen play identification is found in classification-type machine learning, and this feature vector is also commonly used in regression-type machine learning. This is because it has been found that the regression machine learning can obtain a stable evaluation result. The principle of obtaining such a result is not necessarily clear, but it is considered that one of the reasons is that there is no need to change the parameters used for the analysis in the classification and regression of the movement to be evaluated. In addition, there is an advantage that the accuracy of regression type machine learning can be easily improved by setting the feature vector by classification type machine learning in which pass / fail is easy to judge.

E. Machine learning training process:
As described above, in the present embodiment, a common feature vector is used for the support vector machine that is the classification type machine learning and the support vector regression that is the regression type machine learning. Therefore, training is necessary. Below, the training process is demonstrated.
FIG. 6 is a flowchart of the machine learning training process. This is a process for training the identification processing unit 15 and the evaluation processing unit 17 of the motor motion evaluation system 10, and corresponds to the process executed by the machine learning training processing unit 21 in FIG.
When the process is started, the exercise motion evaluation system 10 reads action data (step S51). What is read here is action data prepared for training. The exercise motion evaluation system 10 also reads teacher data (step S52). Teacher data is data in which responses for each of the action data are recorded. As shown in the figure, the classification-type machine learning teacher data is data that records whether or not each action corresponds to screen play. Regression type machine learning teacher data is data that stores the results of evaluating the degree of completion of screen play with respect to each action in terms of scores or the like.

  Then, the exercise motion evaluation system 10 extracts a feature amount for each action (step S53). The feature vectors are the same as those used in the play identification process (FIG. 4) and the play evaluation process (FIG. 5). When setting a feature vector based on the results of machine learning, the machine learning training process (FIG. 6) is repeatedly executed for various feature vectors. Therefore, in step S53, as an evaluation target at the time of execution. What is necessary is just to extract based on the set feature vector.

  Next, the exercise motion evaluation system 10 divides the action into K ways in order to perform training and evaluation using a method called a cross-validation method. K can be arbitrarily set, for example, it can be divided into ten. The division means that a plurality of existing actions are grouped, and does not mean that one action is divided at time intervals. Further, the number of actions included in each divided group does not need to be exactly the same.

When the action division is completed, the exercise motion evaluation system 10 calculates a cross validation score (step S55).
As shown in the figure, first, the divided group 1 is used as test data, and the remainder after group 2 is used as training data. Then, machine learning training is executed using action data and teacher data included in the training data. In the case of categorized machine learning, teacher data is used to tell whether action data corresponds to screen play. When the training is completed, the exercise motion evaluation system 10 identifies whether or not the screen play is for the group 1 prepared as test data, and compares the result with the teacher data for the group 1, The correct answer rate is calculated as a score. In the example in the figure, 0.856 is calculated as the score.
The same processing is executed next using group 2 as test data. In this way, the score is calculated while sequentially replacing the test data from group 1 to group K. And let the average value of the score of the group 1-the group K be a score of a cross-validation.

The cross-validation score described above also varies depending on the value of the limiting parameter C that defines the complexity of the relationship for obtaining the determination result according to the feature vector in machine learning.
Therefore, in this embodiment, the exercise motion evaluation system 10 repeats the calculation of the cross-validation score while changing the value of the restriction parameter C until the cross-validation score obtains the optimum value (step S56).

FIG. 7 is an explanatory view showing an example of evaluation of the limit parameter. The value of the limit parameter C is plotted on the horizontal axis, and the error rate is plotted on the vertical axis. The feature vector is in a constant state and represents the influence of the limiting parameter. As illustrated, when the restriction parameters is 10 ^2, has a error rate minimum, it can be seen the value is restricted parameter optimal. In addition, the cross-validation score obtained with this restriction parameter is the result for the feature vector being used.

If the above machine learning training process is repeatedly executed for various feature vectors, a result for each feature vector can be obtained. Then, the best feature vector can be set by selecting the best one from among them.
In the present embodiment, through such processing, the feature vector in the classification type machine learning is determined as the content described in FIG.

The machine learning training process of FIG. 6 does not always have to execute all the processes. When a feature vector that is considered to be optimal has already been obtained, it is sufficient to perform it based on the feature vector. If the optimum value of the restriction parameter C is also obtained for the feature vector, the process for calculating the cross-validation score in step S55 using the restriction parameter C, or using all groups as training data instead. You may be made to perform the training processing which was.
Further, if the identification processing unit 15, the feature vector storage unit 16, and the evaluation processing unit 17 in a state where the training is completed can be transplanted, the training unit 20 in FIG. 1 is omitted, and the machine learning training process in FIG. 6 is omitted. Thus, the exercise motion evaluation system 10 may be constructed.

FIG. 8 is an explanatory diagram illustrating an example of evaluation of regression results. FIG. 8A shows the generation of regression machine learning teacher data based on coach 1's evaluation for screen play, machine learning training, and machine learning evaluation for a new screen play ( The graph shows the relationship between the vertical axis) and the evaluation by coach 1 (horizontal axis). “C = 10” shown in the figure represents the value of the limiting parameter. Further, “R ² = 0.99” is a determination coefficient, and the closer to value 1, the higher the degree of coincidence between the two. That is, FIG. 8A shows that the exercise motion evaluation system 10 accurately reproduces the same evaluation as the coach 1 with respect to the screen play.
8 (b) to 8 (d) show the results of training and evaluation in the same manner for different coaches 2 to 4, respectively. Different coaches use different teacher data for regression type machine learning, so that the limiting parameters that are optimum values also have different values. That is, the exercise motion evaluation system 10 is trained reflecting the individuality of each coach. However, in any case, the coefficient of determination shows a value almost close to 1, and it can be seen that the motion motion evaluation system 10 reproduces the evaluation of each coach with very high accuracy.

  As described above, according to the present embodiment, it is possible to realize accurate evaluation reflecting the individuality of each coach. In this embodiment, basketball screen play is targeted, but this embodiment can be applied to various games and evaluation of various plays and performances. In addition to strengthening and instructing players during practice, for example, scenes used for performance evaluation in competitions where performances are scored, such as figure skating, can also be considered.

Next, a second embodiment of the present invention will be described. The second embodiment shows a configuration example as a referee ability evaluation system for referee evaluation and the like, not as a system for strengthening / instructing players. The referee ability evaluation system is a system that evaluates the referee's ability on the assumption that the evaluation by the machine learning is a correct answer by comparing the evaluation by the referee for play and performance with the evaluation by the machine learning. It can be used for umpire education / training, and ability evaluation such as whether or not a umpire has the ability to judge international matches.
FIG. 9 is an explanatory diagram showing the configuration of the referee ability evaluation system 10A in the second embodiment. Similarly to the first embodiment, the referee ability evaluation system 10A can also be configured in software by installing a computer program for realizing each function shown in the computer. Of course, some or all of the functions may be configured in hardware.

Among the functions shown in the figure, the motion data input unit 11, the motion data storage unit 12, the action extraction unit 13, the action data storage unit 14, the identification processing unit 15, the feature vector storage unit 16, and the evaluation processing unit 17 are respectively This is the same as in the first embodiment.
The action data reproduction unit 30 has a function of reproducing the action stored in the action data storage unit 14 as a moving image and displaying it on a display D of a personal computer.
The response input unit 31 inputs the referee's response using the keyboard K, mouse M, etc. of the personal computer.
And the pass / fail determination part 32 compares the evaluation result by the machine learning with respect to an action, and the reply of a referee, and performs pass / fail determination.

FIG. 10 is a flowchart of the referee ability evaluation process. When this process is started, the referee ability evaluation system 10A reads an action (step S60). The action used here may be an action prepared for referee training in advance, or may be newly extracted from a player's play or performance. The action can be cut out in the same manner as in the first embodiment.
The referee ability evaluation system 10A reproduces the read action on a computer display or the like (step S61), and inputs an answer to the action (step S62). The answer is the result of the referee evaluating the action.

On the other hand, the referee ability evaluation system 10A extracts a feature amount for the action (step S63) and obtains a reference answer using machine learning (step S64). These processes are the same as the play evaluation process (FIG. 5). That is, an action evaluation result is obtained by support vector regression.
However, as shown in FIG. 8 in the first embodiment, in the case of evaluation by machine learning using support vector regression, if the personality of the coach who created the teacher data is reflected, it can be used as a reference answer Can not. Therefore, in the second embodiment, evaluation results obtained by a plurality of referees who have already been recognized to have a certain ability are used as teacher data for regression machine learning. This makes it possible to create standard teacher data that averages the differences in ability and individuality of each referee, and the evaluation result by support vector regression can be used as a reference answer.

Next, the referee ability evaluation system 10A compares the reference answer with the inputted answer, determines pass / fail (step S65), and outputs the result (step S66). In the pass / fail judgment, if the difference between the reference answer and the answer is within a predetermined range, it is possible to arbitrarily set a reference such as passing. Moreover, you may make it evaluate based not only on the evaluation based on one action but on several actions.
As described above, according to the second embodiment, it is possible to objectively realize the referee's ability evaluation by using regression machine learning.

The first embodiment and the second embodiment have been described above with respect to the present invention. According to these embodiments, the present invention can objectively and accurately realize identification of play and evaluation of completeness in various competitions. Further, as shown in the second embodiment, the present invention can also objectively and accurately evaluate the referee's ability.
The various features described in these embodiments are not necessarily all provided, and the present invention can be implemented by omitting or combining some of them as appropriate. Further, the present invention can also realize various modifications.

  The present invention can be used to evaluate the completeness of a specific play or performance regarding the movement of a player during exercise.

DESCRIPTION OF SYMBOLS 10, 10A ... Exercise motion evaluation system 11 ... Motion data input part 12 ... Motion data storage part 13 ... Action extraction part 14 ... Action data storage part 15 ... Identification processing part 16 ... Feature vector storage part 17 ... Evaluation processing part 20 ... Training Unit 21 ... Machine learning training processing unit 22 ... Educational data storage unit 23 ... Teacher data setting unit 30 ... Action data reproduction unit 31 ... Answer input unit 32 ... Pass / fail judgment unit

Claims (8)

An exercise motion evaluation system for evaluating the completeness of a specific play or performance with respect to the movement of a player during exercise,
An operation data input unit for reading operation data representing the movement of the player;
An exercise motion evaluation system comprising: an evaluation unit that performs the evaluation by regression type machine learning using a feature vector set by classification type machine learning for the play or performance based on the motion data. The motor motion evaluation system according to claim 1,
The motion data is data representing a series of movements including play or performance to be evaluated,
Prior to the evaluation by the evaluation unit, an athletic motion evaluation system including a segmentation unit that extracts, from the motion data, a part corresponding to the play or performance to be evaluated by classification machine learning. It is an exercise | movement movement evaluation system of Claim 2, Comprising:
Prior to the classification type machine learning, the segmenting unit extracts, from the motion data, a section where the movement of the player satisfies a predetermined condition set according to the play or performance, and the extracted section An exercise motion evaluation system that applies the classification type machine learning as a target. It is an exercise | movement movement evaluation system of Claim 3, Comprising:
When the plurality of extracted sections are determined as a series of sections, the segmenting unit further combines the plurality of sections into one section and then applies the classification type machine learning. Evaluation system. It is an exercise | movement movement evaluation system in any one of Claims 1-4,
The feature vector is an exercise motion evaluation system in which a plurality of types of feature vectors and combinations thereof are subjected to the classification-type machine learning and set based on the evaluation of the result. It is an exercise | movement motion evaluation system in any one of Claims 1-5,
The play is a screen play in basketball,
The motion vector evaluation system is set based on the positional relationship of four players related to the screen play without considering the positional relationship with the basketball goal. An exercise motion evaluation method for evaluating the degree of perfection as a specific play or performance with respect to the movement of a player during exercise,
As the steps executed by the computer,
Reading motion data representing the movement of the player;
And a step of performing the evaluation by regression type machine learning using a feature vector set by classification type machine learning for the play or performance based on the action data. A computer program for evaluating by a computer the completeness of a specific play or performance with respect to the movement of an athlete during exercise,
A function of reading motion data representing the movement of the player;
A function of performing classification-type machine learning for the play or performance based on the motion data;
A computer program for realizing, by a computer, the function of performing the evaluation by regression type machine learning using a feature vector set by the classification type machine learning.

Images