CN121527685B - A method and system for predicting sports event training based on AI and image analysis - Google Patents

Abstract

This invention provides a method and system for predicting sports event training based on AI and image analysis. The invention relates to the field of sports event training prediction technology. The method includes: acquiring badminton match videos of athletes; determining multiple hitting errors using an error point determination model based on the badminton match videos; determining a K-value based on the multiple hitting errors; clustering the multiple hitting errors and the K-value to determine K error regions; constructing an error region map; processing the error region map using a graph neural network to determine at least K hitting training points, wherein each error region has at least one hitting training point; and determining a target training plan based on the at least K hitting training points. This method can accurately formulate the optimal targeted training plan based on the athlete's badminton match errors.

Description

Sports event training prediction method and system based on AI and image analysis

Technical Field

The invention relates to the technical field of sports event training prediction, in particular to a sports event training prediction method and system based on AI and image analysis.

Background

In athletic badminton, the training pertinence of sportsmen is directly related with the playing performance, and the accurate control of the mispoints of playing the ball is the core key for improving the training efficiency. The current badminton training is based on subjective observation of coaches or experience summary after the contestant contests, and needs to repeatedly repeat the past athlete badminton contests in a large number, thus being labor-consuming, low in efficiency, and lacking in the depth of actual error data in the contests and the mining and analysis of the system. The traditional training mode often is difficult to accurately and comprehensively capture scattered error points of athletes in a match, so that the training plan is made to lack accurate basis, and further the problems of blind training and fuzzy training key points are easy to occur. Especially for the individualized mistake characteristics of different athletes, the current mode is difficult to realize the pertinence focus, not only wastes training resources, but also is difficult to make up for the short board of technique fast. Meanwhile, the existing training schemes are mostly fixed template contents, can not be dynamically adjusted in combination with error conditions in badminton games of athletes, and are difficult to meet accurate and personalized training requirements of the athletes.

Therefore, how to accurately determine the optimal targeted training scheme according to the error of the badminton match of the athlete is a current urgent problem to be solved.

Disclosure of Invention

The invention mainly solves the technical problem of accurately determining an optimal targeted training scheme according to the error of badminton games of athletes.

According to a first aspect, the invention provides a sports event training prediction method based on AI and image analysis, which comprises the steps of obtaining a shuttlecock match video of a player, determining a plurality of miss points by using a miss point determination model based on the shuttlecock match video of the player, determining K values based on the miss points, clustering based on the miss points and the K values to determine K miss areas, constructing a miss area map, wherein the miss area map comprises K miss area nodes and a plurality of edges between the K miss area nodes, the node characteristics of each miss area node are the miss points of the miss area, the edges between the miss area nodes are straight line distances between the miss areas, processing the miss area map based on a graph neural network to determine at least K ball hitting training points, determining a target training plan based on the at least K ball hitting training points.

In one possible implementation, the determining the target training program based on the at least K batting training points comprises determining a plurality of high-importance batting training points and a plurality of medium-importance batting training points based on the player badminton game video and the at least K batting training points, generating a plurality of sets of training programs based on the plurality of high-importance batting training points, the plurality of medium-importance batting training points and the at least K batting training points, displaying the plurality of sets of training programs on a screen, acquiring a training program selected by a user, and taking the training program selected by the user as the target training program.

In one possible implementation, the fault point determination model is a recurrent neural network model.

In one possible implementation manner, the clustering based on the plurality of misball hitting points and the K value comprises the steps of clustering based on the plurality of misball hitting points and the K value by using a K-means clustering algorithm to determine K clusters, and connecting the misball hitting points of the edges in each cluster to obtain K error areas.

According to a second aspect, the invention provides a sports event training prediction system based on AI and image analysis, which comprises an acquisition module, a fault point determination module, a fault point analysis module, a clustering module and a pattern construction module, wherein the acquisition module is used for acquiring a badminton game video of an athlete, the fault point determination module is used for determining a plurality of play mispoints by using a fault point determination model based on the badminton game video of the athlete, the fault point analysis module is used for determining K values based on the plurality of play mispoints, the clustering module is used for clustering and determining K fault areas based on the plurality of play mispoints and the K values, the pattern construction module is used for constructing a fault area pattern, the fault area pattern comprises a plurality of edges between K fault area nodes and K fault area nodes, the node characteristics of each fault area node are the straight line distance between the fault areas, the training point determination module is used for processing and determining at least K training points based on a graph neural network, and the training determination module is used for determining at least one training point based on the at least K target play points.

In one possible implementation, the training plan determining module is further configured to determine a plurality of high-importance batting training points and a plurality of medium-importance batting training points based on the player badminton game video and the at least K batting training points, generate a plurality of sets of training plans based on the plurality of high-importance batting training points, the plurality of medium-importance batting training points and the at least K batting training points, display the plurality of sets of training plans on a screen, acquire a training plan selected by a user, and use the training plan selected by the user as a target training plan.

In one possible implementation, the fault point determination model is a recurrent neural network model.

In one possible implementation manner, the clustering module is further configured to determine K clusters by using a K-means clustering algorithm based on the multiple play mispoints and the K values, and connect the play mispoints of the edges in each cluster to obtain K error areas.

According to a third aspect, an embodiment of the invention provides an electronic device comprising a processor, a memory, and a computer program, wherein the computer program is stored in the memory and configured to be executed by the processor to implement a method as described above, the method comprising obtaining a player shuttlecock match video, determining a plurality of miss points using a miss point determination model based on the player shuttlecock match video, determining a K value based on the plurality of miss points, clustering based on the plurality of miss points, the K value, determining K zones, constructing a miss zone map comprising a plurality of edges between K miss zone nodes and K miss zone nodes, wherein the node characteristics of each miss zone node are a plurality of miss points for the present miss zone, the edges between miss zone nodes are linear distances between miss zones, processing the miss zone map based on a graph neural network, determining at least K training points, wherein each miss zone has at least one training point, and determining a training ball stroke based on the at least K target training stroke plan based on the at least K target ball stroke points.

According to a fourth aspect, there is provided a computer readable storage medium having stored thereon a computer program which when executed by a processor implements the previously provided method for predicting athletic event training based on AI and image analysis, the method comprising obtaining a player shuttlecock match video, determining a plurality of miss points using a miss point determination model based on the player shuttlecock match video, determining a K value based on the plurality of miss points, clustering based on the plurality of miss points, the K value, determining K miss areas, constructing a miss area map comprising a plurality of edges between K miss area nodes and K miss area nodes, wherein the node characteristics of each miss area node are the plurality of miss points of the present miss area, the edges between the miss area nodes are linear distances between the miss areas, processing the miss area map based on a graph neural network, determining at least K miss points, wherein each miss area has at least one training point, and determining a target hitting plan training based on the at least K miss points.

The method comprises the steps of obtaining a shuttlecock match video of an athlete, determining a plurality of miss points by using a miss point determination model based on the shuttlecock match video of the athlete, determining K values based on the miss points, determining K miss areas by clustering based on the miss points and the K values, constructing a miss area map, wherein the miss area map comprises K miss area nodes and a plurality of edges between the K miss area nodes, the node characteristics of each miss area node are the miss points of the miss area, the edges between the miss area nodes are straight line distances between the miss areas, processing the miss area map based on a graph neural network to determine at least K batting training points, determining a target training plan based on the at least K batting training points, and accurately determining an optimal targeted training scheme according to the shuttlecock match of the athlete.

Drawings

FIG. 1 is a schematic flow chart of a training prediction method for a sporting event based on AI and image analysis according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a high-speed industrial camera according to an embodiment of the present invention;

fig. 3 is a schematic diagram of a mobile high-speed shooting device according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a batting misinformation point at a badminton court according to an embodiment of the present invention;

FIG. 5 is a schematic flow chart of determining K error regions according to an embodiment of the present invention;

FIG. 6 is a schematic flow chart of determining a target training plan according to an embodiment of the present invention;

Fig. 7 is a schematic diagram of a sports event training prediction system based on AI and image analysis according to an embodiment of the present invention.

Detailed Description

The invention will be described in further detail below with reference to the drawings by means of specific embodiments. Wherein like elements in different embodiments are numbered alike in association. In the following embodiments, numerous specific details are set forth in order to provide a better understanding of the present invention. However, one skilled in the art will readily recognize that some of the features may be omitted, or replaced by other elements, materials, or methods in different situations. In some instances, related operations of the present invention have not been shown or described in the specification in order to avoid obscuring the core portions of the present invention, and may be unnecessary to persons skilled in the art from a detailed description of the related operations, which may be presented in the description and general knowledge of one skilled in the art.

In an embodiment of the present invention, there is provided a sports event training prediction method based on AI and image analysis as shown in fig. 1, where the sports event training prediction method based on AI and image analysis includes steps S1 to S7:

Step S1, obtaining a badminton match video of the athlete.

The athlete badminton match video is high-definition high-frame-rate video material shot by professional high-speed shooting equipment deployed in a match field.

The camera shooting equipment can comprise a high-speed industrial camera set fixedly deployed at a competition field and a mobile high-speed shooting equipment matched with the anti-shake tripod head. All the image pickup devices have a global shutter function so as to avoid picture smear under a high-speed motion scene and ensure the restoration precision of video details. Fig. 2 is a schematic diagram of a high-speed industrial camera according to an embodiment of the present invention. Fig. 3 is a schematic diagram of a mobile high-speed shooting device according to an embodiment of the present invention.

The player badminton match video can completely present the match detailed information such as the batting action, the running position, the score, the error and the like of the player in the match.

And S2, determining a plurality of play mispoints by using a fault point determination model based on the player badminton match video.

The fault point determination model is a recurrent neural network model. The input of the error point determining model is the badminton match video of the athlete, and the output of the error point determining model is a plurality of lost ball error points.

The recurrent neural network model comprises a Recurrent Neural Network (RNN), which is a neural network model with sequential data processing capability. The recurrent neural network can process time-series input data by utilizing an internal memory mechanism and can capture time dependency relationship in the data. The cyclic neural network introduces cyclic connection in the network, so that the output at the current moment is not only dependent on the current input, but also influenced by the hidden state at the previous moment. The cyclic neural network can be used for analyzing and processing time sequence data such as video frame sequences, language texts and the like, and can effectively mine dynamic characteristics and rules in the sequence data.

The plurality of lost ball mispoints are specific site positions of the player with the ball striking errors in the game and the set of corresponding error event characteristics, which are identified and extracted after the analysis of the badminton game video of the player through the error point determination model. Fig. 4 is a schematic diagram of a playing error point at a badminton field according to an embodiment of the present invention.

Each lost ball point corresponds to a three-dimensional space coordinate when the lost ball happens, and comprises a time stamp when the lost ball happens, a type of the lost ball, a batting action characteristic corresponding to the lost ball, a stop area of the player when the lost ball happens, and a flight track parameter of the ball when the lost ball happens.

The type of fault is specific tactical classification information for the nature of the fault. The fault types comprise subdivision types such as ball returning, net falling, bottom line out-of-bounds, side line out-of-bounds, killing fault, ball picking fault, receiving fault and the like.

The batting action is characterized by physical characteristics of the motion state of limbs and instruments of the athlete at the moment of failure. The batting action characteristics include swing speed and face angle when the player loses.

Athlete location areas are information describing the specific tactical location of the athlete on the field when the fault occurs. The athlete's stance zone includes whether the athlete is in a top, middle, or bottom field when the error occurs, and the specific orientation in the left and right halves.

The flight track parameters of the shuttlecock are data for recording the space flight state of the shuttlecock in the fault process. The flight trajectory parameters of the ball comprise the initial flight direction, the flight height and the deviation value of the ball drop pre-judgment point and the actual drop point of the ball.

The sportsman badminton match video comprises a sportsman limb action frame sequence which continuously changes along with time and a badminton flying track sequence in the air, the video data comprises dynamic time sequence information such as the time of a sportsman waving, the batting angle, the moving speed, the badminton drop point and the like, and the fault point determining model can analyze the specific position and the moment of the occurrence of the fault of the frame sequence information in the video.

The circulating neural network can carry out frame-by-frame scanning and time sequence analysis on the shuttlecock match video of the athlete through a special circulating structure, and the hidden layer state in the circulating neural network can be continuously updated along with continuous input of video frames, so that continuous characteristics of actions, ball roads and the like at the front and rear moments in the shuttlecock match video of the athlete are memorized and associated. The model can learn the standard high-level badminton action mode and the trajectory of the ball path, and then can extract action characteristics of the player such as swing, pace, batting angle and the like in the current frame and trajectory characteristics of the badminton such as flight direction, speed and the like in real time when processing the video of the badminton match of the player, and calculate the deviation value of the characteristics and the standard mode. When the cyclic neural network detects that the ball receiving position deviation is caused by slow movement of a player and the swing action deformation causes the ball returning to get off the net or get out of the limit, or the ball striking force is used for controlling the flight track of the ball to deviate from the expected abnormal time sequence modes and the like, the spatial mapping relation of the abnormal modes in the video frame can be used for positioning the specific position of the badminton in the moment when the badminton falls on the ground or breaks down by combining the coordinate system of the competition field, and various associated features corresponding to the faults are extracted and integrated, so that the error point is marked as a ball striking error point. The model can continuously traverse the whole competition video sequence, analyze and mark each picture which accords with the fault characteristics, and finally integrate all marking results to output a plurality of ball hitting error points.

In some embodiments, the determining a plurality of play error points based on the player badminton game video includes steps S21 to S23:

step S21, determining a plurality of player action time sequence parameter sets of single incoming ball return shots, a plurality of real-time flight track sequence sets of single incoming ball return shots and video frames corresponding to abnormal actions based on the player badminton match video.

In some embodiments, a recurrent neural network may be used to determine a set of player action timing parameters for a plurality of single ball shots, a set of real-time flight trajectory sequences for a plurality of single ball shots, a video frame for an abnormal action.

The player action time sequence parameters of each single ball return stroke are output through the cyclic neural network, and continuous data of the motion state of each part of the player body along with time change are obtained in each single complete ball return stroke process. The player action time sequence parameters of the single ball return stroke comprise a player joint angle time sequence, a limb position time sequence, a speed time sequence and an acceleration time sequence.

The real-time flying track sequence of the shuttlecock with each single incoming ball strike back is output through a cyclic neural network, and in the process of each single complete incoming ball strike back, the shuttlecock is driven to the next time, and the space position of the shuttlecock before landing or out-of-range continuously changes along with the time three-dimensional coordinate data.

The video frames corresponding to the abnormal actions are determined through the cyclic neural network, all picture frames showing the moment that the player has errors or does not accord with the action in the conventional mode in the badminton match video of the player, and limb gesture picture information and spatial position picture information of the badminton in the moment when the player makes the abnormal actions can be recorded in the picture frames.

The cyclic neural network has the capability of modeling the sequence of time sequence data and the capability of capturing the spatial characteristics and time association in the video, and can accurately identify the pixel change, the human body posture key point association and the object motion trail rule of continuous frames in the badminton match video of the sportsman. The cyclic neural network can extract the time sequence characteristics of the player limb motions from the continuous frames of the player badminton match video and integrate the time sequence characteristics into continuous data of the motion states of all parts of the body in each single incoming ball return stroke process, so that a plurality of player action time sequence parameter sets of single incoming ball return strokes are formed, the model can identify the pixel characteristics of the shuttlecock in the player badminton match video and fit the time sequence changes of the spatial positions of the shuttlecock, and further a plurality of real-time flight track sequence sets of the shuttlecock single incoming ball return strokes are generated. The model can also learn the characteristic mode of normal batting action and label the frame sequence of the deviation mode in the badminton match video of the sportsman so as to determine the video frame corresponding to the abnormal action.

Step S22, determining a player gravity center adjusting frequency time domain value, a player swing angular velocity change curve, player movement starting speed time course data and a single-slapping ball action scoring table based on the player action time sequence parameter sets of the plurality of single-slapping shuttlecocks, the real-time flight track sequence sets of the plurality of single-slapping shuttlecocks and the video frames corresponding to the abnormal actions.

In some embodiments, a recurrent neural network may be used to determine player center of gravity adjustment frequency time domain values, player swing angular velocity profiles, player movement initiation speed time course data, single slap ball action scoring tables.

The gravity center adjusting frequency refers to the number of times that the body gravity center of the athlete completes effective position or posture adjustment in unit time.

The athlete's body barycenter adjustment frequency time domain value is a continuous data sequence of athlete's body barycenter adjustment frequency over time output through a recurrent neural network.

The player swing angular velocity change curve is a continuous curve of the change of the instantaneous angular velocity of the player holding hands with time in the swing process, which is output through the cyclic neural network.

The athlete movement starting speed time course data is dynamic data of the time course of the initial speed of the body or the key part when the athlete starts to move from rest after receiving an incoming ball signal output by the cyclic neural network.

The single-shot action scoring table is a data table for quantitatively evaluating the shot action quality of all single-shot shots in a badminton match through a cyclic neural network. The row dimension in the single-slapping ball action scoring table takes each single ball coming return stroke as an independent row unit, and each row corresponds to one complete ball coming return stroke in the match, and the column dimension in the single-slapping ball action scoring table comprises a basic information column, a scoring Xiang Lie, a scoring result column and a grading result column.

The basic information column covers the time stamp of the single return stroke and the stroke round number.

The scoring item list includes an action normalization index, an action fluency index, and a striking speed index.

The score result column records the specific score and the comprehensive score result of each score item.

The ranking result column presents four grades of excellent, good, pass, and fail defined by composite scores.

The circulating neural network can extract the time sequence change rules of the motion characteristics such as the athlete gravity center adjusting frequency sequence, the athlete swing angular speed change curve and the like from the athlete action time sequence parameter sets of a plurality of single ball return shots so as to generate corresponding characteristic sequences, and meanwhile, the circulating neural network can combine the spatial characteristics of the badminton real-time flight track sequence sets of a plurality of single ball return shots to establish the association of actions and tracks. The circulating neural network can also combine visual features of video frames corresponding to abnormal actions with quantitative features of a plurality of player action time sequence parameter sets of single ball hitting return, so that feature modes of the abnormal actions are learned, and finally, an athlete gravity center adjusting frequency sequence and a single ball hitting action scoring table quantitative analysis result are output.

Step S23, determining a plurality of ball hitting error points based on the shuttlecock real-time flight track sequence set of the plurality of single ball hitting return shots, the athlete gravity center adjusting frequency time domain value, the athlete swing angular velocity change curve, the athlete moving starting speed time course data and the single ball hitting action scoring table.

In some embodiments, a recurrent neural network may be used to determine a plurality of ball play mispoints.

The circulating neural network can extract key characteristics from a plurality of shuttlecocks shot once in real time, namely a real-time flight track sequence set of shuttlecocks shot once, a player gravity center adjusting frequency time domain value, a player swing angular velocity change curve, player movement starting speed time course data and a single-shot action scoring table. The model can extract the characteristics of the deviation of the ball falling point and the stability of the flying track from the real-time flying track sequence set of the badminton, extract the characteristics of average frequency and peak frequency from the time domain value of the gravity center adjusting frequency, extract the characteristics of the maximum angular velocity and the rising rate of the angular velocity from the swing-bat angular velocity change curve, extract the characteristics of the starting time and the average starting velocity from the moving starting velocity time course data, and take various scores in the single-bat ball action scoring table as independent characteristics. The model can conduct time alignment and association modeling on all dimension features corresponding to the same batting action through a time sequence association analysis technology, and then a feature chain is constructed to analyze association among all features. And finally, the model can judge the comprehensive characteristics of each batting according to a preset error judgment rule and a threshold value, when the characteristics of a certain batting meet error judgment conditions, the batting is determined to be a batting error point, and a plurality of batting error points can be determined by the final model through analyzing and judging all single batting one by one.

And S3, determining a K value based on the plurality of ball hitting error points.

In some embodiments, a fault point analysis model may be used to determine the K value. The fault point analysis model is a deep neural network model. The input of the error point analysis model is the multiple ball striking error points, and the output of the error point analysis model is a K value.

The deep neural network model includes a Deep Neural Network (DNN), which is a neural network based on a deep learning architecture, which may contain multiple hidden layers. The deep neural network can perform characteristic abstraction and layer-by-layer mapping on input data through multi-layer nonlinear transformation. The deep neural network can learn the complex distribution rule of data in space and adjust the interlayer weight through a back propagation algorithm, so that the deep neural network has strong function fitting capability and pattern recognition capability.

The K value is an optimal cluster number value for clustering after comprehensively evaluating the distribution characteristics of a plurality of ball playing fault points through a fault point analysis model.

The plurality of lost ball points record multidimensional attribute data including three-dimensional space coordinates, fault types, batting action characteristics, flight track parameters of balls and the like. These data reflect not only the degree of aggregation of errors in physical space, but also the inherent relevance of errors in causes, action patterns and tactical scenarios. The plurality of ball hitting error points can reflect the concentrated trend and the distribution rule of the exercise errors, and the error point analysis model can determine the proper clustering quantity by deep mining of the characteristics.

The deep neural network can analyze the spatial distribution density of the ball hitting error points and identify dense areas with high occurrence of errors on the field. Meanwhile, the model can combine the fault type and the batting action characteristic to mine the hidden action mode similarity behind the space position. For example, the model can find that errors near the bottom line are mostly related to excessive swing speed, while errors in front of the net are mostly not related to swatter angle control. The depth neural network can evaluate how many groups the overall fault is divided into by the cross combination of the characteristics and the depth characteristic extraction, and can ensure the consistency of the fault characteristics in the groups and the difference of the fault characteristics among the groups to the greatest extent. The model can calculate characteristic aggregation index under different division numbers, and finally maps an integer through an output layer, wherein the integer is a K value capable of optimally representing the current error distribution structure.

And S4, clustering based on the multiple ball hitting error points and the K values to determine K error areas.

In some embodiments, fig. 5 is a schematic flow chart of determining K error areas according to an embodiment of the present invention, where the steps S41 to S42 of determining K error areas are:

and S41, carrying out clustering by using a K-means clustering algorithm based on the multiple ball hitting error points and the K value to determine K clusters.

The K-Means clustering algorithm (K-Means) is an iteratively solved clustering analysis algorithm. The K-means clustering algorithm can divide the data set into mutually disjoint subsets by a preset number of clusters so as to minimize the sum of squares of distances between data points in each subset and centroids of the subset. The K-means clustering algorithm can quickly converge and automatically merge data points with similar spatial features into the same cluster.

The K clusters are specific data groups obtained by dividing all the ball hitting fault points according to the feature similarity through a K-means clustering algorithm.

Each cluster contains a set of multiple ball-playing mispoints that have a high degree of similarity in spatial location, type of mishap, or motion characteristics. The different clusters exhibit significant differences in spatial distribution and miss-characteristics.

The clustering process of the multiple mishit points and the K values by the K-means clustering algorithm comprises the steps of firstly extracting three-dimensional space coordinate data in the multiple mishit points to serve as a core clustering basis, and simultaneously reading a preset K value to determine the number of clustered target clusters. The algorithm randomly selects K error points from a plurality of lost-play error points as initial clustering centers, calculates Euclidean distances from each lost-play error point to the K clustering centers, and distributes each lost-play error point to the cluster to which the corresponding clustering center belongs according to the principle of nearest distance. After the first round of distribution is completed, the algorithm recalculates the three-dimensional space coordinate mean value of all the ball hitting error points in each cluster, takes the mean value as a new cluster center, and then executes the operations of distance calculation and error point distribution again. The process is iterated continuously until the change amplitude of the cluster center position obtained by two adjacent times of calculation is smaller than a preset threshold value, the clustering process is stopped at the moment, and K groups are finally formed, namely K clusters.

Multiple complex ball-playing mispoint data can be effectively integrated through clustering. Because the plurality of fault points of playing the ball comprise multidimensional information such as space positions, fault types, batting action characteristics and the like, the quantity is large, the distribution is scattered, and the centralized rule of the faults of the player is difficult to find by directly analyzing the data. And the ball playing error points with similar characteristics can be classified into one type and K clusters are formed through clustering, so that the data structure is greatly simplified, and valuable error distribution and characteristic association information can be conveniently extracted. By dividing the plurality of lost ball points into K clusters, different concentrated areas and characteristic types of player errors on the playing field can be intuitively displayed. Through analysis of the clusters, the dominant type and typical action deviation of errors in each area can be rapidly determined, and the technical shortboards and error cause rules of athletes in different areas can be clearly found through further comparing the characteristic differences of different clusters.

And S42, connecting the ball striking error points of the edges in each cluster to obtain K error areas.

The missing ball points at the edge of each cluster are missing ball points at the peripheral boundary position of the spatial distribution of the cluster in a single cluster obtained by a K-means clustering algorithm, and the three-dimensional space coordinates of the missing points are relatively far from the cluster center of the cluster.

The K error areas are closed field space areas formed by connecting the ball hitting error points of the edges in each cluster in sequence according to the actual space positions of the ball hitting error points in the competition field, each error area corresponds to one cluster obtained by clustering, and the error areas are consistent with the number K of clusters obtained by clustering.

In some embodiments, three-dimensional space coordinates of all the ball-hitting error points in each cluster can be extracted, the ball-hitting error points at the edges forming the cluster distribution profile are screened out, then connection is performed according to the sequence of the ball-hitting error points at the edges in the space position of the competition field, so as to form a closed polygonal area, each cluster correspondingly generates an error area, and finally K error areas matched with the number of the clusters are obtained.

And S5, constructing a fault area map, wherein the fault area map comprises K fault area nodes and a plurality of edges between the K fault area nodes, the node characteristics of each fault area node are a plurality of ball striking fault points of the fault area, and the edges between the fault area nodes are straight line distances between the fault areas.

The map of the error area is a map data structure capable of representing the relation of error space distribution structure and area association in badminton match. The fault region map contains K fault region node information describing region attributes and side information describing region spatial adjacency.

The node characteristics of each miss area node detail records all the ball striking misinformation contained in the area. The edges represent the straight line distance between the error areas, and the edges in the error area map can intuitively display the position characteristics of each error area and the spatial distance relation between the error areas.

And S6, processing the error area map based on the graph neural network to determine at least K batting training points, wherein each error area has at least one batting training point.

The Graph Neural Network (GNN) is a deep learning architecture that can be used exclusively to process atlases, which are data with topology. The graph neural network can realize information interaction and aggregation among nodes through a message passing mechanism. In the graph neural network, each node is able to update its own state representation according to its own characteristics as well as the characteristics of neighboring nodes. The graph neural network can simultaneously capture topological structure information of the graph and attribute characteristics of the nodes. Through multi-layer graph convolution or aggregation operation, the graph neural network can deeply mine the dependency relationship and the global distribution pattern among nodes.

The ball hitting training point is a specific site position for the athlete to train in a specific way, which is determined in the error area after the map of the error area is processed by the map neural network. Wherein each error region at least comprises one batting training point, and K error regions comprise at least K batting training points.

By constructing the fault region map, a spatial association network among K fault regions can be clearly reflected, the association information is very important for determining the batting training points, and the association analysis of fault causes can be influenced due to the spatial position relationship among different fault regions. The data information can be more fully utilized by taking a plurality of ball playing fault points of each fault area as node characteristics of nodes of the fault area and taking straight line distances among the fault areas as edge characteristics. This helps the neural network to better understand the characteristic differences and spatial correlations of the various miss areas, thereby improving the accuracy of the shot training point determination. The complex relation and information transfer between the nodes can be effectively learned based on the graph neural network processing fault region map data, so that the core fault characteristics and training priority of each fault region can be more accurately mined. Compared with the traditional cluster analysis or single feature extraction method, the graph neural network has better characterization capability and learning capability when processing graph data, and can give consideration to the integral association between individual features of the error region and the region.

The graph neural network can aggregate the self characteristics of each node and the characteristics of the neighborhood nodes through the graph convolution layer, and analyze the influence of the space distance between different error areas on the occurrence of errors, such as the relevance of whether error reasons exist in adjacent error areas. The neural network of the graph can deeply excavate the characteristics of each error area node so as to calculate the dense center position of the error points in the error area, wherein the dense center position of the error points in the error area can be the place where the player error is most concentrated, and then the dense center position is used as the primary batting training point. If the area of a certain error area is larger or the distribution of the error points is more scattered, the model can additionally determine auxiliary training points in the area according to the distribution density of the error points and the concentration condition of the error types so as to ensure that at least one batting training point exists in each error area.

And S7, determining a target training plan based on the at least K batting training points.

In some embodiments, fig. 6 is a schematic flow chart of determining a target training plan according to an embodiment of the present invention, where the steps S71 to S73 of determining the target training plan are:

Step S71, determining a plurality of high-importance batting training points and a plurality of medium-importance batting training points based on the player badminton match video and the at least K batting training points.

In some examples, a shot training point determination model may be used to determine a plurality of high importance shot training points, a plurality of medium importance shot training points. The batting training point determining model is a cyclic neural network model. The input of the batting training point determining model is the player badminton match video and the at least K batting training points, and the output of the batting training point determining model is a plurality of high-importance batting training points and a plurality of medium-importance batting training points.

The plurality of high-importance batting training points are training points with decisive influence or serious error consequences on the comparison of the victory and the victory of the match after the batting training points are determined by analyzing the badminton match video of the athlete and at least K batting training points through a batting training point determination model.

The type of error corresponding to the high-importance ball striking training points, which are core problem points in the athlete's technical flitch, can directly affect the winner and the winner of the game.

The important batting training points are training points which are determined by the batting training point determining model, have importance degree inferior to that of the important batting training points, have influence on the game fluency and have relatively low direct scoring probability.

The technical stability of the athlete can be assisted by targeted training of the important hit training points to solve the minor error problem in the game.

The badminton match video of the athlete comprises information such as the occurrence scene of errors, the error frequency, the influence degree of comparison match situations and the like corresponding to each batting training point, the batting training points define specific site positions to be analyzed, the information can provide a complete data analysis basis for determining the importance degree of the model distinguishing training points for the batting training points, and the model can judge the influence weights of different training points on the athlete by analyzing match details in the badminton match video of the athlete.

The recurrent neural network can use its timing analysis capability to trace back in the player's badminton game video the game segments associated with the spatial positions of at least K batting training points. The recurrent neural network may analyze whether the player is in a critical score stage when the player hits the ball in the vicinity of these training points, whether the error is caused by non-forced action deformation, and whether the error directly results in opponent scoring. The cyclic neural network can be related to the contextual information of the game through a memory mechanism, and the error risk cost corresponding to each batting training point is calculated. If a failure of a particular hit training point is likely to occur at the event or result in a succession lose scores, the model may determine that the hit training point has a high training priority and categorize the hit training point as a plurality of high-importance hit training points. If the associated error of a particular shot training point is a forced error and does not result in serious tactical disadvantages, the model classifies the shot training point as a plurality of medium-importance shot training points.

Step S72, generating a plurality of training programs based on the plurality of high importance shot training points, the plurality of medium importance shot training points, and the at least K shot training points.

In some embodiments, multiple sets of training plans may be generated using a plan determination model. The planning determination model is a deep neural network model. The inputs of the plan determination model are the plurality of high-importance batting training points, the plurality of medium-importance batting training points and the at least K batting training points, and the outputs of the plan determination model are a plurality of sets of training plans.

The training programs are various badminton training embodiments output by the program determining model, and each training program comprises training frequency of each batting training point, special action training content aiming at each batting training point, training connection sequence of each batting training point and single training duration of each batting training point.

The training frequency of each hitting training point refers to the number of training times scheduled for a single hitting training point in a single set of training programs. The training frequency of the high-importance batting training points is higher than that of the medium-importance batting training points.

The special action training content for each batting training point is a badminton technical action training project designed according to the technical short plates and the error types corresponding to the batting training points. The training content comprises a racket face angle control exercise of the net front ball rolling, a direction aiming exercise of bottom line buckling, a station position adjusting exercise of receiving and sending the ball, a force handle control exercise of the back ground pull and hanging, a force sending time exercise of the middle ground ball picking and the like.

The training connection sequence of each batting training point refers to the sequence of training special exercises of different batting training points in sequence in the training process.

At least K batting training points cover all of the player's desired training positions, wherein a plurality of high importance batting training points define a training core priority and wherein a plurality of medium importance batting training points define a secondary training direction. The data define the key areas and coverage areas of training, and can provide basis for planning a scientific training plan for a plan determination model, and the model can distribute reasonable training resources according to the importance degree of training points to generate diversified training schemes.

The deep neural network can generate a plurality of sets of training plans, and the deep neural network has the multi-dimensional feature extraction and accurate resource allocation capability. The deep neural network can process input spatial point location and attribute data by utilizing strong feature mapping and nonlinear combination capability. The model first deeply parses the detailed feature data carried behind each shot training point entered, including shot motion features and error types. The model can establish accurate mapping relation between the specific action defect characteristics and a preset kinematics correction strategy through neuron operation of an hidden layer, for example, when the model recognizes that the related characteristic of a certain high important batting training point is that the batting angle is too large and the ball return is too high, the model can be automatically matched and generates 'the batting angle control exercise of the pre-net batting' as the special action training content of the point. Meanwhile, the deep neural network can perform weight distribution calculation according to importance labels of the input points. The model can adjust the value of the output layer through activating the function according to the principle that the weight of a plurality of high-importance batting training points is obviously higher than that of a plurality of medium-importance batting training points, so that higher training frequency and longer single training duration are allocated to the high-importance batting points. When determining the training engagement sequence of each batting training point, the model comprehensively considers the space distance and the rationality of physical energy consumption of the batting training points on the field, and generates the engagement sequence which accords with human engineering and can simulate the actual combat rhythm by simulating the moving path of the athlete. In addition, the model can simulate different training guidance by adjusting internal super parameters or adopting a multi-output structure, for example, the model can generate a set of skill type plans focusing on correcting specific technical actions, another set of physical type plans focusing on full-field high-frequency mobile running, and a comprehensive type plan focusing on alternate training of high-importance batting training points and medium-importance batting training points, and finally, a plurality of different and feasible training plans are output.

And step S73, displaying the training plans on a screen, acquiring the training plan selected by the user, and taking the training plan selected by the user as a target training plan.

The target training program is a scheme selected by a user from a plurality of training programs and used for the practical training of the athlete, and the target training program can be fit with the error characteristics and training requirements of the athlete.

Based on the same inventive concept, fig. 7 is a schematic diagram of a sports event training prediction system based on AI and image analysis according to an embodiment of the present invention, where the sports event training prediction system based on AI and image analysis includes:

an obtaining module 81, configured to obtain a badminton match video of an athlete;

a miss-point determination module 82 for determining a plurality of missed-play mispoints using a miss-point determination model based on the player's shuttlecock match video;

a fault point analysis module 83, configured to determine a K value based on the plurality of ball striking mispoints;

the clustering module 84 is configured to determine K error areas by performing clustering based on the plurality of ball striking error points and the K value;

the map construction module 85 is configured to construct a fault area map, where the fault area map includes K fault area nodes and a plurality of edges between the K fault area nodes, and a node characteristic of each fault area node is a plurality of ball striking error points of the fault area, and edges between the fault area nodes are straight line distances between the fault areas;

a training point determination module 86 for determining at least K shot training points based on processing the fault region map based on a neural network, wherein each fault region has at least one shot training point;

the training program determining module 87 is configured to determine a target training program based on the at least K hitting training points.

It should be noted that in order to simplify the presentation disclosed in this specification and thereby aid in understanding one or more inventive embodiments, various features are sometimes grouped together in a single embodiment, figure, or description thereof. This method of disclosure does not imply that the subject matter of the present description requires more features than are set forth in the claims. Indeed, less than all of the features of a single embodiment disclosed above.

Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments of this specification. Other variations are possible within the scope of this description. Thus, by way of example, and not limitation, alternative configurations of embodiments of the present specification may be considered as consistent with the teachings of the present specification. Accordingly, the embodiments of the present specification are not limited to only the embodiments explicitly described and depicted in the present specification.

Claims (6)

1. A method for predicting sports event training based on AI and image analysis, characterized in that it includes: Obtain videos of athletes playing badminton; Based on the badminton match video of the athlete, a fault point determination model is used to identify multiple hitting fault points. Each fault point corresponds to a three-dimensional spatial coordinate at the time of the fault and includes the timestamp of the fault, the fault type, the characteristics of the hitting action corresponding to the fault, the athlete's standing area at the time of the fault, and the trajectory parameters of the shuttlecock at the time of the fault. The fault point determination model is a recurrent neural network model. The process of determining multiple hitting fault points based on the badminton match video of the athlete includes: Based on the badminton match video of the athlete, determine the time sequence parameter set of multiple single shuttlecock return, the set of real-time flight trajectory sequence of the badminton shuttlecock for multiple single shuttlecock return, and the video frame corresponding to the abnormal action; Based on the set of athlete action timing parameters for multiple single-shot returns, the set of badminton shuttlecock real-time flight trajectory sequences for multiple single-shot returns, the time domain value of athlete's center of gravity adjustment frequency determined by the video frames corresponding to the abnormal actions, the athlete's swing angular velocity change curve, the athlete's movement start speed time history data, and the single-shot action scoring table. Based on the set of real-time flight trajectory sequences of badminton shuttlecocks from multiple single-shot returns, the time domain value of the athlete's center of gravity adjustment frequency, the change curve of the athlete's swing angular velocity, the time history data of the athlete's movement start speed, and the single-shot hitting action scoring table, multiple hitting error points are determined. The K value is determined based on the multiple points of hitting error; Based on the multiple points of hitting errors and the K value, clustering is performed to determine K error regions; Construct a fault region map, which includes K fault region nodes and multiple edges between the K fault region nodes. The node characteristics of each fault region node are multiple hitting errors in this fault region, and the edges between the fault region nodes are the straight-line distances between the fault regions. The error region map is processed using a graph neural network to determine at least K hitting training points, wherein each error region has at least one hitting training point. A target training plan is determined based on the at least K ball-hitting training points. 2. The sports event training prediction method based on AI and image analysis as described in claim 1, characterized in that determining the target training plan based on the at least K hitting training points includes: Based on the badminton match video of the athlete and the at least K hitting training points, multiple high-importance hitting training points and multiple medium-importance hitting training points are determined. Multiple training plans are generated based on the multiple high-importance hitting training points, the multiple medium-importance hitting training points, and the at least K hitting training points; The multiple training plans are displayed on the screen, and the training plan selected by the user is obtained and used as the target training plan. 3. The sports event training prediction method based on AI and image analysis as described in claim 1, characterized in that the step of clustering to determine K error regions based on the multiple ball-playing error points and the K value includes: Based on the multiple points of hitting errors and the K value, the K-means clustering algorithm is used to determine K clusters; Connect the points of hitting errors at the edges of each cluster to obtain K error zones. 4. A sports event training prediction system based on AI and image analysis, characterized in that it includes: The acquisition module is used to acquire videos of athletes playing badminton matches. The error point determination module is used to determine multiple hitting errors based on the badminton match video of the athlete, using an error point determination model. Each hitting error point corresponds to a three-dimensional spatial coordinate at the time of the error and includes a timestamp of the error, error type, the corresponding hitting action characteristics, the athlete's standing area at the time of the error, and the trajectory parameters of the shuttlecock at the time of the error. The error point determination model is a recurrent neural network model. Determining multiple hitting errors based on the badminton match video of the athlete includes: Based on the badminton match video of the athlete, determine the time sequence parameter set of multiple single shuttlecock return, the set of real-time flight trajectory sequence of the badminton shuttlecock for multiple single shuttlecock return, and the video frame corresponding to the abnormal action; Based on the set of athlete action timing parameters for multiple single-shot returns, the set of badminton shuttlecock real-time flight trajectory sequences for multiple single-shot returns, the time domain value of athlete's center of gravity adjustment frequency determined by the video frames corresponding to the abnormal actions, the athlete's swing angular velocity change curve, the athlete's movement start speed time history data, and the single-shot action scoring table. Based on the set of real-time flight trajectory sequences of badminton shuttlecocks from multiple single-shot returns, the time domain value of the athlete's center of gravity adjustment frequency, the change curve of the athlete's swing angular velocity, the time history data of the athlete's movement start speed, and the single-shot hitting action scoring table, multiple hitting error points are determined. The error point analysis module is used to determine the K value based on the multiple hitting error points; The clustering module is used to cluster and determine K error regions based on the multiple hitting error points and the K value; The graph construction module is used to construct a fault region graph, which includes K fault region nodes and multiple edges between the K fault region nodes. The node characteristics of each fault region node are multiple hitting errors in this fault region, and the edges between the fault region nodes are the straight-line distance between the fault regions. The training point determination module is used to process the error region map based on a graph neural network to determine at least K hitting training points, wherein each error region has at least one hitting training point; The training plan determination module is used to determine a target training plan based on the at least K hitting training points. 5. The sports event training prediction system based on AI and image analysis as described in claim 4, wherein the training plan determination module is further configured to: Based on the badminton match video of the athlete and the at least K hitting training points, multiple high-importance hitting training points and multiple medium-importance hitting training points are determined. Multiple training plans are generated based on the multiple high-importance hitting training points, the multiple medium-importance hitting training points, and the at least K hitting training points; The multiple training plans are displayed on the screen, and the training plan selected by the user is obtained and used as the target training plan. 6. The sports event training prediction system based on AI and image analysis as described in claim 4, wherein the clustering module is further used for: Based on the multiple points of hitting errors and the K value, the K-means clustering algorithm is used to determine K clusters; Connect the points of hitting errors at the edges of each cluster to obtain K error zones.