JP7740693B2 - Tactical analysis device, control method thereof, and control program - Google Patents

Description

The present invention relates to a tactical analysis device that analyzes a player's tactics in a competition, as well as a control method and control program for the same.

In recent years, computer-based analysis has become increasingly common in the field of competitive sports. For example, Patent Document 1 discloses an attack pattern extraction device that analyzes attack patterns based on player position information and play history in team sports.

Japanese Patent Application Laid-Open No. 2017-209237 Japanese Patent Application Laid-Open No. 2020-185370

However, Patent Document 1 only identifies past tactics (offensive patterns) and does not predict future tactics from the player's tactical history during the game. Furthermore, during baseball broadcasts, commentators sometimes predict future pitching moves based on the pitching history during the game, but this is based on the commentator's many years of experience and is not done automatically using a computer.

One aspect of the present invention aims to realize a tactical analysis device that can predict a player's future tactics during a game.

In order to solve the above problems, one aspect of the present invention provides a tactical analysis device that analyzes a player's tactics in a competition, and includes a tactical acquisition unit that acquires the player's tactical history in the competition from video of the competition, a learning model of the player in which a strategy, which is a time-series pattern of tactics, is machine-learned using the player's tactical history in past competitions, and a strategy estimation unit that estimates the player's strategy in the competition using the tactical history acquired by the tactical acquisition unit.

Here, tactics refer to the characteristic behavior of the player in a certain unit of behavior in the game. For example, in baseball, the unit of behavior may be a single pitch, and the characteristic behavior may be at least one of the course, type of ball, and ball speed of the pitch. In addition, in table tennis, the unit of behavior may be a single rally, and the characteristic behavior may be at least one of the rally trajectory, serve style, position and rotation of the served ball, and position and posture of the player.

With the above configuration, the tactical history of a player in a competition is acquired from video of the competition, and the acquired tactical history and a learning model of the player are used to estimate the player's strategy for the competition. Therefore, a user can refer to the strategy estimated from video of the competition, i.e., the time series pattern of tactics, and thereby predict the player's future tactics during the competition. Note that the user may be the player's opponent or the opponent's coach.

In the tactic analysis device according to this aspect, the tactic acquisition unit may extract time-series data of representative points related to tactics from the video and acquire the player's tactics by performing singular value decomposition on the extracted time-series data. In this case, the player's tactics can be acquired even if the time-series data of the representative points varies in duration.

In the tactical analysis device according to this aspect, the tactical acquisition unit may extract trajectories of moving objects from the video of the game, acquire the player's tactics based on the extracted trajectories of the moving objects, and repeat this process to acquire the player's tactical history.

Here, a moving object refers to an object that moves during the game, such as a player in the game or a ball used in the game. Furthermore, as a preprocessing step, the tactic acquisition unit may remove scenes unrelated to the acquisition of the tactics (such as close-up footage of the player) from the video of the game as unnecessary scenes.

In the tactic analysis device according to this aspect, the moving object may include the player, and the tactic acquisition unit may estimate the trajectory of the player's skeleton by performing machine learning using video of the game, and acquire at least a portion of the player's tactic based on the estimated trajectory of the skeleton. In this case, the player's behavior can be estimated with high accuracy, and therefore the player's tactic can be estimated with high accuracy.

In the tactical analysis device according to this aspect, the learning model may be created by performing machine learning by integrating and combining multiple models. In this case, the accuracy of estimating the strategy can be improved compared to other learning models.

In the tactical analysis device according to this embodiment, the multiple models may be fuzzy clustering models. In this case, the estimated strategy is provided using fuzzy rules (IF THEN rules). This allows the user to easily predict future tactics.

In the tactical analysis device according to this aspect, the learning model may be created by generating virtual data based on estimated data estimated by one of the plurality of models through machine learning, and then using the virtual data and learning data to create another model of the plurality of models through machine learning. In this case, the amount of data available for machine learning can be increased, further improving the accuracy of estimating the strategy.

A control method for a tactical analysis device according to another aspect of the present invention is a control method for a tactical analysis device that analyzes the tactics of a player in a competition, and includes a tactical acquisition step of acquiring the tactical history of the player in the competition from video of the competition, and a strategy estimation step of estimating the player's strategy in the competition using a learning model in which a strategy, which is a time-series pattern of tactics, is machine-learned using the tactical history of the player in past competitions, and the tactical history acquired in the tactical acquisition step.

The above control method can achieve the same effects as the tactical analysis device described above.

The tactical analysis device according to each aspect of the present invention may be realized by a computer. In this case, the control program for the tactical analysis device, which causes the computer to operate as each unit (software element) of the tactical analysis device, thereby realizing the tactical analysis device on the computer, and the computer-readable recording medium on which the program is recorded, also fall within the scope of the present invention.

One aspect of the present invention is that it is possible to predict a player's future tactics during a match.

1 is a block diagram showing a schematic configuration of a tactical analysis device according to an embodiment of the present invention; 10 is a flowchart showing a process flow in a learning mode in the tactic analysis device. 10 is a flowchart showing a processing flow in a competition mode in the tactic analysis device. FIG. 10 is a diagram showing a black-and-white image converted from a color image for broadcast superimposed thereon. 10 is a diagram showing an example of an image including tactics information acquired by a tactics acquisition unit in the tactics analysis device. FIG. 3 is a block diagram showing a schematic configuration of a strategy estimation unit and a tactics history storage unit in the tactics analysis device. FIG. 10 is a flowchart showing a flow of a learning process in the strategy estimation unit. 10 is a diagram showing, in a table format, an example of a tactics history acquired by the tactics acquisition unit. FIG. 10 is a flowchart showing the flow of processing for extracting a ball from a video in a tactical analysis device according to another embodiment of the present invention. FIG. 10 is a diagram showing an example of a frame image processed by the white blog extraction process in the above processing. FIG. 10 is a diagram showing an example of a frame image processed by the inter-frame difference detection step in the above processing.

Embodiments of the present invention will be described in detail below. For ease of explanation, components that have the same functions as those shown in each embodiment will be denoted by the same reference numerals, and their description will be omitted where appropriate.

[Embodiment 1]
An embodiment of the present invention will be described with reference to FIGS.

(Overview of the Tactical Analysis Device)
FIG. 1 is a block diagram showing the schematic configuration of a tactical analysis device according to this embodiment. The tactical analysis device 10 of this embodiment acquires and analyzes the tactics of an opponent (player) from video and estimates the opponent's strategy. The following description is based on the case where the sport is table tennis, but the present invention is not limited to this. Details of the rules of table tennis, etc., are well known, and therefore will not be described here.

As shown in FIG. 1, the tactical analysis device 10 has a control unit 11, a memory unit 12, and a communication unit 13. The control unit 11 comprehensively controls the operation of the various components of the tactical analysis device 10 and is composed of, for example, a computer including a CPU (Central Processing Unit) and memory. The operation of the various components is controlled by having the computer execute a control program. The memory unit 12 records information and is composed of a storage device such as a hard disk or flash memory. The communication unit 13 is composed of a communication device that sends and receives information via a communication network such as the Internet.

The control unit 11 has a tactic acquisition unit 22 and a strategy estimation unit 23. The memory unit 12 also includes a tactic history memory unit 31 that stores tactic history data. In this embodiment, the memory unit 12 includes a learning mode TM that performs machine learning on the learning model of the strategy estimation unit 23, and a competition mode PM that uses the machine-learned learning model during a competition.

The tactical acquisition unit 22 acquires data on the opponent's tactical history (chronological data on tactics) from the video data acquired via the communication unit 13. The tactical acquisition unit 22 stores the acquired tactical history data in the tactical history storage unit 31.

Specifically, the tactical acquisition unit 22 processes the video to remove unnecessary scenes, and from the processed video, identifies the position and rotation of the ball (moving object), the position and posture of the player (moving object), etc., and uses these as tactical information. Next, the tactical acquisition unit 22 acquires this tactical information for each rally, and creates a tactical history from multiple pieces of tactical information for multiple rallies. The tactical acquisition unit 22 then stores the acquired data on the opponent's tactical history in the tactical history storage unit 31.

The above video can also be broadcast video. In this case, there is no need to use a special camera system. The above video can be video taken with a single camera, or multiple videos taken with multiple cameras.

The strategy estimation unit 23 estimates the opponent's strategy using tactical history data stored in the tactical history memory unit 31. Specifically, the strategy estimation unit 23 has a learning model, and in learning mode TM, machine learning is performed on the learning model using the opponent's past tactical history. On the other hand, in competition mode PM, the strategy estimation unit 23 estimates the opponent's strategy using the opponent's tactical history during the competition and the machine-learned learning model. The strategy estimation unit 23 outputs data on the estimated strategy to an external device such as the user's terminal via the communication unit 13.

(Processing in learning mode)
2 is a flowchart showing the flow of processing in the learning mode TM in the tactic analysis device 10 configured as described above. The processing shown in FIG. 2 is executed for each player.

As shown in FIG. 2, first, the tactics acquisition unit 22 collects video of a player's past games (S11). Next, the tactics acquisition unit 22 collects the player's past tactics history from the collected video and stores it in the tactics history storage unit 31 (S12).

Next, the strategy estimation unit 23 performs machine learning on the learning model using the past tactical history stored in the tactical history storage unit 31 (S13). Note that the strategy estimation unit 23 may estimate the player's strategy using a portion of the past tactical history and the machine-learned learning model, and output data on the estimated strategy to an external device via the communication unit 13. Thereafter, processing in the learning mode TM ends.

Figure 3 is a flowchart showing the processing flow in competition mode PM in the tactical analysis device 10 configured as described above. The processing shown in Figure 3 is executed during the competition.

As shown in FIG. 3, first, the tactics acquisition unit 22 acquires video of the opponent (player) during the competition (S21). Next, the tactics acquisition unit 22 acquires the opponent's tactics history during the competition from the acquired video and stores it in the tactics history storage unit 31 (S22, tactics acquisition step).

Next, the strategy estimation unit 23 estimates the opponent's strategy for the competition using the tactical history during the competition stored in the tactical history storage unit 31 and the opponent's learning model obtained by machine learning in step S13 of FIG. 2 (S23, strategy estimation step). The strategy estimation unit 23 then outputs data on the estimated strategy to the opponent's or their coach's terminal (external device) via the communication unit 13 (S24). Thereafter, processing in competition mode PM ends.

The tactical analysis device 10 of this embodiment therefore acquires the tactical history of the opponent in the competition from video footage of the competition, and estimates the opponent's strategy during the competition using the acquired tactical history and a machine-learned learning model of the opponent. By referring to the estimated strategy, i.e., the time series pattern of the tactics, the opponent or their coach can predict the opponent's future tactics.

(Additional Notes)
The strategy estimation unit 23 may have the learning model previously trained by machine learning. In this case, the learning mode can be omitted.

(Details of the Tactical Acquisition Department)
Next, the tactics acquisition unit 22 and the strategy estimation unit 23 will be described in detail.

The tactical acquisition unit 22 first removes unnecessary scenes such as commercials, close-ups of players, and slow-motion playback images from the 30 fps broadcast video, and then divides the video into individual rally videos. Next, the tactical acquisition unit 22 converts the color video into black and white video.

Figure 4 shows the black-and-white video superimposed on the image. As can be seen from Figure 4, the tactical acquisition unit 22 can identify the positions of moving objects, i.e., the ball and the player, from the black-and-white video and acquire the trajectories of the ball and the player. The tactical acquisition unit 22 can also acquire the player's posture from the black-and-white video. The tactical acquisition unit 22 can also acquire the coordinates of the four corners of the table tennis table from the black-and-white video. The tactical acquisition unit 22 can also identify the game number, rally number, and the score immediately before the current rally from the upper left corner of the black-and-white video.

Next, the tactic acquisition unit 22 performs an affine transformation using the coordinates of the four corners of the table to convert the trajectories of the ball and player from coordinates on the image to coordinates on a plane when the table is viewed from above. Next, the tactic acquisition unit 22 identifies the position where the ball landed on the table (landing point) from the ball's trajectory. The rally trajectory is acquired by connecting the identified positions in chronological order with straight lines.

The tactic acquisition unit 22 also identifies the serve style and ball spin from the trajectories of the ball and the player, and the player's posture. Examples of serve styles include forehand, backhand, hook, etc. Examples of ball spin in the serve include right side spin, left side spin, topspin, underspin, combinations of these, and knuckle (no spin). Note that the hook refers to a serving technique in which the ball is thrown up, the racket is raised toward the falling ball, and the ball is brushed against the racket with either side of the racket.

The tactical acquisition unit 22 then acquires the rally trajectory, serve style, and serve ball spin as tactical information for each rally, and stores multiple pieces of tactical information from multiple rallies as a tactical history. In learning mode TM, the tactical acquisition unit 22 stores multiple pieces of tactical history data from multiple past matches in the tactical history memory unit 31 as learning data, which will be described later. On the other hand, in competition mode PM, data on tactical history during a competition (e.g., the tactical history of one game) is stored in the tactical history memory unit 31 as evaluation data, which will be described later. In the embodiment, these processes were realized using Anaconda Python 3.5.2, OpenCV 3.1.0, easygui, and tqdm programs.

Figure 5 shows an example of an image including the tactical information. As shown in Figure 5, the left part of the image includes a frame showing the table and a rally trajectory, and the center right part of the image includes a legend for the rally trajectory. In the image, the serve rally trajectory is shown with a solid line, the final rally trajectory is shown with a two-dot chain line, and the rally trajectory between the serve rally trajectory and the final rally trajectory is shown with a dashed dot line. The ball spin for the serve is included below the rally trajectory. In the example of Figure 5, the ball spin for the serve is clockwise side back spin.

The top right corner of the image above contains, in order from top to bottom, the server's name, the receiver's name, the game number, and the rally number. The bottom right corner of the image also contains, in order from top to bottom, the score before the rally, the serve style, and the ball rotation of the serve.

(Details of the Strategy Estimation Unit)
6 is a block diagram showing a schematic configuration of the strategy estimation unit 23 and the tactical history storage unit 31. As shown in FIG. 6, the strategy estimation unit 23 includes a learning device group 41, a result acquisition unit 42, a virtual data generation unit 43, and a final decision unit 44. The tactical history storage unit 31 includes a learning data storage unit 51, an evaluation data storage unit 52, and a virtual data storage unit 53.

The learning data storage unit 51 stores data on past tactical history as learning data TRD. The evaluation data storage unit 52 stores data on tactical history during a competition as evaluation data CHD.

In this embodiment, an ensemble learning fuzzy clustering model is used as the learning model. Therefore, the learning device group 41 includes n learning devices (n is an integer greater than or equal to 2) that perform machine learning. Each learning device uses a fuzzy clustering model as its learning model. Fuzzy clustering classifies an input data set into classes using if-then rules and estimates the boundaries of those classes.

In learning mode TM, the first learner 41a performs learning using a set of training data TRD from the training data storage unit 51 and outputs a set of estimated classes as a result. The other learner 41b performs learning using a portion of the set of training data TRD from the training data storage unit 51 and a set of virtual data VD from the virtual data storage unit 53 and outputs a set of estimated classes as a result.

Furthermore, in the competition mode PM, all learned learning devices 41a and 41b estimate classes from the set of evaluation data CHD from the evaluation data storage unit 52 and output the set of estimated classes as a result.

The result acquisition unit 42 acquires the set of classes estimated by each of the learners 41a and 41b in the group of learners 41. The result acquisition unit 42 sends the set of estimated classes to the virtual data generation unit 43 in the learning mode, and to the final determination unit 44 in the competition mode.

The virtual data generation unit 43 generates a set of virtual data VD using a set of training data TRD from the training data storage unit 51 and a set of estimated classes from the result acquisition unit 42. The virtual data generation unit 43 stores the generated virtual data in the virtual data storage unit 53.

Specifically, the virtual data generation unit 43 first compares the actual class contained in each piece of training data TRD with the estimated class, and identifies whether the class is correct. Next, the virtual data generation unit 43 generates a set of virtual data VD based on at least a portion of the incorrectly estimated training data TRD (incorrectly estimated data) and the correctly estimated training data TRD (correctly estimated data), and stores the generated virtual data in the virtual data storage unit 53.

The final determination unit 44 finally determines the most common class from the set of classes estimated by each learning device 41a and 41b for each piece of evaluation data CHD as the estimated class. The final determination unit 44 transmits the set of classes estimated for the set of evaluation data CHD to an external device via the communication unit 13.

~ (learning process)
7 is a flowchart showing the flow of the learning process in the above-configured strategy estimation unit 23. As shown in Fig. 7, first, a variable i is initialized to 1 (S31), and the i-th learner (first learner 41a) performs learning using a set of training data TRD from the training data storage unit 51 (S32).

Next, the virtual data generation unit 43 compares the class estimated by the i-th learner with the actual class for each piece of training data TRD to determine whether the class is correct (S33). Next, if the proportion of correctly estimated classes is equal to or greater than the threshold (YES in S34), the learning process ends.

On the other hand, if the proportion of correctly estimated data is less than the threshold (NO in S34), the virtual data generation unit 43 sets the region in the multidimensional coordinate system (space) that contains the incorrectly estimated data as the (i+1)th low estimation region (S35). Next, the virtual data generation unit 43 generates a set of virtual data VD based on the set of incorrectly estimated data and correctly estimated data located in the (i+1)th low estimation region, and stores the set in the virtual data storage unit 53 (S36).

Next, the (i+1)th learner (another learner 41b) performs learning using the set of incorrectly estimated data and correctly estimated data located in the (i+1)th low estimation region and the set of virtual data VD (S37).

Next, the variable i is incremented by 1 (S38). If the variable i is equal to or less than the predetermined number of times (NO in S39), the process returns to step S12 and the above operation is repeated. If the variable i exceeds the predetermined number of times (YES in S39), the learning process ends.

(Example)
FIG. 8 is a diagram showing, in table form, an example of a tactical history acquired by the tactical acquisition unit 22. The tactical history shown in FIG. 8 includes tactical information for the first to fourth games when the opponent was the server and points were acquired. That is, the portions of FIG. 8 where no tactical information is listed correspond to the cases when the opponent was the server and points were acquired. In addition, in the example of FIG. 8, the tactical information includes the serving style (forehand, backhand, etc.), ball spin, and attack type.

The strategy estimation unit 23 performs a learning process using the tactical history shown in Figure 8 as the learning data TRD, and the following is an example of an estimated set of classes.

If the game number is Small and the rally number is Small, then the (forehand + right side down) serve is Large.
If Game Number is Small and Rally Number is Small, then (N or V) Attack is Large.
If the rally number is Small, then (N or V) the attack is Large.
If Rally Number is Large, then (A or V) Attack is Large.
If Game Number is Large, then Serve Type is Large.
Next, it is assumed that the tactic acquisition unit 22 acquires the tactic histories of the first game and the second game in Fig. 8 from the video of the first game and the second game. Then, it is assumed that the strategy estimation unit 23 estimates the set of classes using the tactic histories of the first game and the second game in Fig. 8 as a set of evaluation data CHD.

In this case, if the opponent or their coach refers to the sets from the above class at the end of the second game, they can predict the opponent's strategy from the third game onwards as follows (a) and (b) below.

(a) In the first half of a rally, players tend to use N or V attacks, and in the second half of a rally, players tend to use A or V attacks.

(b) As the game progresses, the variety of serves increases.

The above (a) is an in-game strategy, and the above (b) is an inter-game strategy. In this way, since the above in-game strategy and the above inter-game strategy can be predicted, it is possible to accurately predict the opponent's future strategy in this match.

Therefore, in the learning unit group 41 of the tactical analysis device 10 of this embodiment, machine learning is performed by integrating multiple learning units 41a and 41b, which allows for improved strategy estimation accuracy compared to other learning models. Furthermore, each of the multiple learning units 41a and 41b uses fuzzy clustering as a learning model, so the estimated strategy is provided using fuzzy rules (IF THEN rules). Therefore, opponents or their coaches can easily predict future tactics.

In addition, virtual data VD is generated based on estimated data estimated by the (i-1)th learner (i is an integer between 2 and n) that has undergone machine learning, and the i-th learner 41b performs machine learning using the virtual data VD and the training data TRD. This allows for an increase in the amount of data available for machine learning, further improving the accuracy of strategy estimation.

[Embodiment 2]
Another embodiment of the present invention will be described with reference to Figures 9 to 11. The tactic analysis device 10 of this embodiment is different from the tactic analysis device 10 shown in Figures 1 to 8 in the operation of the tactic acquisition unit 22 to acquire the trajectories of the ball and players from video, but is otherwise similar.

Figure 9 is a flowchart showing the process flow for extracting the ball from video in the tactical acquisition unit 22 of the tactical analysis device 10 of this embodiment. In this embodiment, the ball's trajectory is estimated from broadcast video automatically divided into rallies by a white image detection process (S40) for the white ball and an inter-frame difference detection process (S45).

As shown in Figure 9, in the white image detection process (S40), first, threshold processing is performed on each of the RGB color scales to extract the white image (S41).

Next, the player's skeletal position is estimated through CenterNet learning (S42). The input to CenterNet is an MP4 video, and the output from CenterNet is 17 two-dimensional skeletal position coordinates and two two-dimensional coordinates specifying the entire skeletal area. Therefore, in the case of two players, the number of output variables is (17 x 2 + 2 x 2) x 2 = 76. Step S42 enables accurate estimation of the player's behavior, and therefore accurate estimation of the player's tactics.

Then, the white blog of the white ball is detected by erasing the white blogs other than the white ball, such as the player's racket, taping, bib number letters, shoes, etc. (S43).

Figure 10 shows an example of a frame image processed by the white blog extraction process (S40). Referring to Figure 10, it can be seen that several white blog areas have been detected, including not only the player's skeleton and the white ball, but also the white frame of the table tennis table, the white part above the net, and the white part of the score box.

On the other hand, in the inter-frame difference detection process (S45), as shown in FIG. 9, first, the difference in the RGB color scale between the image of the current frame and the image of the seven frames prior is calculated (S46). Next, the same process as step S42 of the white image detection process (S40) is performed (S47). Next, white blogs other than the player's white ball are deleted from the player's skeleton position coordinates estimated in step S47, thereby detecting white blogs outside the skeleton area (S48).

Figure 11 shows an example of a frame image processed by the frame difference detection process (S45). Referring to Figure 11, it can be seen that only the player's skeleton and the white ball are detected, and other white areas, such as the white frame of the table, the white part above the net, and the white score box, have been erased. Therefore, the position of the white ball can be detected by the white image detection process (S40) and the frame difference detection process (S45).

After the white image detection process (S40) and the inter-frame difference detection process (S45), the ball position in the next frame is estimated (S51). This estimation is performed by extracting the white blog of the white ball within a circular area with a radius of 80 pixels from the ball position in the current frame, and limiting the white blog portion (ball candidate group) within an inner area that is 10 pixels smaller than the area of the table. If detection is not possible, the radius is expanded by 20 pixels for each frame, and the estimation is performed again. Then, using a Kalman filter, bicube interpolation, etc., the ball position is interpolated when the ball is hidden by a player's body, etc. (occlusion), and the position coordinates of the ball's trajectory are estimated (S52). Subsequent processing is the same as that of the tactical analysis device 10 shown in Figures 1 to 8.

(Additional Notes)
The trajectories of the ball and players are represented by time-series data of positions (representative points), but the length of the time-series data varies depending on the number of rallies, the elapsed time, etc. Therefore, the tactic acquisition unit 22 may acquire, as tactic information for each rally, singular values calculated by singular value decomposition of the time-series data. In this case, it is possible to acquire the opponent's tactics even if the time-series data has different lengths.

The competition mode PM shown in Figure 3 may also be executed after the match. In this case, the opponent's strategy in the current match can be estimated and analyzed, which can be used in the next match against that opponent.

In addition, while the above embodiment is applied to table tennis, the present invention is not limited to this. The above embodiment can be applied to various competitive sports using a ball, such as tennis, baseball, and soccer. In baseball, the pitching technique corresponds to the tactic. In soccer, the tactic is not determined on a player-by-player basis, but on a team-by-team basis. The above embodiment can also be applied to martial arts such as wrestling and judo. In addition, while the above embodiment estimates the strategy of the opponent (enemy), it may also be possible to estimate the strategy of the opponent (ally).

[Software implementation example]
The functions of the tactical analysis device 10 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (particularly each part included in the control unit 11).

In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The functions described in each of the above embodiments are realized by executing the program using this control device and storage device.

The above program may be stored non-transitory on one or more computer-readable storage media. These storage media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

Furthermore, some or all of the functions of each of the above control blocks can be realized by logic circuits. For example, integrated circuits incorporating logic circuits that function as each of the above control blocks are also included in the scope of the present invention. In addition, the functions of each of the above control blocks can also be realized by, for example, a quantum computer.

The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

According to this aspect of the present invention, the following becomes possible, for example: (1) By analyzing the opponent's past data from broadcast footage, the opponent's next strategy can be predicted, allowing the opponent to know what kind of attack and defense they should use. (2) The results of the first game can be analyzed immediately. This allows the opponent's coach to advise the opponent midway through the second game on what kind of attack they should use in the second game. (3) Since the strategies of top players can be referenced, this can also be used to develop players.

10 Tactics analysis device 11 Control unit 12 Memory unit 13 Communication unit 22 Tactics acquisition unit 23 Strategy estimation unit 31 Tactics history memory unit 41 Learning device group 41a and 41b Learning device 42 Result acquisition unit 43 Virtual data generation unit 44 Final decision unit 51 Learning data memory unit 52 Evaluation data memory unit 53 Virtual data memory unit

Claims (9)

A tactics analysis device that analyzes tactics, which are the action content of a player in a specific action unit in a competition, comprising:
a tactic acquisition unit that acquires a tactic history of the player in the competition from a video of the competition;
A tactic analysis device comprising: a learning model of the player in which a strategy, which is a time-series trend of a plurality of tactics executed by the player since the tactic acquisition unit acquired the tactic history, is machine-learned using the tactic history of the player in the past competition ; and a strategy estimation unit that estimates the player's strategy for the competition using the tactic history acquired by the tactic acquisition unit. The tactic analysis device of claim 1, wherein the tactic acquisition unit extracts time-series data of representative points related to tactics from the video and acquires the player's tactics by performing singular value decomposition on the extracted time-series data. The tactic analysis device described in claim 1 or 2, wherein the tactic acquisition unit extracts trajectories of moving objects from the video of the game, acquires the player's tactics based on the extracted trajectories of the moving objects, and repeats this process to acquire the player's tactic history. the moving object includes the player,
4. The tactic analysis device according to claim 3, wherein the tactic acquisition unit estimates a skeletal trajectory of the player by performing machine learning using a video of the game, and acquires at least a part of the tactic of the player based on the estimated skeletal trajectory. The tactical analysis device described in any one of claims 1 to 4, wherein the learning model is created by performing machine learning by integrally combining multiple models. The tactical analysis device of claim 5, wherein the plurality of models are fuzzy clustering models. The tactical analysis device of claim 5 or 6, wherein the learning model is created by generating virtual data based on estimated data estimated by one of the plurality of models through machine learning, and then using the virtual data and learning data to machine-learn another model of the plurality of models. A control method for a tactics analysis device that analyzes tactics, which are the action content of a player in a specific action unit in a competition, comprising:
a tactic acquisition step of acquiring a tactic history of the player in the competition from a video of the competition;
A control method for a tactic analysis device, comprising: a learning model in which a strategy, which is a time-series trend of a plurality of tactics executed by the player after the tactic acquisition step , is machine-learned using the tactic history of the player in past competitions; and a strategy estimation step in which the strategy of the player in the competition is estimated using the tactic history acquired in the tactic acquisition step. A control program for causing a computer to function as the tactical analysis device described in claim 1, the control program causing the computer to function as the tactical acquisition unit and the strategy estimation unit.