JP6553545B2 - Attack pattern extraction apparatus, attack pattern display apparatus, attack pattern extraction method, attack pattern display method, and attack pattern extraction program - Google Patents

Description

  The present invention relates to an attack pattern extraction device, an attack pattern display device, an attack pattern extraction method, an attack pattern display method, and a program for analyzing an attack pattern based on player position information and play history in a group sports competition.

  Conventionally, in team sports competitions such as soccer and basketball, team attack patterns are determined based on machine learning methods such as topic models using player position information and play history of actions such as player passes, shoots, and dribbling. An analysis method is used (see Non-Patent Document 1 and Non-Patent Document 2). In these methods, using the play history of the player who holds the ball as input data, the attack patterns are classified into a plurality of groups based on "where", "which player", and "what kind of play" is easy. In addition, it is possible to automatically extract a plurality of frequently occurring attack patterns for each team.

  As an example of the strategy to be extracted, it is conceivable that the pass from the player A to the player B is easy to pass to the left side, the player B has many dribbling, and the player B has many shots from the center. In particular, in Non-Patent Document 2, when a player is playing according to a certain strategy, the probability indicating the ease of passing between the players, and the probability indicating where the ball recipient is likely to appear on the court Is modeled. A strategy can be automatically extracted by estimating these probabilities based on the observed play history and analyzing the similarity of each attack.

P. Lucey, D. Oliver, P. Carr, J. Roth, I. Matthews, “Assessing Team Strategy using Spatiotemporal Data,” In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Pages. , 2013. Q. Wang, H. Zhu, W. Hu, Z. Shen, and Y. Yao, “Discerning Tactical Patterns for Professional Soccer Teams: An Enhanced Topic Model with Applications,” In Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Pages. 2197-2206, 2015. Y. Cheng, Mean shift, mode seeking, and clustering, In IEEE trans.Pattern Anal. And Machine Intell., 17 (8), pp. 790-799, 1995. Kevin P. Murphy, Machine Learning: A Probabilistic Perspective, The MIT Press, pp. 387-391, 2012. Kevin P. Murphy, Machine Learning: A Probabilistic Perspective, The MIT Press, pp. 352.

  In the above-described conventional technology, similarities of each attack are analyzed based on the position information of the player holding the ball and the pass history information between the players for soccer, basketball, etc., and a plurality of attack patterns are automatically detected. Extraction is realized.

  However, the attack pattern does not necessarily depend only on the player holding the ball, but may also depend on the movement of the player not holding the ball. For example, in soccer, it may be considered that "a player who does not hold the ball moves to attract the defender, thereby creating a space for shooting."

  In the above-described prior art, as in the above-described example, there is a problem that it is not possible to extract an attack pattern in which the movement of players other than the ball holder is important. This is because in the conventional technology, only the position information of the ball holder is handled independently with respect to time, and it is a technical problem that the position information of the player who does not hold the ball cannot be handled. is there.

  In the above-described conventional technology, when quantizing player position information, the court is divided into a predetermined number of meshes, and the position information is quantized based on which mesh the player is in. The method was used. However, this method has a problem that the number of mesh divisions must be appropriately determined according to the target sport. Further, in the prior art, when the position information is quantized, a method of equally dividing the coat is used. However, depending on the target sport, there is a bias in the position where players are likely to gather. For example, in the case of basketball, it is easy for players to gather at a position having an inclination of 45 degrees with respect to the goal. As described above, in the prior art, there is a problem that the characteristics of each type of sport are not taken into consideration, and flexible and automatic quantization can not be performed.

  The present invention has been made in view of the circumstances as described above, and an attack pattern extraction device capable of extracting a more accurate attack pattern in consideration of the movements of all players participating in sports matches, An object is to provide an attack pattern display device, an attack pattern extraction method, an attack pattern display method, and a program.

  In order to achieve the above object, the attack pattern extraction device of the present invention quantizes the position information of each player in the plurality of attack opportunities based on the position history representing the position information of each player in the plurality of attack opportunities. The player moves between the quantized position information in each of the plurality of attack opportunities using the position information quantization unit to be integrated and the position information quantized by the position information quantization unit For each of the plurality of attack opportunities, the player's movement and action are generated from the generated number vector representing the number of times and the generated number vector and the play history representing the action of each player in the plurality of attack opportunities A feature vector generation unit for generating a feature vector; and the feature vector generation unit for each of the plurality of attack opportunities generated by the feature vector generation unit. Based on the vector, it comprises a plurality of attack opportunities, and attack pattern classifying unit for classifying a plurality of attack patterns, the.

  The position information quantization unit causes the position information of each player to correspond to the position at which the density function representing the density of the player at each position, which is obtained using the mean shift clustering method, is a maximum value. Thus, the position information may be quantized.

  Further, the feature vector generation unit generates a connected vector connecting the frequency vector and a vector representing the number of times of each action generated from the play history for each of the plurality of attack opportunities, The feature vector for each of the plurality of attack opportunities may be generated by performing feature conversion on the concatenated vector for each of the plurality of attack opportunities using a feature conversion method that projects to a low-dimensional space. good.

  The attack pattern classification unit may classify the plurality of attack opportunities into a plurality of attack patterns using a clustering method.

  In order to achieve the above object, according to the attack pattern display device of the present invention, an attack pattern extraction device, a reception unit for receiving an input of an attack pattern number of an attack pattern to be output, and the attack pattern extraction device Heat representing the tendency of the action based on the play history at the attack opportunity or the quantized position information at the attack opportunity classified into the attack pattern indicated by the attack pattern number accepted by the unit An attack pattern generation unit that generates at least one of a map and a tendency of movement between the quantized position information of each player.

  In order to achieve the above object, an attack pattern extraction method of the present invention is an attack pattern extraction method of an attack pattern extraction device including a position information quantization unit, a feature vector generation unit, and an attack pattern classification unit, A position information quantization unit quantizing position information of each player in the plurality of attack opportunities based on position history representing position information of each player in the plurality of attack opportunities; and the feature vector generation unit Using the position information quantized by the position information quantization unit to generate a number vector representing the number of times the player has moved between the quantized position information in each of the plurality of attack opportunities Each of the plurality of attack opportunities from the generated frequency vector and a play history representing the action of each player in the plurality of attack opportunities. Generating a feature vector related to player movement and action, and the attack pattern classification unit generates the plurality of the plurality of attack opportunities based on the feature vector for each of the plurality of attack opportunities generated by the feature vector generation unit. And classifying the attack opportunities into a plurality of attack patterns.

  In order to achieve the above object, an attack pattern extraction method of the present invention is an attack pattern display method of an attack pattern display device comprising an attack pattern extraction device, a reception unit, and an attack pattern generation unit, wherein the reception unit Receiving an attack pattern number of an attack pattern to be output, and the attack pattern generation unit is classified by the attack pattern extraction device into an attack pattern indicated by the attack pattern number received by the reception unit. Further, based on the play history at the attack opportunity, or the quantized position information at the attack opportunity, a heat map representing the action tendency, and the quantized position information of each player Generating at least one of the movement tendencies between.

  In order to achieve the above object, a program of the present invention is a program for causing a computer to function as each part of the attack pattern extraction device or each part of the attack pattern display device.

  According to the present invention, it is possible to extract a more accurate attack pattern in which the movements of all players participating in a sport match are considered.

It is a block diagram which shows an example of a structure of the attack pattern extraction apparatus which concerns on embodiment. It is a schematic diagram which shows an example of a structure of the data stored in the player position log | history storage part which concerns on embodiment. It is a schematic diagram which shows an example of a structure of the data stored in the play log | history storage part which concerns on embodiment. It is a schematic diagram which shows an example of the representative point extracted by the mean shift clustering method in the attack pattern extraction apparatus which concerns on embodiment. It is a flowchart which shows the flow of the extraction process performed by the attack pattern extraction apparatus which concerns on embodiment. It is a flowchart which shows the flow of the 1st calculation process performed by the attack pattern extraction apparatus which concerns on embodiment. It is a flowchart which shows the flow of the 2nd calculation process performed by the attack pattern extraction apparatus which concerns on embodiment. It is a schematic diagram which shows an example of the output screen output from the output part which concerns on embodiment.

  Hereinafter, embodiments of an attack pattern extraction apparatus according to the present invention will be described with reference to the drawings.

  Note that the attack pattern extraction apparatus according to the present embodiment extracts attack patterns for various team sports, particularly ball games (soccer, basketball, etc.), and can be applied flexibly according to the type of sport. Moreover, the attack pattern extraction apparatus according to the present embodiment can be flexibly applied to observable action types such as shoots and passes. For example, when basketball is targeted, a play history relating to actions such as shoot, pass, dribble, rebound, and screen is used.

  In this embodiment, a case where a team attack pattern is extracted under the condition that the target is basketball, and in addition to the position information of all players, the play history regarding the pass, dribble, and shoot actions is given. explain.

  As shown in FIG. 1, the attack pattern extraction device 10 according to the present embodiment includes a player position history storage unit 12, a play history storage unit 14, an operation unit 16, a reception unit 18, an attack pattern extraction unit 20, and an attack pattern generation unit. 30 and an output unit 32.

  The player position history storage unit 12 is a database server or the like equipped with a database, and stores position history information indicating the position history of each player to be analyzed by the attack pattern extraction unit 20.

Here, it is assumed that position history information of each player is segmented at every attack opportunity. If the attack opportunity is i, the player is u, the time is t, and the position is (x, y), the position information of each player can be expressed as (i, u _ij , t _ij , x _ij , y _ij ) it can. That is, when the position information index when the position information at the i-th attack opportunity is arranged in time series is j, the position information (i, u _ij , t _ij , x _ij , y _ij ) In the attack opportunity, it indicates that the player u _ij was present at the time t _ij at the position (x _ij , y _ij ).

  That is, as shown in FIG. 2 as an example, each player's position history information is information in which an attacking opportunity number, an event number, a player, and a position are associated with each other. In the example shown in FIG. 2, the attack opportunity number “1”, the event number “1”, the player “U1”, and the position “(12.0, 54.5)” are associated with each other, and the attack opportunity number “ 1 ”, event number“ 1 ”, player“ U2 ”, and position“ (5.0, 60.4) ”are associated with each other. In this case, at the “1” event of the “1” attack opportunity, the “U1” player is at the position (12.0, 54.5), and the “U2” player is (5 .0,60.4).

  Further, in the example shown in FIG. 2, the attack opportunity number “1”, the event number “2”, the player “U1”, and the position “(10.0, 90.5)” are associated with each other. In this case, the player of “U1” who was at the position of (12.0, 54.5) in the “1” event of the “1” attack opportunity is changed to (10. (0,90.5).

  The player position history storage unit 12 reads position history information for each player according to a request from the attack pattern extraction unit 20, and outputs the read position history information to the attack pattern extraction unit 20. Also, according to the request from the attack pattern generation unit 30, the player position history storage unit 12 receives position history information for each player of the attack opportunity i corresponding to the attack pattern indicated by the attack pattern number received by the reception unit 18. The read position history information is output to the attack pattern generation unit 30.

  The play history storage unit 14 is a database server or the like equipped with a database, and stores play history information indicating the play history of each player to be analyzed by the attack pattern extraction unit 20.

The play history can be represented as (i, a _ij , u _ij , u _{ij + 1} , t _ij ), where i represents an attack opportunity, u represents a player, a represents an action, and t represents time. When the play history index when the play history at the i-th attack opportunity is arranged in time series is j, (i, a _ij , u _ij , u _{ij + 1} , t _ij ) is At time t _ij , it indicates that the player u _ij has performed the action a _ij . However, when the action a _ij is a pass, the player u _{ij + 1} represents the receiver of the pass, and when the action a _ij is other than a pass, u _{ij + 1} represents an empty element.

  As shown in FIG. 3 as an example, the play history information is information in which an attack opportunity number, an event number, a player, a player (pass receiver), and an action are associated with each other. In the example shown in FIG. 3, the attack opportunity number “1”, the event number “1”, the player “U1”, the player (pass recipient) “U2”, and the action “pass” are associated with each other. The number "1", the event number "2", the player "U2", the player (pass recipient) "-", and the action "dribble" are associated with each other. In this case, the "U1" player passes the "U2" player at the "1" event of the "1" attack opportunity, and the "2" event of the "1" attack opportunity. In this event, it is shown that the player of “U2” dribbling.

  The play history storage unit 14 reads the play history information in accordance with a request from the attack pattern extraction unit 20 and outputs the read play history information to the attack pattern extraction unit 20. Further, the play history storage unit 14 reads and reads the play history information of the attack opportunity i corresponding to the attack pattern indicated by the attack pattern number received by the reception unit 18 according to the request from the attack pattern generation unit 30. The play history information is output to the attack pattern generation unit 30.

  The operation unit 16 is connected to display means such as a display for displaying a menu screen and the like and input means such as a keyboard, a mouse, and a touch panel, and is stored in the player position history storage unit 12 and the play history storage unit 14. Various operations using the input means from the user for received data are accepted. The various operations include operations for registering information, correcting and deleting stored information, and the like. The operation unit 16 is realized by display control software of the display unit and a device driver of the input unit.

を一意に識別するための選手ｕのＩＤ等の情報と、選手位置履歴格納部１２に格納されている位置履歴情報と同じ形式の位置履歴情報と、プレイ履歴格納部１４に格納されているプレイ履歴情報と同じ形式のプレイ履歴情報である。 The receiving unit 18 is connected to display means such as a display for displaying a menu screen and the like and input means such as a keyboard, a mouse, and a touch panel, and shows an attack pattern that the user wants to visualize, that is, an attack pattern to be output. The input of the attack pattern number and the information on the player u to be analyzed is received. Information about the player u is stored in the player position history storage unit 12 and the play history storage unit 14.

Information for uniquely identifying the player u, position history information in the same format as the position history information stored in the player position history storage unit 12, and the play stored in the play history storage unit 14 The play history information has the same format as the history information.

  The attack pattern extraction unit 20 includes a position information quantization unit 22, a feature vector generation unit 24, an attack pattern classification unit 26, and an attack pattern number storage unit 28.

  The position information quantization unit 22 acquires the position history information of each player u stored in the player position history storage unit 12. Further, the position information quantization unit 22 identifies a representative point which is a position where the player u is likely to exist based on the acquired position history information, and represents the position of each player u in a plurality of attack opportunities i as one of the representative points. By corresponding to the points, the position information of each player u in the plurality of attack opportunities i is quantized. Further, the position information quantization unit 22 associates the position information (for example, information indicating a representative point to be described later) obtained by quantizing the position information with each piece of position information in the position history information, and thus the player position history. The data is stored in the storage unit 12.

  Here, specifying a representative point is equivalent to solving a mode search problem that is a method for searching for a mode value. That is, it is assumed that each position of the player u at each time is plotted as a point on the court based on the acquired position information. In this case, a function representing how a collection of plotted points is distributed is called a density function. The density function is a function representing the density of a point (player) at any position (x, y) in space, and many points gather around the high value position, and around the low value position It shows that there are not many points.

  Further, in the mode search problem, the mode is defined as the maximum value of the density function, and the point corresponding to the maximum value indicates a position where the density of points is locally high in space. Further, the position information (x, y) is quantized by making the position of each player u at each time point correspond to any of the extracted L types of representative points. For example, in the mode search problem, mean shift clustering that estimates the position at which the density function has a maximum value from the position of the player u, as a method of making each point corresponding to the maximum value correspond to any representative point It is conceivable to use the method. The advantage of using the mean shift clustering method is that the number L of representative points does not have to be set, and automatic determination can be made from data of a set of points.

  By using the mean shift clustering method, it is not necessary to determine the number of divisions of space in advance, and quantization is performed flexibly and automatically according to the distribution of the position of each player u be able to.

  The mean shift clustering method is described in detail in Non-Patent Document 3 above. Further, the method of quantizing the position information of each player at a plurality of attack opportunities i is not limited to the mean shift clustering method, and an existing clustering method such as the k-means method may be used.

  FIG. 4 shows the result of specifying representative points using the player's position history information in an actual basketball game by means of the mean shift clustering method. In FIG. 4, the player 34's own goal 34a is provided on the right side of the court 34 in front view, and the enemy 34's goal 34b is provided on the left side in front view. Also, the black points 36 in FIG. 4 represent the positions of the players at each time, and the hatched points 38 in FIG. 4 represent the identified representative points. According to FIG. 4, in basketball, when player u attacks, it is understood that a position having a 45 degree inclination with respect to the enemy's goal 34b is a position where the player is likely to be present.

  The feature vector generation unit 24 acquires the play history information of the player u stored in the play history storage unit 14. Further, the feature vector generation unit 24 uses the position information quantized in the position information quantization unit 22 to determine the number of times the player u has moved between the quantized position information in each of the plurality of attack opportunities i. For each of the plurality of attack opportunities i, the feature vector relating to the movement and action of the player u is generated from the generated number vectors and the play history representing the action of each player u in the plurality of attack opportunities i. Is generated.

とする。 Specifically, first, the player's position information at the ith attack opportunity is quantized using L types of representative points. Using this, a matrix representing where each player moved from where, that is, a tendency of movement between quantized location information of each player u.

And

とし、Ｌ^２次元の回数ベクトルとする。 Each element x ⁽ⁱ⁾ _{ll '} of matrix X ⁽ⁱ⁾ represents that the number of times of movement from representative point l to representative point l' was x ⁽ⁱ⁾ _{ll '} times at the ^ith attack opportunity . Also, transform the matrix X ⁽ⁱ⁾ into a vector with K ² elements

And an L ^two- dimensional frequency vector.

とする。ここで、出場選手の人数をＱとした。行列Ｍ^（ｉ）の各要素ｍ^（ｉ） _ｑｑ’は、ｑ＝ｑ’の場合は、ｉ番目の攻撃機会において選手ｑがドリブルを行った回数を表す。また、ｑ≠ｑ’の場合は、ｉ番目の攻撃機会において選手ｑから選手ｑ’へのパスが通った回数を表す。さらに、行列Ｍ^（ｉ）をＱ^２個の要素を持つベクトルに直したものを

とし、Ｑ^２次元ベクトルとする。 Next, a method of generating a vector using a play history (pass, dribble) will be described. For the i-th attack opportunity, the number of passes passed between each player and the number of times each player dribble, that is, a matrix indicating the action tendency of each player u

And Here, the number of participating players is Q. Each element m ⁽ⁱ⁾ _{qq ′} of the matrix M ⁽ⁱ⁾ represents the number of times the player q dribbles at the i-th attack opportunity when q = q ′. Also, in the case of q ≠ q ', it represents the number of times the pass from the player q to the player q' has been taken at the ith attack opportunity. Furthermore, the matrix M ⁽ⁱ⁾ is converted to a vector with Q ² elements

Let Q be a ^two- dimensional vector.

と表す。ここで、Ｉは攻撃機会の総数である。 Furthermore, the number vector x ^ ^{(i) of} the i-th attack opportunity and the vector m ^ ⁽ⁱ⁾ are connected to form one connected vector d ^ (i) = (x ^ ⁽ⁱ⁾ , m ^ ⁽ⁱ⁾ It is expressed as). Here, the symbol attached with "^" is a vector, and hereinafter, the vector is represented by adding "^". Also, the connected vector d ^ ⁽ⁱ⁾ is a vector of (L ² + Q ² ) dimensions. Generate a connected vector d ^ ⁽ⁱ⁾ for all attack opportunities i,

It expresses. Here, I is the total number of attack opportunities.

とする。 When the size of quantization is large, or when the number of players u is large, it is assumed that the connected vector d ^ ⁽ⁱ⁾ has a high dimension. Therefore, it is effective to use a feature conversion method for projecting high-dimensional space data to a low-dimensional space. For example, it is conceivable to use principal component analysis as a feature conversion method. Principal component analysis determines variables (features) of a space to be projected under the condition that information loss is minimized when data of a high-dimensional space is projected to a low-dimensional space. This can be realized by setting the principal component in the direction in which the variance of the data is maximized. However, the feature conversion method is not limited to this, and a feature conversion method based on factor analysis or the like may be used. A vector obtained by performing feature conversion on the connected vector d ^ ⁽ⁱ⁾ is defined as a feature vector f ^ ^(i), and I attack opportunities are summarized.

And

  In this way, by using the feature conversion method, not only at which position player u can easily play but also analysis of attack patterns taking into consideration how each player u moves based on the movement information of all players u It becomes possible.

  The feature conversion method is described in detail in Non-Patent Document 4 above.

  Also, when the feature conversion method is not used, and the data to be analyzed has high dimensions, deterioration of the clustering accuracy occurs as shown in the above non-patent document 5, but the feature conversion method By using it, this deterioration of clustering accuracy can be avoided.

The attack pattern classification unit 26 acquires the feature vector F ^ ⁽ⁱ⁾ generated by the feature vector generation unit 24. In addition, the attack pattern classification unit 26 converts a plurality of (here, I times) attack opportunities into a plurality (here, K types) of attack patterns based on the degree of similarity of the acquired feature vector F ^ ^(i). Classify. Note that the number K of attack patterns needs to be set in advance.

As a method for classifying I attack opportunities into K types of attack patterns, for example, a clustering method based on machine learning such as a k-means method may be used. In the k-means method, a process of assigning K cluster centers to a feature vector on a feature space and assigning each feature vector to one of the K clusters, and a feature assigned to each cluster. K cluster centers are determined by repeating the process of calculating the centroid of the vector. Under the present circumstances, it is estimated that the attack pattern finally assigned to the same cluster is a similar attack pattern. The method of classifying I attack opportunities into K types of attack patterns is not limited to the k-means method, and an existing clustering method such as a clustering method using a mixed Gaussian distribution may be applied. The attack pattern classification unit 26 outputs the cluster number k ^ ⁽ⁱ⁾ assigned to the i-th attack opportunity and stores it in the attack pattern number storage unit 28.

  The k-means method is described in detail in Non-Patent Document 5 above.

  Moreover, the clustering method based on the existing machine learning is applicable by using a feature vector as an input. Thereby, an attack pattern can be extracted based on the similarity of feature vectors.

Attack pattern number storage unit 28 is a database server or the like having a database, classified by attack pattern classifying unit 26, the cluster number corresponding to the attack pattern number assigned to each attack opportunities i k ^ ^{( i)} is stored.

  The attack pattern generation unit 30 acquires, from the attack pattern number storage unit 28, the attack opportunities i classified into the attack pattern number k accepted by the reception unit 18. In addition, the attack pattern generation unit 30 is based on the acquired play history information at the attack opportunity i and the quantized position information at the acquired attack opportunity i, and a matrix representing the action tendency of each player u, and each A matrix representing a tendency of movement between quantized position information of the player u is generated.

及び、各々の選手がどこからどこへ移動したかを表す行列（代表点間の移動のしやすさを表す行列）

を生成して出力する。 That is, the attack pattern generation unit 30 corresponds to the attack pattern number k received by the reception unit 18 and represents a matrix (pass) indicating the number of times that each player has passed and the number of times each player has dribbled. (Matrix indicating ease of passage and ease of dribbling)

And a matrix indicating where each player has moved from where (a matrix indicating ease of movement between representative points)

Generate and output

  In addition, about the 1st calculation process which calculates | requires the matrix showing the tendency of the action of each player u, and the 2nd calculation process which calculates | requires the matrix showing the movement tendency between the quantized positional information of each player u, This will be described later with reference to the flowcharts shown in FIGS.

The output unit 32 outputs, based on the matrix M ^ output by the attack pattern generation unit 30, a heat map indicating the tendency of actions between players in the attack pattern number k received by the reception unit 18. Further, the output unit 32 generates a plan view showing what movement trajectory each player can easily draw based on the matrix X ^ output by the attack pattern generation unit 30, and the movement trajectory pattern of each player. Output.
At this time, in order to make it easy to visually recognize the movement trajectory of the player u, a threshold value is set as to whether or not to output each element of the matrix X ^ ^u related to the player u of the matrix X ^, and the element of the matrix X ^ ^u Only the portion exceeding the threshold may be output.

  Here, the term "output" is a concept including display by display means such as a display, printing by an image forming apparatus such as a printer, voice output by a speaker or the like, transmission to an external device, and the like. However, the output unit 32 may or may not include an output device such as a display, a printer, or a speaker. That is, the output unit 32 is realized by drive software of an output device, or driver software and an output device of an output device.

  The attack pattern extraction device 10 according to the present embodiment is configured by, for example, a computer device provided with a central processing unit (CPU), a random access memory (RAM), and a read only memory (ROM) for storing various programs. . The computer constituting the attack pattern extraction apparatus 10 may include a storage unit such as a hard disk drive or a nonvolatile memory. In the present embodiment, when the CPU reads and executes the program stored in the storage unit such as the ROM and the hard disk, the above-described hardware resource and the program cooperate with each other to realize the above-described function.

  Here, as described above, the attack pattern extraction device 10 according to the present embodiment performs an extraction process of extracting an attack pattern based on the position history information and the play history information. A flow of processing when the attack pattern extraction device 10 according to the present embodiment performs an extraction process of extracting an attack pattern will be described using a flowchart shown in FIG.

  In the present embodiment, the extraction processing program is started at a timing when the operation unit 16 is operated in advance.

  In step S101, the position information quantization unit 22 acquires position history information from the player position history storage unit 12, and acquires play history information from the play history storage unit 14.

  In step S103, the position information quantization unit 22 quantizes the position information of each player based on the position history information.

In step S105, the feature vector generation unit 24, for each of the plurality of attack opportunities i, based on the position information of each player quantized in step S103, a frequency vector (x ^) representing the number of times each player has moved. Generate ⁽ⁱ⁾ .

In step S107, the feature vector generation unit 24 generates a vector (m ⁽ⁱ⁾ ) representing the number of actions of each player based on the play history information for each of the plurality of attack opportunities i.

In step S109, the feature vector generation unit 24, for each of a plurality of attack opportunities i, a frequency vector representing the number of times each player has moved in step S107, and the number of actions of each player generated in step S109. A connected vector (d ^ ⁽ⁱ⁾ ) is generated by connecting a vector representing.

In step S111, the feature vector generation unit 24 generates a feature vector (f ^ ⁽ⁱ⁾ ) for each of the plurality of attack opportunities i from the concatenated vector generated in step S109.

In step S113, the attack pattern classification unit 26 classifies the plurality of attack opportunities i into K types of attack patterns based on the feature vector (f ^ ⁽ⁱ⁾ ) generated for each of the plurality of attack opportunities i. .

In step S115, the attack pattern classification unit 26 causes the attack pattern number storage unit 28 to store the cluster number (k ^ ⁽ⁱ⁾ ) assigned to the attack opportunity i as a classification result.

  As described above, the attack pattern extraction device 10 according to the present embodiment uses the position information quantization unit 22 to each of the plurality of attack opportunities based on the position history representing the position information of each player in the plurality of attack opportunities i. Quantize the position information of players. In addition, the attack pattern extraction device 10 uses the position information quantized by the feature vector generation unit 24 to indicate the number of times the player has moved between the quantized position information in each of the plurality of attack opportunities. A vector is generated, and a feature vector relating to player's movement and action is generated for each of the plurality of attack opportunities i from the generated number vector and the play history representing the action of each player in the plurality of attack opportunities i. Then, the attack pattern extraction apparatus 10 classifies the plurality of attack opportunities i into a plurality of attack patterns based on the feature vector for each of the generated plurality of attack opportunities i by the attack pattern classification unit 26.

Moreover, the attack pattern extraction apparatus 10 performs the 1st calculation process which calculates | requires the matrix M ⁽ⁱ⁾ showing the tendency of each player's action with respect to the i-th attack opportunity. A flow of processing when the attack pattern extraction device 10 according to the present embodiment performs the first calculation processing will be described using the flowchart shown in FIG.

  In the present embodiment, the program of the first calculation process is started at a predetermined operation timing using the operation unit 16.

  In step S201, the attack pattern generation unit 30 acquires the attack pattern number k of the attack pattern that the user wants to visualize, that is, the attack pattern number k of the attack pattern to be output target accepted by the accepting unit 18.

In step S203, the attack pattern generation unit 30 initializes an attack opportunity i, which is a temporary variable, to “0”. Further, the attack pattern generation unit 30 initializes each element m ^ _{qq ′} of the matrix M ^ to “0”.

  In step S205, the attack pattern generation unit 30 determines whether the attack opportunity i is equal to or less than the total number I of attack opportunities i. When it is determined in step S205 that the attack opportunity i is equal to or less than the total number I of attack opportunities i (S205, Y), the process proceeds to step S207. When it is determined in step S205 that the attack opportunity i is larger than the total number I of the attack opportunities i (S205, N), the execution of the program of the first calculation process is ended.

In step S207, the attack pattern generation unit 30 determines whether or not the cluster number k ^ ⁽ⁱ⁾ assigned to the i-th attack opportunity i is the attack pattern number k. When it is determined in step S207 that the cluster number k ^ ⁽ⁱ⁾ is the attack pattern number k (Y in S207), the process proceeds to step S211. If it is determined in step S207 that the cluster number k ^ ⁽ⁱ⁾ is not the attack pattern number k (S207, N), the process proceeds to step S209.

  In step S209, the attack pattern generation unit 30 adds “1” to the attack opportunity i, which is a temporary variable, and proceeds to step S205.

  In step S211, the attack pattern generation unit 30 initializes the index j of the position information when arranging the position information in the i-th attack opportunity, which is a temporary variable, in time series to “0”.

In step S213, the attack pattern generation unit 30, an index j of the position information when arranged in time series position information in the i-th attack opportunities, obtained by subtracting "1" from the total number J _i of the index j 'J _i −1 ”or less is determined. If it is determined in step S213 that the index j is equal to or less than “J _i −1” obtained by subtracting “1” from the total number J _{i of the} index j (S213, Y), the process proceeds to step S215. If it is determined in step S213 that index j is larger than "J _i -1" obtained by subtracting "1" from the total number J _{i of} index j (S213, N), the process proceeds to step S209.

In step S215, the attack pattern generation unit 30 adds “1” to the element m ^ _{qq ′} of the matrix M ^.

  In step S217, the attack pattern generation unit 30 adds “1” to the index j of the position information when arranging the position information in the i-th attack opportunity, which is a temporary variable, in time series, and proceeds to step S213. To do.

In addition, as described above, the attack pattern extraction device 10 according to the present embodiment is a matrix X ⁽ⁱ representing a tendency of movement between the quantized position information of each player u with respect to the ith attack opportunity ^{). )} performs a second calculation process for obtaining the. The flow of processing when the attack pattern extraction device 10 according to the present embodiment performs the second calculation processing will be described using the flowchart shown in FIG. 7.

  In the present embodiment, the program for the second calculation process is started at a timing at which a predetermined operation is performed using the operation unit 16.

  In step S301, the attack pattern generation unit 30 acquires the attack pattern number k of the attack pattern that the user wants to visualize, that is, the attack pattern number k of the attack pattern to be output target received by the reception unit 18.

In step S303, the attack pattern generation unit 30 initializes the attack opportunity i, which is a temporary variable, to “0”. In addition, the attack pattern generation unit 30 initializes each element x ^ ^u _{ll '} of the matrix X ^ ^u to "0".

  In step S305, the attack pattern generation unit 30 determines whether the attack opportunity i is equal to or less than the total number I of attack opportunities i. If it is determined in step S305 that the attack opportunity i is less than or equal to the total number I of attack opportunities i (S305, Y), the process proceeds to step S307. When it is determined in step S305 that the attack opportunity i is larger than the total number I of the attack opportunities i (S305, N), the execution of the program of the second calculation process is ended.

In step S307, the attack pattern generation unit 30 determines whether or not the cluster number k ^ ⁽ⁱ⁾ assigned to the i-th attack opportunity i is the attack pattern number k. If it is determined in step S307 that the cluster number k ^ ⁽ⁱ⁾ is the attack pattern number k (S307, Y), the process proceeds to step S311. If it is determined in step S307 that the cluster number k ^ ⁽ⁱ⁾ is not the attack pattern number k (S307, N), the process proceeds to step S309.

  In step S309, the attack pattern generation unit 30 adds “1” to the attack opportunity i, which is a temporary variable, and proceeds to step S305.

  In step S311, the attack pattern generation unit 30 selects one player u from all the players u.

  In step S313, the attack pattern generation unit 30 initializes an index j of position information when the position information at the i-th attack opportunity, which is a temporary variable, is arranged in time series, to “0”.

In step S315, the attack pattern generation unit 30, an index j of the position information when arranged in time series position information in the i-th attack opportunities, obtained by subtracting "1" from the total number J _i of the index j 'J _i −1 ”or less is determined. If it is determined in step S315 that the index j is equal to or less than "J _i -1" obtained by subtracting "1" from the total number J _{i of} index j (S315, Y), the process proceeds to step S317. If it is determined in step S315 that index j is greater than "J _i -1" obtained by subtracting "1" from the total number J _{i of} index j (S315, N), the process proceeds to step S321.

In step S317, the attack pattern generation unit 30 adds “1” to each element x ^ ^u _{lijlij + 1} of the matrix X ^ ^u .

  In step S319, the attack pattern generation unit 30 adds “1” to the index j of the position information when the position information at the i-th attack opportunity, which is a temporary variable, is arranged in time series, and the process proceeds to step S315. To do.

  In step S321, the attack pattern generation unit 30 determines whether there are unprocessed players u among all the players U, that is, players u who have not performed the processes of steps S311 to S319. If it is determined in step S321 that there is an unprocessed player u (S321, Y), the process proceeds to step S311. When it is determined in step S321 that there is no unprocessed player u (S321, N), the process proceeds to step S305.

  As described above, the attack pattern extraction apparatus 10 according to the present embodiment receives the input of the attack pattern number of the attack pattern to be output by the receiving unit 18. In addition, the attack pattern extraction device 10 is based on the play history in the attack opportunity i classified by the attack pattern extraction unit 20 or the quantized position information in the attack opportunity i according to the attack pattern indicated by the received attack pattern number. Then, a matrix representing a tendency of each player's action and a matrix of a tendency of movement between the quantized position information of each player are generated.

  That is, in the present embodiment, a plurality of (here, K types) attack patterns are used with the information stored in the player position history storage unit 12 and the information stored in the play history storage unit 14 as input data. Is extracted, a plurality of attack opportunities i are assigned to any one of the extracted K types of attack patterns. Thereby, the input of the attack pattern number k that the user wants to visualize is received via the receiving unit 18, and the heat map indicating the pass between the players and the dribbling tendency of each player according to the received attack pattern number k. And the movement trajectory pattern of each player can be obtained as an output.

  FIG. 8 shows an output screen 40 by the output unit 32 when “attack pattern 1” is accepted as an attack pattern. In the example shown in FIG. 8, the accepted attack pattern (here, “1”) is shown in the display column 42 on the front view upper side. Further, a graph 44 in the lower left of the front view shows a tendency of passing between players and a tendency of dribbling of each player. In this graph, the vertical axis and the horizontal axis respectively represent players (here, q1, q2, etc.). In this graph, when the player on the vertical axis is different from the player on the horizontal axis, a pass is shown, and when the player on the vertical axis is the same as the player on the horizontal axis, dribble is shown. Also, a plan view 46 in the lower right of the front view shows a tendency of movement of each player (here, “Player A”, “Player B”, etc.) (here, representative movement of each player in the attack pattern 1). Series).

  Using these heat maps and movement trajectory patterns, it is possible to analyze what kind of attack pattern the analysis target team is good at. For example, for a team to be analyzed, it is possible to visualize what kind of pass or dribble the ball is brought to near the goal and where it is likely to be shot.

  Moreover, in this embodiment, the movement of the player who does not hold | maintain a ball is also contained in the attack pattern visualized. This makes it easy for players who do not hold the ball to deceive, attract the defending player, gather the ball where space is available, and shoot easily. It is also possible to visualize the activities of.

  In addition, this embodiment is not specialized for basketball as described above, but is intended for general sports. For example, the present invention can be applied not only to team sports (soccer, rugby, etc.) similar in form to basketball but also to competitions such as badminton and tennis performed in individual matches and doubles.

  Further, in the present embodiment, the player position history storage unit 12, the play history storage unit 14, and the attack pattern number storage unit 28 are database servers etc. provided with a database, but the present invention is not limited thereto. It may be a general-purpose storage device (a memory or a hard disk device) provided in advance, and it may be any storage means that can store various data and can be restored.

  Further, in the present embodiment, the operation of the components of the functions shown in FIG. 1 is constructed as a program and is installed and executed on a computer used as the attack pattern extraction apparatus 10, but is not limited to this. It may be distributed.

  Further, the constructed program may be stored in a portable storage medium such as a hard disk, a flexible disk, a CD-ROM, etc., and may be installed in a computer or distributed.

DESCRIPTION OF SYMBOLS 10 Attack pattern extraction apparatus 12 Player position history storage part 14 Play history storage part 16 Operation part 18 Reception part 20 Attack pattern extraction part 22 Position information quantization part 24 Feature vector generation part 26 Attack pattern classification | category part 28 Attack pattern number storage part 30 Attack pattern generation unit 32 Output unit

Claims (8)

A position information quantizing unit that quantizes position information of each player in the plurality of attack opportunities based on a position history representing position information of each player in a plurality of attack opportunities;
The position information quantized by the position information quantization unit is used to generate a number vector representing the number of times the player has moved between the quantized position information in each of the plurality of attack opportunities, A feature vector generation unit that generates, for each of the plurality of attack opportunities, a feature vector relating to the player's movement and action from the generated number vector and the play history representing the action of each player in the plurality of attack opportunities. When,
An attack pattern classifying unit that classifies the plurality of attack opportunities into a plurality of attack patterns based on the feature vector for each of the plurality of attack opportunities generated by the feature vector generating unit;
Attack pattern extraction device equipped with. The position information quantization unit corresponds the position information of each player to a position where the density function representing the density of the player at each position is obtained using the mean shift clustering method is a maximum value. The attack pattern extraction device according to claim 1, wherein the position information is quantized. The feature vector generation unit generates, for each of the plurality of attack opportunities, a connected vector in which the number-of-times vector and a vector representing the number of times of each action generated from the play history are connected. The feature vector for each of the plurality of attack opportunities is generated by performing feature transformation on a connection vector for each of the attack opportunities using a feature transformation method that projects to a low-dimensional space. Attack pattern extraction device. The attack pattern extraction device according to any one of claims 1 to 3, wherein the attack pattern classification unit classifies the plurality of attack opportunities into a plurality of attack patterns using a clustering method. The attack pattern extraction device according to any one of claims 1 to 4,
A reception unit that receives an input of an attack pattern number of an attack pattern to be output;
Based on the play history at the attack opportunity or the quantized position information at the attack opportunity classified by the attack pattern extraction device into the attack pattern indicated by the attack pattern number accepted by the reception unit. An attack pattern generation unit that generates at least one of a matrix representing the action tendency and a movement tendency matrix between the quantized position information of each player,
Attack pattern display device equipped with An attack pattern extraction method of an attack pattern extraction device comprising a position information quantization unit, a feature vector generation unit, and an attack pattern classification unit,
The position information quantization unit quantizes the position information of each player in the plurality of attack opportunities based on a position history representing the position information of each player in a plurality of attack opportunities;
The feature vector generation unit uses the position information quantized by the position information quantization unit to determine the number of times the player has moved between the quantized position information in each of the plurality of attack opportunities. For each of the plurality of attack opportunities, a feature vector relating to the player's movement and action is generated from the generated number vectors and the generated play vectors representing the actions of the respective players in the plurality of attack opportunities. Generating the
The attack pattern classification unit classifying the plurality of attack opportunities into a plurality of attack patterns based on the feature vector for each of the plurality of attack opportunities generated by the feature vector generation unit;
Attack pattern extraction method to do. An attack pattern display method of an attack pattern display device comprising the attack pattern extraction device according to any one of claims 1 to 4, a reception unit, and an attack pattern generation unit,
The receiving unit receiving an input of an attack pattern number of an attack pattern to be output;
The attack pattern generation unit is classified by the attack pattern extraction device into the attack pattern indicated by the attack pattern number received by the reception unit, the play history at the attack opportunity, or the quantization at the attack opportunity Generating at least one of a matrix indicating the tendency of the action and a matrix indicating a tendency of movement between the quantized position information of each of the players based on the position information obtained;
How to display attack patterns.   The program for functioning a computer as each part of the attack pattern extraction apparatus in any one of Claims 1-4, or each part of the attack pattern display apparatus in any one of Claim 5.

Images