KR20230062289A - Method for providing tactical analysis related to team sports - Google Patents

Abstract

The present invention generally relates to a sports analysis field and, more specifically, to analysis on tactical changes in team formations and roles in team sports. In accordance with one aspect of the present disclosure, a tactical analysis method for team sports includes the following steps of: obtaining a plurality of player tracking datasets for a plurality of players in a team sports match; obtaining a plurality of position distributions for each of the plurality of players; assigning a plurality of role identifiers to the plurality of position distributions, respectively; obtaining a plurality of player-role assignments for each of the plurality of time points; obtaining a plurality of role arrangements for each of the plurality of time points; detecting a change point in the sequence of the plurality of role arrangements; dividing the team sports match into at least first and second formation periods based on the change point; and determining a first team formation played in the first formation period from the role arrangements corresponding to the first formation period, and determining a second team formation played in the second formation period from the role arrangements corresponding to the second formation period.

Description

METHOD FOR PROVIDING TACTICAL ANALYSIS RELATED TO TEAM SPORTS}

This description relates generally to the field of sports analysis, but more specifically to the analysis of tactical changes in team formations and roles in team sports, without limitation.

Training in modern team sports is largely divided into physical training and tactical training. While physical training aims to improve the player's basic physical abilities such as strength, endurance, and quickness, tactical training aims to improve the ability to perform tactics to obtain overall or local gains during the game.

In tactical training, practical exercises to master local tactics performed by small groups, response in specific situations such as set-pieces, and overall strategies for team formation are important, but to improve team-level understanding of these Tactics analysis and review are also important elements.

Team formation and player positions are the most basic elements in tactical analysis, which are essential for higher-level tactical analysis. However, from a sports analysis point of view, team sports, especially in fluid sports where players interact in real time, such as soccer or basketball, have high-dimensional characteristics that change very fluidly, so most modern sports teams still have no knowledge of formation or position. Analysis is completely dependent on the intuition of a person such as a video analyst.

One object of the present disclosure is to provide a method for tactical analysis of team sports that performs tactical analysis for team sports from player tracking data.

One task of the present disclosure is to provide a method for tactical analysis of team sports that provides information about team formations and/or player positions.

One task of the present disclosure is to provide a method for performing higher-level tactical analysis using information about team formations and/or player positions.

According to one aspect of the present disclosure, obtaining a plurality of player tracking data sets for each of a plurality of players, each player tracking data set including the corresponding player's positions at a plurality of time points within a target session step; acquiring a plurality of position distributions for each of the plurality of players, each position distribution being generated based on a corresponding player tracking data set; assigning a plurality of role identifiers to each of the plurality of location distributions; a plurality of player-role assignments for each of the plurality of time points, each assignment reflecting a relationship between the plurality of players and the plurality of role identifiers at a first corresponding time point, and the plurality of location distributions and Obtaining - generated by assigning each of the plurality of role identifiers to the plurality of players based on positions of the plurality of players at the first corresponding time point among the plurality of player tracking data sets; A plurality of role assignments for each of the plurality of viewpoints, each role assignment reflecting a positional relationship between the plurality of role identifiers at a second corresponding viewpoint, and a player-role at the second corresponding viewpoint obtaining an allocation and generated based on the positions of the plurality of players at the second corresponding point in time; and dividing the sequence of the plurality of role assignments into at least two role assignment groups using a changepoint detection algorithm.

According to another aspect of the present disclosure, obtaining a plurality of player tracking data sets for each of a plurality of players, each player tracking data set including the corresponding player's positions at a plurality of time points within a target session. step; acquiring a plurality of position distributions for each of the plurality of players, each position distribution being generated based on a corresponding player tracking data set; assigning a plurality of role identifiers to each of the plurality of location distributions; a plurality of player-role assignments for each of the plurality of time points, each assignment reflecting a relationship between the plurality of players and the plurality of role identifiers at a first corresponding time point, and the plurality of location distributions and Obtaining - generated by assigning each of the plurality of role identifiers to the plurality of players based on positions of the plurality of players at the first corresponding time point among the plurality of player tracking data sets; and acquiring a main player-role assignment from the plurality of player-role assignments.

According to an embodiment of the present disclosure, it is possible to detect a change point in a team formation from player tracking data, search a point where a team formation changes during a game, and identify a team formation for each period.

According to an embodiment of the present description, a change in a player's position can be identified from player tracking data, and a play style of a team can be extracted.

According to an embodiment of the present disclosure, a role switch may be identified from player tracking data, and a set-piece situation may be retrieved.

The problems, solutions, and effects to be solved by the present invention are not limited to the above-mentioned matters, and matters not mentioned are clearly apparent to those skilled in the art from this specification and the accompanying drawings. You will be able to understand.

1 is an example of a timeline of formation and role changes according to the results of SportsCPD according to an embodiment of the present disclosure.
2 and 3 are graphs of role adjacency according to an embodiment of the present disclosure.
4 relates to an example formation group according to an embodiment of the present disclosure.
5 is a diagram related to formation prediction according to an embodiment of the present disclosure.
6 is a graph of role substitution sequences for each formation period according to an embodiment of the present disclosure.
7 is a diagram related to an example of tactic analysis using a role substitution sequence according to an embodiment of the present disclosure.
8 is a diagram illustrating a relationship between a switch rate and set-piece occurrence according to an embodiment of the present disclosure.
9 is a block diagram of a system for providing tactical analysis of team sports according to an embodiment of the present disclosure.
10-17 are flow charts of examples of methods for providing tactical analysis of team sports according to embodiments of the present disclosure.

The embodiments described in this description are intended to clearly explain the spirit of the present invention to those skilled in the art to which the present invention belongs, and the present invention is not limited by the examples described in this description, and the present invention The scope of should be construed as including modifications or variations that do not depart from the spirit of the present invention.

The terminology used in this description is a selection of general terms that are currently widely used as much as possible in the technical field to which the present invention belongs, but these terms are selected according to the intention, custom, or the emergence of new technologies of those skilled in the art in the technical field to which the present invention belongs. meaning may vary. However, in the case where a specific term is defined and used in an arbitrary meaning, the meaning of the term will be separately described. Therefore, the terms used in this description should be interpreted based on the actual meaning of the term and the overall content of this description, not the simple name of the term.

The drawings attached to this description are for easily explaining the present invention, and the shapes shown in the drawings may be exaggerated or abbreviated as necessary to help the understanding of the present invention, so the present invention is not limited by the drawings. .

In the present description, if it is determined that a detailed description of a known configuration or function related to the present invention may obscure the gist of the present invention, a detailed description thereof will be omitted if necessary.

According to one aspect of the present disclosure, obtaining a plurality of player tracking data sets for each of a plurality of players, each player tracking data set including the corresponding player's positions at a plurality of time points within a target session step; acquiring a plurality of position distributions for each of the plurality of players, each position distribution being generated based on a corresponding player tracking data set; assigning a plurality of role identifiers to each of the plurality of location distributions; a plurality of player-role assignments for each of the plurality of time points, each assignment reflecting a relationship between the plurality of players and the plurality of role identifiers at a first corresponding time point, and the plurality of location distributions and Obtaining - generated by assigning each of the plurality of role identifiers to the plurality of players based on positions of the plurality of players at the first corresponding time point among the plurality of player tracking data sets; A plurality of role assignments for each of the plurality of viewpoints, each role assignment reflecting a positional relationship between the plurality of role identifiers at a second corresponding viewpoint, and a player-role at the second corresponding viewpoint obtaining an allocation and generated based on the positions of the plurality of players at the second corresponding point in time; and dividing the sequence of the plurality of role assignments into at least two role assignment groups using a changepoint detection algorithm.

Here, the method further comprises: obtaining a primary player-role assignment from the plurality of player-role assignments; and determining a non-regular player-role assignment of at least one of the plurality of player-role assignments based on the main player-role assignment, wherein the sequence of plurality of role assignments comprises: Among the batches, only the remainder except for the at least one non-regular player-role assignment may be included.

Here, the method comprises: assigning the target session to at least two time periods corresponding to each of the at least two role placement groups - the at least two time periods, wherein the main formation of a team participating in the target session is a first team and a second time period in which the primary formation of the team is a second team formation.

wherein the method further comprises determining the first team formation from at least one first role assignment with respect to the first time period and the second team formation from at least one second role placement with respect to the second time period. Step; may further include.

According to another aspect of the present disclosure, obtaining a plurality of player tracking data sets for each of a plurality of players, each player tracking data set including the corresponding player's positions at a plurality of time points within a target session. step; acquiring a plurality of position distributions for each of the plurality of players, each position distribution being generated based on a corresponding player tracking data set; assigning a plurality of role identifiers to each of the plurality of location distributions; a plurality of player-role assignments for each of the plurality of time points, each assignment reflecting a relationship between the plurality of players and the plurality of role identifiers at a first corresponding time point, and the plurality of location distributions and Obtaining - generated by assigning each of the plurality of role identifiers to the plurality of players based on positions of the plurality of players at the first corresponding time point among the plurality of player tracking data sets; and acquiring a main player-role assignment from the plurality of player-role assignments.

Here, the main player-role assignment may be the most frequent player-role assignment among the plurality of player-role assignments.

Here, the method may further include determining a plurality of positions for each of the plurality of players based on the main player-role assignment.

Here, the method may further include selecting a secondary player-role assignment from the plurality of player-role assignments.

Here, the secondary player-role assignment may be the second most frequent player-role assignment among the plurality of player-role assignments.

Here, the method may further include acquiring information on position conversion by comparing the main player-role assignment and the secondary player-role assignment.

Here, the method may further include acquiring switch rate information regarding the plurality of players-role assignments based on the main player-role assignments.

Here, the method may further include detecting an irregular situation in the target session based on the switch rate information.

1. Introduction

In fluid team sports such as football and basketball, analyzing team formations is one of the most intuitive ways to understand tactics from the practitioner's point of view. However, existing approaches have limitations in not fully reflecting the actual game situation even if they assume that the team formation is consistent throughout the game or assign the formation for each scene. With this problem in mind, in this article, SportsCPD (Sports Change-Point Detection), a change point detection framework that distinguishes tactically intended changes in formations and player positions during team sports games from temporary changes using spatio-temporal tracking data, is developed. Suggest.

Here, the term "player position" can be interpreted as referring to a role assigned to a player in team sports (eg, center forward or central defender in soccer), and the term "team formation" refers to a team's affiliation. Some form of combination of roles assigned to all or most of the players (e.g. field players excluding the goalkeeper in soccer) (e.g. "4-4-2" or "3-4-3" in soccer). ) can be interpreted as referring to

Here, as an example of "intentional change", a team that used a "4-4-2" formation in the first half of a soccer match adopts a "3-4-3" formation in the second half under the tactical judgment of the coaches. An example of a "temporary change" is a temporary change, such as playing the role of an attacker for a short period of time by overlapping a defender in a soccer game. , random or temporary occurrence of such changes.

SportsCPD according to an embodiment of the present disclosure may detect a point at which a team formation changes (hereinafter referred to as a “formation change point”). Hereinafter, this is referred to as Formation Change-Point Detection (FormCPD). Illustratively, FormCPD first acquires information reflecting the roles assigned to players for each scene, obtains information reflecting the positional relationship between the roles, and then obtains information about the positional relationship between the roles. This can be done by detecting formation change points from the sequence.

In addition, SportsCPD according to an embodiment of the present disclosure can detect a player position, that is, a point at which a role assigned to a player changes (hereinafter referred to as a “role change point”). Hereinafter, this is referred to as Role Change-Point Detection (RoleCPD). Illustratively, RoleCPD may be performed by detecting a role change point from a sequence of information reflecting roles assigned to players.

Here, information reflecting roles assigned to players may be referred to as "player-role assignments". For example, player-role assignment can be expressed as a role permutation, which will be described in more detail later.

Also, here, information reflecting the positional relationship between roles may be referred to as "role arrangement". For example, the role arrangement may be expressed as a role-adjacency matrix, which will be described in detail later.

According to one example, detection of change points may be performed using a nonparametric change point detection technique. However, the change point detection technique is not limited thereto.

According to the result of evaluating the method according to the embodiment of the present disclosure using actual measurement data recorded by workers in the field, the method according to the embodiment of the present disclosure can accurately detect tactical change points and predict the formation of each segment. Confirmed. In addition, in the embodiments of the present description, practical examples that can be interpreted and utilized by those skilled in the art will be mentioned.

2. Overview

In fluid team sports such as soccer or basketball, players interact with each other while maintaining their roles within a specific formation. Therefore, analyzing team formations is one of the most intuitive ways to understand tactics from the point of view of practitioners in the field. Since it is clear that team formation has a significant effect on player positions during a game, there have been previous studies in the field of sports science to estimate team formation during a game using spatio-temporal tracking data.

However, the results of previous approaches differed from formation changes in actual matches. Many approaches assume that team formations are consistent throughout a game or session, but this does not reflect real-life situations in which coaches often change team formations in response to various in-game situations during a game. Other approaches have considered changing formations, but have assigned formations for each scene or for each offensive/defensive period. When this allocation method is used, considering that a coach usually instructs only a few changes during a game, there is a problem in that team formation changes too frequently, contrary to the coach's intention.

To fill this gap, an embodiment of the present disclosure proposes a changepoint detection framework called SportsCPD that distinguishes tactically intended changes in formation and role from temporary changes in team sports games.

SportsCPD may be exemplarily performed as follows.

First, roles are assigned to players for each scene. Illustratively, role assignment to players for each scene may be performed as follows. In the assignment of roles to players described later, the concept of role representation presented in Balkowski's thesis "Large-scale analysis of soccer matches using spatiotemporal tracking data (2014)" was partially borrowed.

Player tracking data sets can be acquired. Each player tracking data set relates to the location of an individual player, and may include information about the locations of the corresponding player within a session to be analyzed (hereinafter referred to as “target session”).

Here, the player tracking data set may be obtained, for example, through a positioning sensor (eg, a GPS sensor or an LPS sensor) worn and/or attached to the player or through image analysis of a sports video, and the stadium It may be expressed according to a field coordinate system. The stadium coordinate system may refer to a coordinate system having a longitudinal direction and a width direction of the stadium as axes with a point (for example, a center or a corner point of the stadium) as an origin in a field where a sports game is held. In addition, the player tracking data set may include positions in a time-series form. In this case, the time interval may be the same as the sampling interval of the sensor or image analysis, but this is not necessarily the case.

Roles can be assigned to each player based on the player tracking data set. Here, role assignment may be performed by assigning a role identifier to a player. First, under the assumption that the players have a certain role throughout the session, each of the player's position distributions during the session obtained from the player tracking data sets can be determined as the initial position distributions of the roles over the entire session. Now, we can assign roles to players according to their positions, given the initial position distribution of roles. As a simple example, roles can be assigned to players on a scene-by-scene basis such that the sum of the distances between the average position of the initial distribution of positions of roles assigned to an individual player over the players and the position of that individual player in the scene is minimized. there is. However, instead of simply minimizing the sum of distances, various techniques may be used. For example, a Hungarian algorithm such as a cost matrix based on log probability density may be used. Next, the location distribution of the roles can be newly created considering the roles assigned to the players for each scene. The initial position distribution is assumed to have a player in only one role throughout the session, so it is only built from positions pertaining to one player over the entire session, while the newly constructed position distribution is based on the player performing one role scene-by-scene. Since it can be different, it can be created from the position of a different player for each scene. If the location distributions of the roles are newly created, the roles can be assigned to the players by considering the locations of the players for each scene. Now, if the above operation is repeated a sufficient number of times or until the role allocation or position distribution is sufficiently converged, the final player-role allocation can be obtained.

On the other hand, the process of acquiring the above player-role assignments can be understood as effective in a situation where each of the players is playing a specific position on a team formation. Thus, an irregular situation in which the team formation is broken, such as a set-piece situation, may be treated as tainted data in assigning roles. Therefore, it is possible to improve the accuracy of player-role assignment by performing pre-processing to remove non-regular situations among scenes. Specifically, if an irregular situation is detected, player-role assignment may be performed except for this. Specifically, the position distribution of the roles may be calculated using only the player tracking data set of the scenes other than the scene corresponding to the non-regular situation.

Detection of non-canonical situations can be in a variety of ways.

For example, an irregular situation is a situation in which a formation is destroyed. Since a significant number of irregular situations may occur during set-pieces, detection of irregular situations may be performed using characteristics occurring in set-pieces. For example, in soccer, in a set-piece situation such as a free kick, corner kick or penalty kick, the position of the ball is temporarily fixed (and the fixed position is specified in the case of a corner kick or penalty kick), and the players move in a relatively dense positional relationship. none or very small Therefore, if the position of the ball is fixed for a certain period of time or below the threshold (in the case of a corner kick or penalty kick, a condition based on a specific position can be added), the players must be in a fixed position for a certain period of time or below the threshold. When there is movement (in the case of corner kicks or penalty kicks, a condition based on a specific position can be added), and/or when players are concentrated for a certain amount of time (this is illustratively, a predetermined number or more players within a predetermined radius) It can be determined that the scene exists within the distance or the sum of the distances between the players is less than a predetermined distance), the scene can be judged to correspond to an irregular situation.

As another example, an irregular situation may be detected based on the extent to which players deviate from a designated role. Specifically, when roles are assigned to players for each scene, dominant player-role assignment can be determined. Here, the main player-role assignment is a role-assignment mainly used by a team within the target session, that is, a role-assignment according to an indicated team formation. Primary player-role assignments can be determined based on frequency from scene-by-scene player-role assignments in the target session. For example, determining the most frequent player-role assignment as the primary player-role assignment, given that team formations typically dictated in team sports are one or less than a few over target sessions, or a predetermined number of players in order of frequency. Player-role assignments may be determined as primary player-role assignments. Here, when the formation period described later is determined, it may be desirable to assume that there is one main player-role assignment for each formation period, but note that the main player-role assignment does not necessarily have to be one for the entire target session. Should be. When the main player-role assignment is determined, an irregular situation may be detected by comparing the main player-role assignment with the player-role assignment for each scene. For example, when the player-role assignment is expressed as a permutation sequence as described later, if the distance between the permutation sequence of the main player-role assignment and the permutation sequence of the player-role assignment for each scene is greater than a threshold value, the scene corresponds to a non-regular situation. It can be judged that In addition, if there are multiple main player-role assignments, if the minimum distance between the main player-role assignment and the player-role assignment for each scene is greater than or equal to a threshold value, the scene may be determined to correspond to an irregular situation.

As another example, considering that the team formation is suddenly destroyed in an irregular situation, a case where a change in the sequence of player-role assignment for each scene exceeds a certain level may be detected as an irregular situation. Specifically, a case where the distance of the substitution sequence of adjacent player-role assignments is equal to or greater than a threshold value may be determined as an irregular situation.

After assigning roles to players for each scene, change point detection is performed in the following two steps. (1) Formation changing points are detected for each session, and (2) role changing points are detected for each segment obtained from the results. Hereinafter, each step will be described in detail.

At the FormCPD stage, the formation period can be determined in the target session by detecting formation change points using player-role assignment. Here, the formation period may mean a period during which the indicated formation is constant. In addition, formations can be identified by formation period.

For example, in FormCPD, formation change points can be detected from role arrangements.

Here, the role arrangement relates to a positional relationship between role identifiers, and may be generated from player-role assignment and player tracking data sets. In addition, since the role assignment can be generated for each scene from the player-role assignment for each scene and the positions of the players in the scene extracted from the player tracking data set, a sequence of role assignments can be obtained for the target session.

For example, the role arrangement in a specific scene can be expressed as coordinates between roles on the stadium coordinate system, which is the positional coordinates of the players in the scene extracted from the player tracking data set through player-role assignment in the scene can be created by assigning them to roles.

As another example, role placement may calculate an adjacency relationship between roles. Here, the adjacency relation relates to positional adjacency between roles and may be referred to as “role topology”. In the role topological positional relationship, the positional relationship between roles can be determined based on the adjacency between roles based on contextual adjacency instead of actual values such as location on the playing field or distance accordingly. For example, the role phase position relationship is a role-adjacency matrix calculated from Delaunay triangulation with reference to what was presented in Narizuka and Yamazaki's paper "Characterization of the formation structure in team sports (2017)" can be expressed as A role adjacency matrix using this can be obtained as follows. First, positions of roles may be obtained for each scene. In this case, the position may be based on a reference coordinate system. Next, Delaunay triangulation, that is, triangulation is performed so that the minimum angle of all angles is maximized on the polynomial with each role as a vertex, and the roles sharing the triangle are A role adjacency matrix, which is a matrix defining adjacent roles and defining roles that are not adjacent as non-adjacent roles, may be obtained.

Next, considering that the role arrangement is time-series information provided in the form of a sequence for each scene, a change point may be detected from the calculated role arrangement. In this case, it goes without saying that various change point detection algorithms may be used. The change point detection algorithm may include an online change point detection algorithm that detects a change point from an input sequence in real time and an offline change point detection algorithm that detects a change point from a completed sequence. Examples of online algorithms include sequential hypothesis testing, which detects change points by considering factors such as false alarm rate, misdetection rate, detection delay, etc. streaming algorithms, etc., examples of off-line algorithms include change-in-mean detection according to hypothesis verification method, maximum likelihood estimation according to change over time, or anomaly regression method (two -phase regiression) and the like. Therefore, in the present description, change point detection should be understood as a comprehensive concept including any sequence segmentation or statistical analysis.

According to an example, a change point may be detected from a sequence of a role adjacency matrix. In this case, a nonparametric change point detection algorithm may be applied to detect the change point, without limitation, considering that the role adjacency matrix is high-dimensional data having repetitive observations. As a non-parametric algorithm, for example, discrete g-segmentation can be used. Discrete g-segmentation can be performed by referring to Song and Chen's paper "Asymptotic distribution-free change-point detection for data with repeated observations (2020)".

The detected change point may divide the game into several formation periods in which it is assumed that the team formation is maintained within one formation period.

In addition, FormCPD obtains a formation structure for each of the team formation periods, that is, the segments of the target session, and from this, it is possible to identify the formation in the corresponding segment. For example, the formation structure in each segment may be expressed as an average role adjacency matrix during that segment. For another example, the formation structure in each segment may be expressed as position information (eg, average position or position distribution) for each role of a scene corresponding to the main player-role assignment during the segment. In addition, the expressed formation structure can be interpreted in the language of those working in the field (eg, "4-4-2" or "3-4-3"). For example, the formation structure expressed for each segment can be interpreted as a formation identifier by clustering the average role adjacency matrix, labeling a formation identifier for each cluster, or using an artificial neural network trained with data in which position information is labeled with a formation identifier. You will be able to.

In the RoleCPD step, a change point within the formation period can be detected. At this time, the change point can be detected from a sequence of role permutation within the formation period, and various change point detection algorithms described above can be used. For example, similar to the FormCPD step, a role change point can be obtained by applying discrete g-segmentation to a sequence of permutations with a pairwise Hamming distance. Role change points can divide each formation period into several role periods. Here, the role period may refer to a period during which the role assigned to the player is maintained.

Also in the RoleCPD phase, each player is assigned one dominant role per role period to find out their actual position (e.g. "Center Forward" or "Right Midfielder") for each role period. can

Also, in the RoleCPD step, the positional transposition of the players can be calculated. Here, the position conversion may be derived from the role substitution sequence in each formation period.

On the other hand, some high-level tactical analysis can be performed using SportsCPD.

According to an embodiment of the present disclosure, analysis of a switch pattern may be performed using SportsCPD. Here, the switch may mean exchanging roles between players. Analysis of switch patterns can be performed by considering player-role assignments. In more detail and without limitation, the role exchange between players, that is, a position change, may include only a temporary role exchange between players while the same formation is maintained, except for a position change that occurs when the team formation is changed. there is. In this case, analysis of the switch pattern may be performed in consideration of player-role assignment within the formation period.

Specifically, the analysis of the switch pattern may be performed as follows.

For example, first, a main player-role assignment and a secondary player-role assignment may be acquired within a formation period. Since formations are fixed within a formation period, only one primary player-role assignment can be determined. For example, among the player-role assignments corresponding to the formation period, the most frequent player-role assignment may be determined. Next, the secondary player-role assignments can be determined. For example, among the player-role assignments corresponding to the formation period, whether the number of players excluding the main player-role assignment is within a predetermined number in order of highest frequency, whether the number is greater than or equal to the predetermined frequency, and the ratio of the occurrence frequency during the formation period Sub player-role assignment can be determined according to whether or not the number is above a certain level and the combination of these conditions. Once the primary player-role assignments and secondary player-role assignments are determined, the switch pattern can be analyzed based on them. For example, a player participating in a switch or a role participating in a switch may be determined based on a difference between a primary player-role assignment and a secondary player-role assignment, that is, a role for each player that has changed between the two assignments. As another example, the number of trials or the tendency of trials of a corresponding switch may be determined based on the occurrence frequency of secondary player-role assignment. Meanwhile, in the case of a target session having a plurality of formation periods, a switch pattern for the entire target session may be analyzed by performing the above-described process for each formation period.

As another example, a switch pattern may be analyzed based on a divided role period within a formation period. First, representative player-role assignments can be extracted for each role period. The representative player-role assignment for each role period may be the most frequent player-role assignment within the role period. If the representative player-role assignment is extracted, it is possible to determine the player and/or role participating in the switch based on the main player-role assignment of the formation period and the representative player-role assignment by role period, or to determine the number of attempts or the tendency of attempts. can Alternatively, after selecting the most frequent assignment among representative player-role assignments, the above-described information may be obtained based on other representative player-role assignments. Alternatively, the above-described information may be obtained based on representative player-role assignments adjacent to each other. Meanwhile, in the case of a target session having a plurality of formation periods, a switch pattern for the entire target session may be analyzed by performing the above-described process for each formation period.

In addition, according to an embodiment of the present disclosure, it is possible to perform detection of an irregular situation such as a set-piece using SportsCPD. For example, based on the main player-role assignment for each scene based on the main player-role assignment during the formation period, it may be determined whether the corresponding scene is an irregular situation.

SportsCPD according to an embodiment of the present disclosure may have the following several advantages. First, it is possible to detect the team's basic tactics and their changes in an unsupervised way by applying an advanced changepoint detection method to sports tracking data. Second, the formation structure and role switch for an arbitrary time interval can be effectively expressed as an average role adjacency matrix and a sequence of permutations, respectively. Third, it can be used as a practical application that can be easily understood and used by workers in the field, such as switch pattern search and automated set-piece detection.

3. Related technology

3.1 Non-parametric change point detection

Change-point detection (CPD) is a technique for finding the point at which dominant parameters in a series of processes change. CPD can be classified into a parametric method and a nonparametric method according to whether parameters are estimated in detail or points of change are directly found.

In particular, the graph-based non-parametric method is considered to be more flexible in practical application because it uses only similarity information and can be applied to high-dimensional data or even non-Euclidean data. An example of CPD is g-segmentation, which defines a graph-based scan statistic. Using this, it is possible to draw a similarity graph such as a minimum spanning tree for the observations and check the number of segments connected to the observations before and after each time t. In this case, when the number of line segments is significantly small, t may be regarded as a potential change point. In addition, techniques for correcting g-segmentation errors that occur when there is a change in both the average and the deviation or when the change point is not in the center of the sequence can be considered.

However, the above-described methods have a limitation in that they can be applied only to continuous data. For example, a similarity graph for data having repeated observations may not be uniquely determined by the graph-based method described above. To address this and obtain a unique scan statistic, a combination of multiple optimal similarity graphs or a discrete g-segment using the average of the resulting statistic can be applied. It has been experimentally demonstrated that discrete g-segmentation works well for discrete data, such as network data.

3.2 Estimation of formations and roles in team sports

The following several studies dealt with spatio-temporal tracking data to estimate team formation and find positions, that is, players' roles in sports games. Balkowski's thesis "Large-scale analysis of soccer matches using spatiotemporal tracking data (2014)" proposes a role representation algorithm that dynamically assigns a unique role to each player for each scene. This algorithm is similar to the K-means algorithm, except that the Hungarian algorithm is used in the E-stage to satisfy the constraint that no two players play the same role at the same time. Narizukawa Yamazaki's paper "Characterization of the formation structure in team sports (2017)" presents the provisional formation as an undirected graph using Delaunay triangulation and clustering the graph for a single game to characterize the formation structure. do. Narizuka and Yamazaki's follow-up paper "Clustering algorithm for formations in football games (2019)" combines role representation techniques to cluster formations across multiple matches.

The approaches described above cannot distinguish between a fixed role change by tactical instruction and a temporary role swap, which is mostly unintended from a macro perspective. In an actual team sports game environment, changes according to the coach's instruction are generally at most several times throughout the game or not, and role changes are not instructed every short time interval (eg, every second). Applying the change point detection algorithm according to an embodiment of the present disclosure to the tracking data can detect “actual” role changes between players.

4. Define the problem

In team sports, coaches often change team formations a few times during a game, or instruct players to change roles without changing formations. Therefore, their decision-making process can be traced by following two steps. First, find the point of change in formation, and then find the point of change in the next role. In this section, each step will be treated as a separate change point detection problem with reference to FIG. 1 .

1 is an example of a timeline of formation and turnover changes according to the results of SportsCPD according to an embodiment of the present disclosure. In Figure 1, the upper figure shows the team formation for each formation period. Each colored dot represents the scene-specific coordinates of a specific role normalized by the team's average position. White circles are the average positions of individual roles, and the thickness of the line connecting the circles corresponds to the corresponding element of the average role adjacency matrix. The timeline at the bottom shows the roles assigned to players per role period.

4.1 Formation Change Point Detection

The team formation can be considered using the probability distribution of the role phase position relationship. The goal of FormCPD is to find the point of change in the distribution for a set of observations in the form of bow trajectories. Therefore, it is possible to first find an effective representation of the role-phase-location relationship from raw observations, and then perform CPD using the sequence of feature factors. After finding the feature factor, the change point detection can be performed using the sequence of feature factors.

Expressing this as an equation, when the sequence of feature factors {A(t)} _t∈T representing the role phase position relationship is known, several distributions F ₁ , . . . _, T ₁ < _. <T _m can be obtained. Here, each interval T _i may be referred to as a formation period in which a team has a unique formation represented by a distribution F _i within that period.

4.2 Detecting Role Change Points

Although Balkowski's thesis described above proposes a role expression that allocates roles for each scene to field players excluding the goalkeeper for a soccer game, long-term tactical changes and temporary role exchange cannot be distinguished. Given that players typically play fixed roles directed by coaches over a period of time, the goal of RoleCPD is to find points of change where long-term tactical changes occur between before and after. That is, the formation period T _i is several time intervals T _i,1 < . . . in which players have fixed roles. <T _i,ni , which may be referred to as a role period. Then, mutual change between roles for each scene according to the result of role expression within the role period may be regarded as a temporary swap.

Expressing this as a formula, for a set P of players at game time T, a set of roles satisfying the following formula χ={X ₁ , . . . , X _N } and the player's temporary role map (P-TR map: Player-to-Temporary-Role map) {β _t :P→χ} _t∈T .

β _t (p)≠β _t (q) ∀p≠q∈P, t∈T

Here, the temporary role map {β _t } _t∈T is a player's designated role map (P-IR map: Player-to-Instructed-Role map) {α _t :χ→χ} symmetric group for _t∈T and χ ( It can be expressed as a combination of temporary role permutation {σ _t :χ → χ} _t∈T with σ _t ∈S(χ) for the symmetric group S(χ). Specifically, β _t = σ _to α _t In other words, for a P-TR map {β _t } _t∈T for a formation period T _i , T _{i ,1} < . <T _i,ni and the P-IR map {α _t } _t∈T can be found.

The first is periodwise consistency, in which the indicated role for every player is fixed in each role period T _i,j . For example, for some X _p ^{(i, j)} ∈ χ, the following equation can be satisfied.

α _t (p)=X _p ^(i,j) ∀t∈T _i,j

The second is uniqueness: no two players are assigned the same role within a single role period. For example, the following formula can be satisfied.

X _p ^(i,j) ≠X _q ^(i,j) ∀p≠q∈P

The third is the existence of role change. A change in the duration of a role can mean a change in instruction. For example, the following formula can be satisfied.

X _p ^(i,j) ≠X _p ^(i,j+1)

The last condition can only be valid for changes within a single formation period. That is, there may be identical player-role assignments in adjacent role periods of distinct formation periods. An example of this is shown in the first two role periods in FIG. 1 .

5. Methodology

This section describes in turn the detailed process of dividing a match session into multiple formation periods (FormCPD) and each formation period into multiple role periods (RoleCPD). It should be noted in advance that the methodology described below is a non-limiting example.

5.1 Detection of formation change points based on role adjacency matrix

First, for each scene, the role phase position relationship can be expressed as a role adjacency matrix calculated by Delaunay triangulation. Then, a formation change point can be found using the pairwise distance between the role adjacency matrices through discrete g-splitting, which is a non-parametric change point detection algorithm. The formations in each segment can be represented by an averaged role adjacency matrix, which is then computed by dividing the entire data set with labels used in the field, such as "3-4-3" or "4-4-2", assigned to individual formation periods. can be clustered according to

5.1.1 Creating a sequence of role adjacency matrices

First, roles may be assigned to players for each scene according to a role expression method. Here, the role expression may be exemplarily performed by referring to the concept presented in the above-mentioned Balkowski paper. An adjacency matrix can then be obtained from the role representation. Here, as an example, the adjacency matrix may be {A(t)} _t∈T ⊂ R ^N×N obtained between role labels according to Delaunay triangulation. At this time, the component a _kl (t) is the role group χ={X ₁ , . . . ,X _N }, it may be 1 if X _k and X _l are adjacent at t, and may be 0 otherwise. Figure 2 shows an example of a Delaunay graph for role positions.

2 and 3 are graphs of role adjacency according to an embodiment of the present disclosure. FIG. 2 is a temporary Delaunay graph for player trajectories in an arbitrary scene, and FIG. 3 is a weighted graph drawn from average role positions and an average role adjacency matrix for a formation period. The numbers in each circle indicate role labels 1 to 10, and the numbers under the circle in FIG. 2 are the corresponding player's uniform number.

The application of Delaunay triangulation to player tracking data was first proposed by Narizuka and Yamazaki. However, they used the back number as the index of the adjacency matrix, and applied the role representation only if the back numbers across multiple games matched a single set of indices. According to this, position exchange patterns can be explored well, but as a result player adjacency matrices will be greatly affected by irregular situations such as temporary switches or set-pieces where the original team formation collapses. can Therefore, a more robust representation of the formation structure can be obtained by using role labels as an index of the adjacency matrix instead of these player labels.

5.1.2 Application of discrete g-partitioning to matrix sequences

In order to find formation change points, CPD can be applied to sequences of role adjacency matrices obtained from each session (eg, first half or second half) of a game, respectively. Among various CPD methods, discrete g-segmentation has the distinction that it can be applied to high-dimensional or non-Euclidean data with repeated observations. Given that the sequences of the embodiments of the present disclosure are frequently repeated high-dimensional observations (e.g., a 10×10 binary matrix computed from 10 field players for soccer), there is high confidence in applying the discrete g-segment. can be expected

First, g-segmentation may be exemplarily performed as follows. First, we can construct a similarity graph G, such as a minimum spanning tree, for given observations. Then, for each time stamp, the lines are a type-0 edge connecting observations before and after t, and a type-1 edge connecting observations before t. It can be classified into three types: , and type-2 edges connecting observations after t. Accordingly, R ₀ (t), R ₁ (t), and R ₂ (t), which are the numbers of corresponding types of line segments, can be obtained. From this, the generalized line-count scan statistic can be defined as follows when Σ(t) is the covariance matrix of (R ₁ (t), R ₂ (t)) ^T under the null distribution obtained by mixing the time index .

S(t)=(R ₁ (t)- Ε [R ₁ (t)], R ₂ (t)- Ε [R ₂ (t)]) ^T Σ ^-1 (t)(R ₁ (t)- Ε [R ₁ (t)], R ₂ (t) - Ε [R ₂ (t)]

Since similar observations are more likely to be connected by line segments than distinct observations, a large S(t) can mean small intra-group distances and large inter-group distances (where "groups" is a time interval separated by t). indicated). In this way, according to the algorithm, when S(τ) exceeds a predefined threshold, τ=arg max _t S(t) may be determined as a change point.

However, the similarity graph may not be uniquely defined in g-segmentation, and thus g-segmentation may not be suitable for data processing with repeated observations. Therefore, in the discrete g-partition, instead of (R ₁ (t), R ₂ (t)), averaging statics (R _1,(a) (t), R _2,(a) (t) or combination statistics ( union statics (R _1,(u) (t), R _2,(u) (t)) can be used to overcome the drawback of the g-partition. and a similarity graph G ₀ created by the number of each subsequent observation.

When obtaining a sequence of matrices by applying discrete g-partition, the following L1,1 matrix norm (hereinafter referred to as 'Manhattan distance') can be used to measure the distance between role adjacency matrices.

d _M (A(t), A(t'))=||A(t)-A(t')|| _1,1 =Σ _k=1toN Σ _l=1toN |A _kl (t)-A _kl (t')|

Also, if the temporary role assigned to the player in each scene t is different from their most frequent role throughout the session, it can be determined that the player is switching roles in each scene t. The switch rate at t can be calculated from the number of players switching, and scenes with a high switch rate (e.g., 0.7 or higher) have ratios such as set-pieces in which players do not stay in formation at all. may be excluded under valid circumstances. In summary, discrete g-segmentation can return the predicted change points among sequences of remaining valid adjacency matrices using pairwise Manhattan distances.

Once the change points are found for a given time period, the algorithm can determine their significance according to the three conditions individually or in combination. As a specific example, (1) the p-value of the scan statistic is less than 0.01, (2) two segments last more than a minimum time (eg, 5 minutes), and (3) the calculated Only when the average role adjacency matrices are far enough from each other (i.e., the Manhattan distance is large) can the estimated change point τ be considered significant. The critical distance for (3) can be empirically set to 7.0.

Since there may be more than one formation changepoint during a session, a recursive framework can be built to find multiple changepoints. First, if there is a significant change point within a given period, the CPD algorithm can be applied to sequences before and after the change point, respectively. Each branch CPD (branch Change-Point Detection) may end when there are no significant change points in the segment of interest. Consequently, for a given session T, several formation periods T ₁ <... <T _m .

5.1.3 Formation clustering based on the average role adjacency matrix

Based on the above, the formation in the formation period T _i is the average role position V(T _i )=|T' _i | ^-1 Σ _t∈T'i V(t), and as an edge matrix, the adjacency matrix A(T _i )=|T' _i | When ^-1 Σ _t∈T'i A(t), it can be calculated as a weight graph F(T _i )=(V(T _i ), A(T _i )). Here, T′ _i is a set of effective time points with a low switch rate among T _i , and V(t)=(v ₁ (t), …, v _N (t)) ^T ∈R ^N×2 is each scene t is a two-dimensional position of N roles normalized such that the average is '0' (eg, Σ _{k = 1toN} v _k (t) = (0,0)) in . 3 is an example of a formation graph.

Here, the distance between arbitrary formation pairs may be defined by arranging the roles and calculating the Manhattan distance between the aforementioned role adjacency matrices. Specifically, let's assume that any pair of formations F=(V, A) and F'=(V', A') in the formation period. Here, F and F' do not necessarily belong to the same game, and they may be calculated in different games. Next, the rows and columns of A can be rearranged into the rows and columns of A'. For example, a Hungarian algorithm can be used for rearrangement. In this case, the pairwise Euclidean distance between V and V' ) can be used. In other words, the assignment cost c(V, V'; Q)=Σ _k=1toN ||(QV) _k -v' _k || An optimal permutation matrix Q that minimizes ₂ can be calculated. For example, the distance between formations F and F' is d(F, F')=d _M (QAQ ^T _, A') QAQ ^T and, as in the case of A' can be used.

Depending on these distances, we can perform clustering of formations to determine whether the formations are identical or different using the entire data set. If agglomerative clustering is applied to the pairwise formation distance, the formation graph can be grouped into a plurality of clusters. It was labeled in English familiar to those working in the field.

4 relates to an example formation group according to an embodiment of the present disclosure. Figure 4 shows the average role position of each formation group, and terms used in the field of soccer are labeled for this, and the numbers in parentheses are the proportion of the total playing time. The result of FIG. 4 is calculated using positional data obtained from GPS data as raw data, that is, player tracking data. Specifically, the results of FIG. 4 show that for 809 sessions of 2.33 million seconds, SportsCPD is applied to divide them into 866 formation periods and 2,158 role periods, and then cluster the 866 formation periods into 32, which It was merged again into a group of 7 formations.

While the method proposed by Balkowski uses Earth Mover's Distance (EMD) between role locations, the clustering according to the above-described embodiment of the present disclosure may use the average distance between role adjacency matrices. In addition, in the prior art, only formations for half of the game with an almost constant length of time were dealt with, but in the embodiment of the present description, the formation period is shorter than this or the length varies widely, which is likely to cause distortion of the formation graph. You can handle the formation. This is because the average role adjacency matrix based on adjacency relationships rather than absolute position is more robust against such distortions.

5.2 Role Change Point Detection Based on Role Substitution Sequence

RoleCPD can be performed in a similar way to FormCPD, except that a role permutation sequence is used instead of a role adjacency matrix. For each formation period, a discrete g-segment can be applied recursively to find role changes. Next, the original role indicated to each player for each role period can be found, and the player-role assignment for each scene can be expressed as a substitution sequence of the designated role.

5.2.1 Generation of role substitution sequences

In the above-described role expression, each player p may be assigned an initial role X _p according to a standard number such as a back number, for example. Using this, the players at time t p ₁ , … , the temporary role assignment for p _N is the substitution matrix for the initial role assignment (X _P1 , … , X _PN ) π _t =(π _t (X _P1 ), … , π _t (X _PN ))∈S( χ) can be expressed as Expressing this again as a formula, it can be as follows.

β _t (p)=π _t (X _P )

5.2.2 Application of discrete g-partitioning to sequences of substitutions

Similar to the application to the role adjacency matrix, g-segmentation can be applied to the sequence of role substitution sequences {π _t } _t∈Ti for each formation period T _i . According to this, the Hamming distance d _H (π _t, π _t' )=|{X:π _t (X)≠π _t' (X), X∈χ}| is used as the distance between the role substitution sequences A predicted change point can be determined among sequences of effective substitution sequences (for example, substitution sequences when the switch rate is 0.7 or less).

Inspection of significant change points can also be performed similarly to FormCPD, where the last condition applied to FormCPD can be excluded. Since the goal of RoleCPD is to find when the main role assignment is changed, the change point τ can be judged to be significant only when the most frequent substitution sequence before and after τ is different for the change point τ.

Finally, by applying a recursive CPD to the sequence in each formation period T _i , the partition T _i,1 < . . . of T _i . <T _i,ni can be obtained.

5.2.3 Derivation of designated roles for players by role period

For each role period T _{i,j ,} the most frequent permutation sequence for the player can be set as the designated role, and all temporary roles obtained from the role expression can be expressed as the permuted sequence of the designated role. Expressed as a formula, the P-IR map {α _t :P→χ} _t∈Ti,j is exemplarily, π _(i,j) ∈S(χ) as {π _t } _t∈Ti,j As a frequent substitution sequence, it can be expressed as α _t (p)=π _i,j (X _p ), which is a fixed value at t∈T _i,j . In t∈T _i,j, the role substitution sequence σ _t (X _p )=β _to α ^-1 can be obtained as follows.

σ _t (X _p )=β _t( α _t ^-1 (X _p ))=π _t (π _(i,j) ^-1 (X _p ))

Accordingly, the P-IR map may satisfy the three conditions described above. Specifically, α _t (p) is constant at t∈T _i,j and it is distinct across p∈P because it is a substitution sequence with a distinct initial role and thus has temporal consistency and uniqueness. In addition, the existence of a role change between adjacent role terms can be secured due to the significance test.

6. Experiment

The accuracy of SportsCPD can be evaluated by comparing the result according to the embodiment of the present disclosure with the measured value recorded by an expert in the field. In addition, according to an embodiment of the present disclosure, an interpretation technique for identifying play patterns of a team in terms of a role switch may be provided. In addition, the switch rate according to the embodiment of the present disclosure can be used as a simple and highly reliable index for set-piece detection.

6.1 Model evaluation

The performance of SportsCPD can be measured in two aspects. One is the accuracy of prediction formation by FormCPD, and the other is the robustness of role change point detection according to RoleCPD. Regarding the former, it is possible to calculate the ratio of correctly detected segments per minute to the total number of segments (eg, total play minutes). Regarding the latter, it is measured by how accurately RoleCPD can detect the change point, given that change point detection is about the classification of each time point between the change point and the non-change-point. You can try it.

For both evaluations, ground truth labels recorded by video analysts or subject matter experts working as coaches for professional or semi-professional sports teams may be used. Since it is very exhausting to record formations and roles by minutes, the recording of one game by an expert was used for the evaluation according to the embodiment of the present disclosure.

For formation prediction, comparison was made with labels for each minute segment instead of each formation period, because the method according to the embodiment of the present disclosure and the change point detected by a human expert are slightly different. Out of a total of 2,680 minutes of playing time, the record for 2,047 minutes and the result of FormCPD were consistent, showing a degree of agreement of 76.4%. 5 shows detailed results as a chaos matrix.

5 is a diagram related to formation prediction according to an embodiment of the present disclosure. Figure 5 expresses the accuracy of formation prediction as a confusion matrix, which compares the predicted formation in minutes with the actual recorded formation.

On the other hand, RoleCPD detected 96 change points in 28 games, 86 change points were located within 5 minutes of the recorded change points, and there were a total of 7 missing change points (false negatives). Therefore, the recall points and F1 scores for accuracy of RoleCPD are 0.8958 (86/96), 0.9247 (83/93), and 0.9101 (2·(precision×recall)/(precision+recall)), respectively.

6.2 Switch Pattern Search

The role substitution sequence obtained as described above may indicate a temporary role switch between players. In the case where all players maintain their assigned original roles at time t, the role permutation sequence σ _t can be an identity permutation. In other words, the non-identity role substitution sequence may mean that a switch play occurred at that time. Therefore, by analyzing the non-identity role substitution sequence, the play pattern of the team can be identified.

6 is a graph of role substitution sequences for each formation period according to an embodiment of the present disclosure. In FIG. 6, the period of the top 5 role substitution sequences with the highest frequency for each formation period is expressed in seconds. 7 is a diagram related to an example of tactic analysis using a role substitution sequence according to an embodiment of the present disclosure.

Referring to the results of FIGS. 6 and 7 for the game shown in FIG. 1 , it is clear that a team's tactical pattern can be identified from the role substitution sequence for a specific game. Figure 6 shows the most frequent 5 role substitution sequences by formation period, and the following analysis was performed from them. In the following analysis, the formation period and the role period are referred to as FP (Formation Period) and RP (Role Period), respectively, and a player assigned a specific role k will be denoted as Rk.

However, players actively perform role switches in FP1 (in this interval, the sum of the time of the top 3 role permutation sequence is 140 seconds out of 18 minutes), whereas in FP3 (the top 3 most frequent role in this interval) The total time of the heat of substitution was only 35 seconds out of the total 13 minutes), almost no switch was made.

Second, the role substitution sequence (3 5) occurred frequently in FP1, which is a "false-9 play" in which the center forward shown in Figure 7 falls into the back of the midfielder to provide space for fellow players. " means However, as the team formation changed to 4-3-3 in FP2, the play was reduced.

Third, unlike other FPs, in FP1, the fullbacks (R1, R8) took on the attacking role along the side and actively overlapped with the wingers (R1, R10). Here, it is revealed that the overlap is different on both sides. On the left side, R6 covers the position of R8, resulting in a 3 cycle of (6 8 9), while on the right side, R4 does not cover the position of R1. A two-factor cycle of (1 10) occurred.

Fourth, in FP2, left winger R9 (6 9) repeatedly attempted a cut-in play towards the penalty box via a switch.

Fifth, in FP4, after adding the center forward R5, the team can switch between the center forwards (R3 R5) or between the central midfielders (R4 R6) except for the usual switches (i.e. switches within the same position). kept the formation constant.

6.3 Set-piece detection

In a sports game, a set-piece may refer to a situation in which a game is stopped and resumed again. In particular, in soccer, in set-piece situations, you can send the ball into the scoring area, and set-pieces such as free kicks and corner kicks are considered good opportunities to score points. The review is meticulous, and special tactical training for set play is also conducted.

SportsCPD's output automatically detects set-piece situations using statistical switch rates, allowing sports teams to easily extract and manipulate set-play data. The term switch rate has been defined as the ratio of players switching in the above, but it can also be defined as a value obtained by dividing the distance to the identity substitution sequence (eg, Hamming distance) by the number of roles. Because players are completely mixed up in set-piece situations, switch rates in these situations can be close to 1.0. Thus, using the switch rate provides a simple, completely unsupervised, yet highly accurate set-piece detection model. For example, a set-piece may be detected by selecting a situation where the switch rate is greater than or equal to a threshold value (eg, 0.9). 8 shows a strong cross-correlation between switch rate and set-piece.

8 is a graph illustrating a relationship between a switch rate and set-piece occurrence according to an embodiment of the present disclosure. FIG. 8 is a time-sequential diagram of switch rates and set-piece occurrences during the first half of the game shown in FIG. 1 . In FIG. 8, "C" and "F" indicate a corner kick and a free kick that can be kicked into the opponent's penalty box, respectively. The white square means the case where the team that calculated the switch rate in the situation is the attacking team, and the black square means the case the defending team.

7. Implementation

Hereinafter, a device, system, and method for implementing SportsCPD according to an embodiment of the present disclosure described above will be described.

9 is a block diagram of a system for providing tactical analysis of team sports according to an embodiment of the present disclosure.

Referring to FIG. 9 , a system 1000 may include a sensor device 1200 , an analysis device 1400 and a user device 1600 . Hereinafter, the components of this system will be described in more detail.

The sensor device 1200 may be provided in the form of an attachable device. Here, the term "attachable device" may refer to a device directly or indirectly attached to the player. For example, the attachable device may be a pocket device inserted into a player's clothes or a wearable device in the form of a band wrapped around a sports player's body part, such as a wrist or ankle. Since the sensor device 1200 in the form of an attachable device is attached to each player, there may be a plurality of them in team sports.

For example, the sensor device 1200 may perform global localization on a player. In this case, the player tracking module 1220 may be a global navigation satellite system (GNSS) module including a global positioning system (GPS) module. For example, the GPS module performs global positioning of the player by calculating the GPS global position (eg, latitude, longitude) through a triangulation technique by a GPS processor using GPS signals received from navigation satellites through a GPS antenna. can do. The calculated global location may be transmitted to the analysis device 1400 through the communication module 1240 . Accordingly, the analysis device 1400 may obtain a player tracking data set. In addition, during this process, the analysis device 1400 may convert the global position received from the sensor device 1200 into a field coordinate system.

As another example, the sensor device 1200 may perform local positioning on a player. In this case, the sensor device 1200 may operate as a tag node of a Local Positioning Network (LPS) and transmit/receive an LPS signal with an anchor node fixedly installed in a positioning area. In the case of performing global positioning, the sensor device 1200 can independently perform positioning, whereas in the case of performing local positioning, the sensor device 1200 simply functions as a transmitter/receiver for the LPS signal of the local positioning network. And the positioning result may be calculated in an external device. For example, the sensor device 1200 may transmit or receive an LPS signal to an anchor node, and positioning of a player may be performed by an external device using a result of transmitting and receiving an LPS signal between nodes in a network. At this time, the analysis device 1400 receives the result of transmitting and receiving the LPS signal and directly performs positioning to generate player tracking data, or may receive the positioning result from a separate external device that performs positioning using the result of transmitting and receiving the LPS signal. .

Meanwhile, unlike shown in FIG. 9 , in the tactical analysis system 1000, the sensor device 1200 may be replaced with a video analysis platform. The video analysis platform may be composed of a camera that captures an image and a video analysis device that calculates the position of an object (eg, a player or a ball) in an image through video analysis using top-view conversion and deep learning algorithms. . At this time, the analysis device 1400 may directly receive an image from a camera and perform image analysis to generate a player tracking data set, or may receive a positioning result from a separate external device that performs image analysis.

The analysis device 1400 may perform various operations described above in relation to SportsCPD according to an embodiment of the present disclosure. Basically, unless otherwise noted, various operations described above in relation to SportsCPD can be interpreted as being performed by the analysis device 1400, and among them, it is revealed that they can be interpreted as being performed by the controller 1440. . For example, the analysis device 1400 receives information directly or indirectly reflecting the player's position from the sensor device 1200, generates a player tracking data set therefrom, and uses it to generate the above-described SportsCPD and related data sets. It can perform various applications (detection of irregular situations, etc.). This analysis device 1400 may be provided in the form of a personal computer or a local or remote server, and may not necessarily be physically provided as a single entity.

The analysis device 1400 may include a communication module 1420 , a controller 1440 , and a memory 1460 .

The communication module 1420 may transmit/receive data between the analysis device 1400 and an external device (eg, the sensor device 1200 or the user device 1600). For example, the analysis device 1400 may collect data from an attachable device through the communication module 1420, receive an image from a camera, or transmit various types of information to the user device 1600 through the web.

The controller 1440 may control overall operations of the analysis device 1400 . The controller 1440 may be implemented as a hardware configuration, a software configuration, or a combination thereof. From a hardware point of view, the controller 1440 may be provided in various forms capable of performing calculations or data processing, including an electronic circuit, an integrated circuit (IC), a microchip, and a processor. In addition, since the physical configuration of the controller 1440 is not necessarily limited to a single physical entity, the controller 1440 is a single processor that integrates all processes of the analysis device 1400 or It may be provided as a plurality of processors each performing different functions, or may be provided in a form combined with some of the other components of the analysis device 1400.

The memory 1460 may store various data related to the operation of the analyzer 1400 . The memory 1460 may be provided with various volatile and non-volatile memories.

The user device 1600 may function as a user interface that provides various data or information collected or calculated by the system 1000 to the user or receives user input from the user. For example, the user device 1600 receives a tactical analysis result from the analysis device 1400 through the communication module 1620 under the control of the controller 1640 and provides it to the user as visual information through the display 1660. can do. The user device 1600 may be a smart device such as a smart phone or tablet, a personal computer such as a laptop or desktop, or an electronic device similar thereto.

Hereinafter, various examples of the aforementioned analysis method according to an embodiment of the present disclosure will be described with reference to FIGS. 10 to 18 . 10-18 are flow charts of examples of methods for providing tactical analysis of team sports according to embodiments of the present disclosure.

The methods described below may be used alone or in combination with each other, and since not all of the steps of the methods are essential, the methods may be performed including some as well as all of the steps. In addition, since the order in which each step in the method is mentioned is only for convenience of explanation, the steps do not necessarily have to be performed in the order in which they are described. In addition, hereinafter, each method will be described based on what is performed by the above-described system and device, but since this is also only for convenience of explanation, each method is a tactical analysis system 1000 and a tactical analysis device 1400 It is not something that can only be done by

Referring to FIG. 10 , an example of a method for analyzing tactics of team sports according to an embodiment of the present disclosure includes obtaining a player tracking data set (S1100), obtaining player-role assignments (S1200), and assigning roles. It may include a step of obtaining (S1300), a step of detecting a change point from a sequence of role arrangements (S1400), and a step of dividing a target session into formation periods (S1500). Hereinafter, each step of this example will be described.

The analysis device 1400 may obtain a player tracking data set (S1100).

Specifically, the controller 1440 may receive a player tracking data set from the sensor device 1200 through the communication module 1420 . When the sensor device 1200 performs global positioning, the analysis device 1400 may receive player position information expressed as a global position during the target session and convert it into a stadium coordinate system. When the sensor device 1200 performs regional positioning, the analysis device 1400 may receive a transmission/reception result of the LPS signal from the local positioning network, calculate the position of the player using the LPS signal, and obtain a tracking data set. Alternatively, the analysis device 1400 may obtain a player tracking data set by receiving an image captured by a camera and calculating a player position therefrom, or by receiving a player position calculated by an external device from a camera image.

On the other hand, the player position corrects the position of each player using the positions of all or part of the players participating in the target session in the scene (for example, the same team, field players excluding the goalkeeper, etc., or a combination thereof) can do. For example, each player's corrected position may be corrected as an average position in a corresponding scene of field players of the same team.

Next, the analysis device 1400 may use the player tracking data set to obtain player-role assignments (S1200). For example, the controller 1440 initially assigns a single random number to the players throughout the session. Allocate roles, obtain position distribution of roles based on this, perform player-role assignment for each scene in consideration of position distribution of roles, recalculate position distribution of roles in consideration of player-role assignment for each scene again, The player-role assignment can be obtained through a process of re-executing the player-role assignment for each scene based on the recalculated position distribution. In this process, the controller 1440 may exclude scenes corresponding to non-regular situations when calculating the location distribution of roles. Whether the scenes correspond to non-regular situations will be described in detail in an embodiment to be described later.

Next, the analysis device 1400 may obtain a role arrangement using the player-role assignment (S1300). For example, the controller 1440 may obtain the role arrangement using the location distribution of the roles, which has been described in detail above.

When the role arrangement is acquired, the analysis device 1400 may detect a change point from the sequence of the role arrangement (S1400), and may divide the target session into formation periods using this (S1500). Specifically, the controller 1440 may apply a change point detection algorithm to a sequence of role assignments to detect one change point, and based on this, divide the target session into pre- and post-periods, and divide the divided periods again. The formation period can be obtained by performing re-division by applying the changing point detection algorithm. Furthermore, the controller 1440 may determine whether the change point detected by the change point detection algorithm is valid based on several conditions, which have been described in detail above.

Referring to FIG. 11 , an example of a team sports tactics analysis method according to an embodiment of the present disclosure includes obtaining a player tracking data set (S2100), obtaining a player-role assignment (S2200), and a player-role assignment (S2200). It may include detecting a change point from the role assignment sequence (S2300), and dividing the target session into role periods (S2400). Hereinafter, each step of this example will be described.

First, the analysis device 1400 may obtain a player tracking data set (S2100). A process of acquiring player tracking data may be performed similarly to step S1100. However, at this time, the player tracking data set may be related to the entire target session, or may be a single formation period. If the entire session is divided into a plurality of formation periods, the method of this example may be repeatedly performed for each formation period.

Next, the analysis device 1400 obtains the player-role assignment (S2200), detects a change point from the sequence (S2300), and divides the target session into role periods (S2400). Specifically, the controller 1440 may obtain the player-role assignment for each scene similarly to that described in step S1200, and may determine the changing point by applying a changing point detection algorithm to the sequence. Even at this time, the controller 1440 may determine the validity of the change point through several conditions (detailed descriptions thereof have been described above), and may divide the role period based on the valid change point.

Referring to FIG. 12, an example of a method for analyzing tactics of team sports according to an embodiment of the present disclosure is, in the method according to FIG. 11, the main player for the role period-determining the role assignment (S3500), Determining whether the role assignment is a non-regular player-role assignment (S3600), calculating a switch rate for the non-regular player-role assignment (S3700), and detecting a non-regular situation in the target session based on the switch rate (S3800) It may further include a step including. Hereinafter, each step of this example will be described.

In this example, the analysis device 1400 may determine the main player-role assignment for each session divided into role periods (S3500). Specifically, the controller 1440 may determine the most frequent player-role assignment during the role period as the main player-role assignment.

When the main player-role assignment is determined, the analysis device 1400 determines whether the player-role assignment for each scene is a non-regular player-role assignment based on this (S3600), and calculates a switch rate for the non-regular player-role assignment. (S3700) Specifically, the controller 1440 may determine whether the corresponding scene corresponds to an irregular situation based on the difference in player-role assignment for each scene with respect to the main player-role assignment. For example, when a player-role assignment is expressed as a role permutation sequence, when the distance between the two (eg, Hamming distance, etc.) is greater than or equal to a threshold value, or when the two are different, the player-role assignment is converted to a non-regular player-role. It can be judged by assignment. The controller 1440 may calculate a switch rate for the non-regular player-role assignment, and the switch rate is a value reflecting the difference between the non-regular player-role assignment and the main player-role assignment. A detailed description of this is described above. there is a bar

If the switch rate for each scene is calculated, the analysis device 1400 may detect an irregular situation in the target session based on it (S3800). For example, the controller 1440 may determine a scene having a switch rate equal to or greater than a threshold value as an irregular situation.

Referring to FIG. 13 , an example of a method for analyzing team sports tactics according to an embodiment of the present disclosure is, in the method according to FIG. 10 , obtaining location information for each role regarding a formation period (S4600), and for each role A step of identifying a formation based on location information (S4700) may be further included. Hereinafter, each step of this example will be described.

In this example, when the formation period in which the analysis device 1400 divides the target session is obtained, position information for each role may be obtained (S4600). Here, the acquisition of the formation period may be performed, for example, by the analysis device 1400 according to the method described with reference to FIG. 10 , and the analysis device 1400 based on this obtains location information for each role over the entire session. can be obtained For example, the controller 1440 may refer to the player location tracking data set and the player-role assignment for each scene, determine the location of the role for each scene, and obtain location information for each role throughout the session using the role for each scene. can Here, the location information for each role may be a location distribution or average location of roles over the entire session.

Now, the analysis device 1400 may identify a formation based on location information for each role (S4700). For example, the controller 1440 inputs the location information of the role into a deep learning algorithm trained using training data labeled with a formation identifier recognizable by a worker in the relevant field to determine the formation of the corresponding session. can As another example, location information of roles obtained from different sessions may be clustered, a group including the location information of this session may be determined, and a formation identifier assigned to the group may be determined as a formation identifier for this session. A detailed description of this has already been described above.

Referring to FIG. 14, an example of a method for analyzing tactics of team sports according to an embodiment of the present disclosure is, in the method according to FIG. The method may further include normalizing location information for each role using the relevant role arrangement (S5700) and identifying formations based on the normalized location information for each role (S5800). Hereinafter, each step of this example will be described.

In this example, when the formation period in which the analysis device 1400 divides the target session is obtained, location information for each role may be obtained (S5600), and location information for each role may be normalized using the role arrangement for the formation period. Yes (S5700). Here, normalization may be performed by reallocating roles according to location information for each role obtained in different formation periods. Given that the initial role assignments in the player-role assignments in this description are arbitrary, the player-role assignments defined in this session may be different from the player-role assignments defined in other sessions. That is, for example, when the player-role assignment is expressed as a role substitution sequence, since the order of vectors between different sessions may not match, it is necessary to normalize the order within vectors between similar roles. Compatibility between player-role assignments between different sessions (eg, different games) can be ensured by using location information and role arrangement for each role. This task may be performed by the controller 1440, and for this purpose, the analysis device 1400 clusters or classifies data on player-role assignments, role arrangements, and position distribution by role obtained from various sessions in advance. Through this, it is possible to use an algorithm that can match the order of players in the player-role assignment.

Now, the analysis device 1400 may identify a formation based on location information for each role (S5800). For example, the analysis device 1400 may identify a formation for a corresponding session using a deep learning algorithm from location information for each role or a classification algorithm obtained in a clustering process. Since a detailed description of this has been described above, it will be omitted here.

Referring to FIG. 15 , an example of a team sports tactics analysis method according to an embodiment of the present disclosure includes obtaining a player tracking data set (S6100), obtaining a player-role assignment (S6200), and a main player - Obtaining a role assignment (S6300), and identifying a role for each player based on the main player-role assignment (S6400). Hereinafter, each step of this example will be described. Hereinafter, each step of this example will be described.

In this example, the analysis device 1400 may obtain a player tracking data set (S6100), obtain a player-role assignment (S6200), and obtain a main player-role assignment (S6300). Here, the player tracking data set is related to the entire session or the formation period, and in the latter case, the role of each player maintained constant during the formation period is obtained.

Next, the analysis device 1400 may identify a role for each player based on the main player-role assignment (S6400). Accordingly, the role of each player can be finally interpreted as a position name that can be identified by those working in the field. Specifically, the controller 1440 may determine the main player's main role according to the main player-role assignment, determine the location information of the role as the player's location information, and identify the player's position therefrom. In this case, a deep learning algorithm that receives position information of individual players or other players and outputs a position label may be used, or other classification algorithms may be used. In addition, in team sports having a more complex formation, the formation identifier according to the above example may be additionally used.

Referring to FIG. 16 , an example of a method for analyzing tactics of team sports according to an embodiment of the present disclosure includes obtaining sub player role assignments in the method according to FIG. 15 (S7400), and main player-role assignments and sub players. It may include obtaining role change information in consideration of player role assignment (S7500). Hereinafter, each step of this example will be described.

In this example, when the main player role assignment is completed according to the above-described several examples, the analysis device 1400 obtains the sub player role assignment (S7400), and the role is switched in consideration of the main player-role assignment and the sub player role assignment. Information can be obtained (S7500). Here, the secondary player role may be a player-role assignment with a high frequency except for the main player role. The controller 1440 may detect the sub player-role assignment for the session by applying the above-described sub player role determination condition.

Next, the analysis device 1400 may acquire role change information in consideration of the main player-role assignment and the secondary player-role assignment. For example, controller 1440 may determine how often two player-role assignments occur based on the frequency of primary and secondary player-role assignments, or in primary and secondary player-role assignments. In consideration of the changed role of each player, roles or players whose roles are to be switched can be determined. In addition, a detailed description of the role change information has been described above.

Referring to FIG. 17 , an example of a team sports tactics analysis method according to an embodiment of the present disclosure includes acquiring an image of a target session (S9100) and acquiring a player tracking data set of the target session (S9100). It may include (S9200), obtaining tactical analysis information from the player tracking data set (S9300), and bookmarking the video timeline using the tactical analysis information (S9400). Hereinafter, each step of this example will be described.

First, the analysis device 1400 may obtain an image of the target session (S9100) and a player tracking data set (S9200). Here, the image is an image captured by an external camera and may be an image of a sports game or training, and the player tracking data set may be player location information for the same session.

Next, the analysis device 1400 may obtain tactical analysis information from the player tracking data set (S9300). Here, the tactical analysis information may include the above-described role switching information, whether or not it is an irregular situation, and information about a formation period or a role period. For example, the controller 1400 uses the information obtained from the above-described several examples (eg, the main player-role assignment and the secondary player-assignment comparison result) to determine when a position switch between players occurs or when a position switch occurs. The time period for which it is performed can be specified. For another example, the controller 1400 may specify the occurrence time or occurrence period of the irregular situation through detection of the irregular situation based on the switch rate. For another example, the controller 1400 may detect a formation change point or a role change point to specify a formation period or a role period.

When the tactical analysis information is acquired, the analysis device 1400 may perform bookmarking on the video timeline using the tactical analysis information. For example, the controller 1440 may bookmark time information related to a position switch, time information related to a role period, time information related to a formation period, and the like to a time line of a video. To this end, the analysis device 1400 may perform time synchronization between the video and the player tracking data set in advance. On the other hand, here, when bookmarking, information identifying the type of event that occurred (for example, "overlap of fullback", "forward-midfielder position swap", "corner kick", etc.) is not simply marked with a timestamp. can also be displayed together.

On the other hand, it should be noted in advance that the methods according to the embodiments of the present disclosure described above are described/shown as using some prior methods according to other embodiments of the present disclosure only for convenience of description.

Embodiments of the present disclosure described above may be provided as a computer software program that implements all or part of the above-described method or operation. For example, a computer software program may include program code provided in a form tangibly recorded on a machine-readable medium, and the program code may include instructions for executing steps or operations of a method according to an embodiment of the present disclosure. can include Such a computer program may be downloaded and installed from a network through a communication unit, and/or installed from a removable medium, and may be executed by a processor capable of reading the program code, so that the embodiments of the present description Followed steps or actions can be performed.

In addition, the embodiments of the present description described above may be provided in the form of a computer-readable recording medium on which all or part of the above-described software program is recorded. The type of recording medium may be a conventional memory device such as a floppy disk, a hard disk, a CD-ROM, or a memory card, or may be downloaded through the Internet or a computer communication network. You can take it and use it.

8. Conclusion

An embodiment according to the present disclosure provides a changepoint detection framework, SportsCPD. SportsCPD can distinguish tactically intended changes in formations and roles in team sports from temporary changes. First, the temporary role position phase relationship and position transition can be expressed as a binary matrix and a sequence of permutations, respectively. Using this, a non-parametric change point detection algorithm (eg, discrete g-segment) can be applied to high-dimensional or non-Euclidean data with frequent repetitions to find formation and role assignment change points. Since the concept of formation and role is the most basic and intuitive way to express the tactics of a sports team, tracking and summarizing changes in the team's formation and role is valuable in itself to practitioners in the field. Furthermore, a switch pattern search or set-piece detection may be performed using additional information such as a temporary substitution sequence. Accordingly, embodiments of the present disclosure may be widely used in the field of team sports.

Embodiments disclosed in this description are only illustrative of the technical idea of the present invention, so the scope of the technical idea of the present invention is not limited to these embodiments. Therefore, the embodiments of the present disclosure described above are not only individual implementations and combinations, but also various modifications and variations that can be made to those skilled in the art without departing from the essential characteristics of the present invention. It will be understood that it belongs to the technical spirit of the present invention. Therefore, the protection scope of the present invention should be construed by the following claims, and all technical ideas within the equivalent range should be construed as being included in the scope of the present invention.

1000: tactical analysis system
1200: sensor device
1400: analysis device
1600: user device

Claims (12)

obtaining a plurality of player tracking data sets for each of the plurality of players, each player tracking data set including positions of the corresponding player at a plurality of points in time within the target session;
acquiring a plurality of position distributions for each of the plurality of players, each position distribution being generated based on a corresponding player tracking data set;
assigning a plurality of role identifiers to each of the plurality of location distributions;
a plurality of player-role assignments for each of the plurality of time points, each assignment reflecting a relationship between the plurality of players and the plurality of role identifiers at a first corresponding time point, and the plurality of location distributions and Obtaining - generated by assigning each of the plurality of role identifiers to the plurality of players based on positions of the plurality of players at the first corresponding time point among the plurality of player tracking data sets;
A plurality of role assignments for each of the plurality of viewpoints, each role assignment reflecting a positional relationship between the plurality of role identifiers at a second corresponding viewpoint, and a player-role at the second corresponding viewpoint obtaining an allocation and generated based on the positions of the plurality of players at the second corresponding point in time; and
Dividing the sequence of the plurality of role assignments into at least two role placement groups using a changepoint detection algorithm;
A method of providing tactical analysis of team sports.
According to claim 1,
obtaining a primary player-role assignment from the plurality of player-role assignments; and
Further comprising determining at least one non-regular player-role assignment among the plurality of player-role assignments based on the main player-role assignment;
The sequence of the plurality of role assignments includes only the remainder of the plurality of role assignments except for the at least one non-regular player-role assignment.
A method of providing tactical analysis of team sports.
According to claim 1,
the target session to at least two time periods corresponding to each of the at least two role placement groups, the at least two time periods comprising a first time period in which a primary formation of a team participating in the target session is a first team formation; and a second period of time wherein the primary formation of the team is a second team formation.
A method of providing tactical analysis of team sports.
According to claim 3,
determining the first team formation from at least one first role assignment during the first time period and the second team formation from at least one second role placement during the second time period.
A method of providing tactical analysis of team sports.
obtaining a plurality of player tracking data sets for each of the plurality of players, each player tracking data set including positions of the corresponding player at a plurality of points in time within the target session;
acquiring a plurality of position distributions for each of the plurality of players, each position distribution being generated based on a corresponding player tracking data set;
assigning a plurality of role identifiers to each of the plurality of location distributions;
a plurality of player-role assignments for each of the plurality of time points, each assignment reflecting a relationship between the plurality of players and the plurality of role identifiers at a first corresponding time point, and the plurality of location distributions and Obtaining - generated by assigning each of the plurality of role identifiers to the plurality of players based on positions of the plurality of players at the first corresponding time point among the plurality of player tracking data sets; and
obtaining a primary player-role assignment from the plurality of player-role assignments;
A method of providing tactical analysis of team sports.
According to claim 5,
The main player-role assignment is the most frequent player-role assignment among the plurality of player-role assignments.
A method of providing tactical analysis of team sports.
According to claim 5,
Further comprising determining a plurality of positions for each of the plurality of players based on the main player-role assignment.
A method of providing tactical analysis of team sports.
According to claim 5,
selecting a secondary player-role assignment from the plurality of player-role assignments;
A method of providing tactical analysis of team sports.
According to claim 8,
The secondary player-role assignment is the second most frequent player-role assignment among the plurality of player-role assignments.
A method of providing tactical analysis of team sports.
According to claim 8,
Comparing the main player-role assignment and the secondary player-role assignment to obtain information about position conversion; further comprising
A method of providing tactical analysis of team sports.
According to claim 5,
Acquiring switch rate information about the plurality of player-role assignments based on the main player-role assignment; further comprising
A method of providing tactical analysis of team sports.
According to claim 11,
Detecting an irregular situation in the target session based on the switch rate information; further comprising
A method of providing tactical analysis of team sports.

Images