A sports event risk prediction method and system based on image vision

Abstract

The application provides a sports event risk prediction method and system based on image vision, and relates to the technical field of sports event risk prediction based on image vision. The method comprises the following steps: acquiring historical match video of an athlete's event opponent, pre-match training video of the athlete, and shoulder muscle information after the pre-match training of the athlete; based on multiple match simulation video information of the current competition, using a stretching information processing model to determine shoulder multiple stretching information of each match simulation video; based on a shoulder muscle soreness degree distribution map and the shoulder multiple stretching information of each match simulation video, determining a shoulder strain probability distribution map of each match simulation video; and based on a graph neural network, determining a shoulder strain risk value of the current competition from the shoulder strain probability distribution map of each match simulation video. The method can accurately predict the shoulder strain risk of the event by combining the action features of the opponent and the state of the athlete.

Description

Technical Field

[0001] This invention relates to the field of sports event risk prediction technology based on image vision, and specifically to a sports event risk prediction method and system based on image vision. Background Technology

[0002] In two-person competitive sports such as badminton and tennis, the risk of shoulder muscle injury is a key factor affecting athletes' performance. Athletes' shoulders frequently perform high-intensity, high-frequency movements such as serving, hitting, and defending. The cumulative muscle fatigue and workload from long-term training and competition easily lead to sports injuries such as shoulder strains. This can not only interrupt the current competition but also cause chronic damage, affecting the athlete's long-term competitive state. Currently, the assessment of athletes' shoulder injury risk during the preparation stage of sports events relies heavily on coaches' experience-based analysis or basic physiological indicator monitoring, which has significant limitations. Manual analysis, relying too heavily on experience, requires a significant investment of time reviewing historical matches and training videos and compiling data. It is also highly subjective, and the assessment results are easily influenced by individual perceptions and differences in experience, making it difficult to establish a unified and objective judgment standard. Traditional methods can only consider the athlete's own training status or the opponent's tactical style, failing to comprehensively combine both. This makes it difficult to accurately identify high-risk action scenarios during the competition and to predict the specific load these scenarios will place on the athlete's shoulder. Existing solutions struggle to systematically analyze an athlete's pre-competition muscle condition and the characteristics of their opponent's movements, resulting in insufficient accuracy in assessing the risk of muscle strains and a lack of scientific rigor and foresight in risk warnings. Ultimately, they fail to provide athletes with precise references for competition decisions or targeted protective recommendations.

[0003] Therefore, how to accurately predict the risk of shoulder strain during a match by combining the opponent's movement characteristics with the athlete's own condition is an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem this invention addresses is how to accurately predict the risk of shoulder strain during a competition by combining the opponent's movement characteristics with the athlete's own condition.

[0005] According to a first aspect, the present invention provides a method for predicting the risk of sports events based on image vision, comprising: acquiring historical head-to-head videos of an athlete's opponents, pre-match training videos of the athlete, and shoulder muscle information of the athlete after pre-match training; generating multiple simulated head-to-head videos for the current match based on the historical head-to-head videos of the athlete's opponents, the pre-match training videos of the athlete, and the shoulder muscle information of the athlete after pre-match training; determining a shoulder muscle soreness distribution map based on the shoulder muscle information of the athlete after pre-match training; determining multiple shoulder stretching information in each simulated head-to-head video using a stretching information processing model based on the multiple simulated head-to-head videos for the current match; determining a shoulder strain probability distribution map for each simulated head-to-head video based on the shoulder muscle soreness distribution map and the multiple shoulder stretching information in each simulated head-to-head video; determining a shoulder strain risk value for the current match based on the shoulder strain probability distribution map in each simulated head-to-head video using a graph neural network; and sending a risk warning signal to a control terminal if the shoulder strain risk value for the current match is greater than a preset threshold.

[0006] In one possible implementation, determining the shoulder strain risk value for this competition based on the shoulder strain probability distribution map of each simulated match video using a graph neural network includes: constructing a strain analysis map, which includes multiple simulated match nodes and multiple edges between these nodes; the node features of each simulated match node include information from each simulated match video, the shoulder strain probability distribution map of each simulated match video, the shoulder muscle information of the athlete after pre-match training, and the shoulder muscle soreness distribution map; and the edges between the simulated match nodes represent the credibility difference between the simulated match videos; and processing the strain analysis map using a graph neural network to obtain the shoulder strain risk value for this competition.

[0007] In one possible implementation, determining the shoulder strain probability distribution map for each sparring simulation video based on the shoulder muscle soreness distribution map and the shoulder multiple stretching information of each sparring simulation video includes: clustering the shoulder multiple stretching information of each sparring simulation video to obtain K clusters; determining multiple high-risk stretching information based on the K clusters; and determining the shoulder strain probability distribution map for each sparring simulation video based on the multiple high-risk stretching information and the shoulder muscle soreness distribution map.

[0008] In one possible implementation, the information for each simulated match in this competition includes the simulated match video and the credibility of the simulated match video.

[0009] According to a second aspect, the present invention provides a sports event risk prediction system based on image vision, comprising: an acquisition module for acquiring historical head-to-head videos of an athlete's opponents, pre-match training videos of the athlete, and shoulder muscle information of the athlete after pre-match training; a simulation module for generating multiple simulated head-to-head video information for the current match based on the historical head-to-head videos of the athlete's opponents, the pre-match training videos of the athlete, and the shoulder muscle information of the athlete after pre-match training; a soreness distribution map determination module for determining a shoulder muscle soreness distribution map based on the shoulder muscle information of the athlete after pre-match training; and a stretching information determination module. The system comprises the following modules: a shoulder strain probability distribution map determination module, which uses a stretching information processing model to determine the shoulder strain probability distribution map for each simulated match based on the shoulder muscle soreness distribution map and the shoulder strain probability distribution map for each simulated match; a shoulder strain risk value determination module, which uses a graph neural network to determine the shoulder strain risk value for this match based on the shoulder strain probability distribution map for each simulated match; and an early warning module, which sends a risk warning signal to the control terminal if the shoulder strain risk value for this match is greater than a preset threshold.

[0010] In one possible implementation, the strain risk determination module is further configured to: construct a strain analysis graph, which includes multiple simulated competition nodes and multiple edges between them; each simulated competition node's node features include information from each simulated competition video, a shoulder strain probability distribution map for each simulated competition video, shoulder muscle information of the athlete after pre-competition training, and a shoulder muscle soreness distribution map; and the edges between the simulated competition nodes represent the credibility difference between the simulated competition videos; and process the strain analysis graph using a graph neural network to obtain the shoulder strain risk value for this competition.

[0011] In one possible implementation, the strain probability distribution map determination module is further configured to: cluster K clusters based on the shoulder stretching information of each combat simulation video; determine multiple high-risk stretching information based on the K clusters; and determine the shoulder strain probability distribution map of each combat simulation video based on the multiple high-risk stretching information and the shoulder muscle soreness distribution map.

[0012] In one possible implementation, the information for each simulated match in this competition includes the simulated match video and the credibility of the simulated match video.

[0013] According to a third aspect, embodiments of the present invention provide an electronic device, including: 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 the method as described above, the method including: acquiring historical head-to-head videos of an athlete's opponents in a competition, pre-competition training videos of the athlete, and shoulder muscle information of the athlete after pre-competition training; generating multiple simulated head-to-head video information for the current competition based on the historical head-to-head videos of the athlete's opponents in a competition, the pre-competition training videos of the athlete, and the shoulder muscle information of the athlete after pre-competition training; based on the... The shoulder muscle information after the athlete's pre-competition training is used to determine the distribution map of shoulder muscle soreness; based on multiple simulated match videos of the competition, a stretching information processing model is used to determine the multiple shoulder stretching information in each simulated match video; based on the shoulder muscle soreness distribution map and the multiple shoulder stretching information in each simulated match video, a shoulder strain probability distribution map for each simulated match video is determined; based on the shoulder strain probability distribution map of each simulated match video, a graph neural network is used to determine the shoulder strain risk value for the competition; if the shoulder strain risk value for the competition is greater than a preset threshold, a risk warning signal is sent to the control terminal.

[0014] According to the fourth aspect, this embodiment provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the program implements the aforementioned image-vision-based sports event risk prediction method. The method includes: acquiring historical head-to-head videos of an athlete's opponents, pre-match training videos of the athlete, and shoulder muscle information of the athlete after pre-match training; generating multiple simulated head-to-head video information for the current match based on the historical head-to-head videos of the athlete's opponents, the pre-match training videos of the athlete, and the shoulder muscle information of the athlete after pre-match training; and generating multiple simulated head-to-head video information for the current match based on the athlete's pre-match training... After the competition, shoulder muscle information is used to determine the distribution map of shoulder muscle soreness. Based on multiple simulated match videos of this competition, a stretching information processing model is used to determine the multiple shoulder stretching information in each simulated match video. Based on the shoulder muscle soreness distribution map and the multiple shoulder stretching information in each simulated match video, a shoulder strain probability distribution map for each simulated match video is determined. Based on the shoulder strain probability distribution map of each simulated match video, a graph neural network is used to determine the shoulder strain risk value for this competition. If the shoulder strain risk value for this competition is greater than a preset threshold, a risk warning signal is sent to the control terminal.

[0015] This invention provides a method and system for predicting sports event risks based on image vision. The method includes acquiring historical head-to-head videos of an athlete's opponents, pre-match training videos of the athlete, and shoulder muscle information of the athlete after pre-match training; generating multiple simulated head-to-head videos for the current match based on the historical head-to-head videos, the pre-match training videos, and the shoulder muscle information of the athlete after pre-match training; determining a shoulder muscle soreness distribution map based on the shoulder muscle information of the athlete after pre-match training; and generating multiple simulated head-to-head videos for the current match. The method uses a stretching information processing model to determine multiple shoulder stretches in each simulated match video. Based on the shoulder muscle soreness distribution map and the multiple shoulder stretches in each simulated match video, it determines the shoulder strain probability distribution map for each simulated match video. Based on the shoulder strain probability distribution map of each simulated match video, it determines the shoulder strain risk value for this match based on a graph neural network. If the shoulder strain risk value for this match is greater than a preset threshold, a risk warning signal is sent to the control terminal. This method can accurately predict the shoulder strain risk in a match by combining the opponent's movement characteristics and the athlete's own condition. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a sports event risk prediction method based on image vision provided in an embodiment of the present invention;

[0017] Figure 2 A flowchart illustrating the process of determining the probability distribution of shoulder strain in each simulated combat video, as provided in an embodiment of the present invention;

[0018] Figure 3 A schematic diagram of a process for determining the risk value of shoulder strain in this competition, provided as an embodiment of the present invention;

[0019] Figure 4 A schematic diagram of a control terminal provided in an embodiment of the present invention;

[0020] Figure 5 This is a schematic diagram of a sports event risk prediction system based on image vision, provided as an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0022] In this embodiment of the invention, the following are provided: Figure 1 The image-based vision-based sports event risk prediction method is shown, comprising steps S1 to S7:

[0023] Step S1: Obtain historical head-to-head videos of the athlete's opponents, pre-match training videos of the athlete, and shoulder muscle information after the athlete's pre-match training.

[0024] Athlete competitions can include two-person competitive sports such as badminton and tennis.

[0025] Historical head-to-head videos of athletes' opponents refer to video records of past matches between the athletes who will soon be playing against them.

[0026] Historical match videos of athletes against their opponents can be used to analyze the opponent's tactics and playing habits.

[0027] Historical match videos of athletes against their opponents can include full match footage or clips of the opponent in different events.

[0028] The athletes' pre-competition training video is a video taken during the training process of the athletes before the start of the official competition.

[0029] The pre-match training videos record footage of athletes practicing serving, receiving, and footwork. These videos can be used to capture the athletes' current competitive state and the fluidity of their movements.

[0030] The shoulder muscle information obtained by athletes after pre-competition training refers to the physiological state data of the shoulder muscle groups acquired by athletes through physiological monitoring equipment after completing pre-competition training. This information accurately reflects the physiological activation state, fatigue level, and muscle function performance of the athlete's shoulder muscles after pre-competition training.

[0031] Step S2: Based on the athlete's historical head-to-head videos against opponents, the athlete's pre-match training videos, and the athlete's shoulder muscle information after pre-match training, generate multiple simulated head-to-head video information for this match.

[0032] In some embodiments, a head-to-head simulation model can be used to generate multiple head-to-head simulation video information for this match. The head-to-head simulation model is a gated loop unit. The inputs to the head-to-head simulation model are the athlete's historical head-to-head videos against opponents, the athlete's pre-match training videos, and the athlete's shoulder muscle information after pre-match training. The output of the head-to-head simulation model is multiple head-to-head simulation video information for this match.

[0033] A Gated Recurrent Unit (GRU) can be used to process sequential data and temporal information. The GRU controls the flow and forgetting of information through reset and update gates. The reset gate determines whether to ignore historical information, while the update gate adjusts the fusion ratio of historical and current input information. GRUs can more efficiently capture long-term dependencies in sequential data.

[0034] The multiple simulated match videos for this competition were generated using a simulated match model, describing the possible confrontations between the player's team and their opponent in the upcoming match. Each simulated match video in this competition includes a simulated match video and its corresponding credibility value.

[0035] The simulated match video is a video of a game between your team's athletes and their future opponents. The simulated match video demonstrates the hitting trajectories, movement, and tactical maneuvering of both sides.

[0036] The credibility of simulated match videos is a numerical metric used to measure how closely the generated simulated match videos match real-world conditions. The credibility of the simulated match videos reflects the probability of the simulated confrontation occurring.

[0037] Historical match videos of athletes against their opponents provide insights into the opponents' tactical characteristics, while pre-match training videos reveal the athlete's current movement habits. Post-match shoulder muscle information defines the athlete's physiological limits in the simulation. The opponent's behavioral patterns in historical match videos provide the adversarial logic for the model's simulation, the athlete's pre-match training videos provide the model's movement benchmark, and the shoulder muscle information, as a physiological constraint, prevents the generation of movements beyond the athlete's current physical capabilities, thus ensuring the rationality of the simulated match content.

[0038] Gated recurrent units (ROUs) excel at processing sequential data, enabling in-depth analysis of temporal information in historical head-to-head videos of athletes' matches and pre-match training videos. This allows them to capture patterns in movement sequences and tactical variations. Furthermore, the model can incorporate physiological constraints reflected in the athlete's shoulder muscle information after pre-match training, simulating multiple realistic match scenarios to generate multiple simulated match videos for the current competition. By analyzing the completeness, relevance, and rationality of the input data, the GU calculates the credibility of each simulated match video, thus obtaining the information for multiple simulated match videos for the entire competition.

[0039] In some embodiments, generating multiple simulated match video information for this competition based on the athlete's historical head-to-head videos against opponents, the athlete's pre-match training videos, and the athlete's shoulder muscle information after pre-match training includes steps S21-S23:

[0040] Step S21: Based on the historical match videos of the athlete against the opponent, determine the opponent's various offensive action segments, the opponent's various defensive action segments, the frequency of occurrence of each action of the opponent, and the threat score of each action of the opponent.

[0041] In some embodiments, a gated loop unit can be used to determine multiple offensive action segments of the opponent, multiple defensive action segments of the opponent, the frequency of occurrence of each action of the opponent, and the threat score of each action of the opponent.

[0042] The various offensive action segments of the opponent are extracted from the historical head-to-head videos of the athletes' matches against their opponents by the gated loop unit. It is a collection of multiple independent video segments corresponding to different types of offensive behaviors initiated by the opponent. Each video segment records a specific offensive action with a clear offensive intention, execution logic and unique action type.

[0043] The opponent's various defensive action segments are extracted from the historical head-to-head videos of athletes' matches against their opponents by the gated loop unit. It is a collection of multiple independent video segments corresponding to different types of defensive behaviors made by the opponent to resist the attack and protect their own defensive area. Each video segment records a specific defensive action with a clear defensive direction, execution logic and unique action type.

[0044] Each offensive move and each defensive move may occur once or multiple times.

[0045] The frequency of each opponent's move is a quantitative indicator calculated by a gating loop unit from historical match videos of athletes' opponents, representing the frequency of various offensive and defensive moves made by the opponent in historical games. The frequency of each opponent's move can be represented by the percentage of each type of move in the total number of moves.

[0046] The threat score for each opponent's move is a quantitative score derived from analyzing historical match videos of athletes against their opponents using a gated loop unit. This score represents the degree of threat posed by various offensive and defensive moves by the opponent during the match. The higher the score, the stronger the threat level of that type of move.

[0047] The gated loop unit can process temporal data from historical match videos of athletes facing opponents through reset and update gates. The reset gate filters key action-related features from video frames, such as limb movement trajectories and ball movement direction, and accurately locates the start and end frames of actions with offensive or defensive intent, thus forming complete action sequences and ultimately obtaining various offensive and defensive action segments from the opponent. The update gate memorizes the frequency of different action sequences in the video and calculates the proportion of each type of action based on the total video duration or total number of actions. The model can also quantify the threat score of each opponent action based on memorized historical features such as action execution speed and attack range, combined with data on the defensive pressure exerted on the athlete by that type of action in historical match videos.

[0048] Step S22: Based on the athlete's pre-competition training video and the athlete's shoulder muscle information after the pre-competition training, determine the athlete's set of advantageous training movement segments, the athlete's set of disadvantageous training movement segments, the types of movements the athlete's shoulder can withstand, and the athlete's maximum shoulder range of motion.

[0049] In some embodiments, a gated loop unit may be used to determine the set of advantageous movement segments for an athlete's training, the set of disadvantageous movement segments for an athlete's training, the types of movements an athlete's shoulder can withstand, and the maximum range of motion of an athlete's shoulder.

[0050] The set of athlete training strengths segments is extracted from athletes' pre-competition training videos using a gated loop unit. It consists of multiple independent video segments corresponding to different types of movements that the athletes perform exceptionally well in. Each video segment records a specific training movement that is characterized by a high success rate, strong movement standardization, excellent completion efficiency, and is unique in type.

[0051] The set of athlete training weakness segments is extracted from athletes' pre-competition training videos using a gated loop unit. It consists of multiple independent video segments corresponding to different types of movements that the athletes performed poorly. Each video segment records a specific training movement with low execution success rate, weak movement standardization, poor completion efficiency, and unique movement type.

[0052] The types of movements that an athlete's shoulder can withstand are a set of movement categories that the athlete's shoulder muscles and joints can safely bear without the risk of excessive injury, as determined by the gating loop unit.

[0053] The maximum range of motion of an athlete's shoulder is the limit value that the athlete's shoulder joint can reach in each direction of movement (flexion, extension, abduction, adduction, and rotation), determined by the gating loop unit.

[0054] The gated loop unit excels at capturing temporal data correlations. By resetting the gates, it can filter key information such as movement execution details and physiological characteristics in muscle electromyography (EMG) signals from athletes' pre-competition training videos. It can accurately distinguish between high-success-rate, well-proportioned movements and low-success-rate, poorly-proportioned movements, thereby determining the corresponding set of movement segments. By updating the gate memory of the correlation between physiological data such as muscle strength and endurance and movement execution, the gated loop unit can determine the types of movements that the shoulder can safely bear, and by combining this with the temporal characteristics of joint movement, it can determine the athlete's maximum shoulder range of motion.

[0055] Step S23: Based on the opponent's various offensive action segments, the opponent's various defensive action segments, the frequency of occurrence of each action of the opponent, the threat score of each action of the opponent, the athlete's set of training advantage action segments, the athlete's set of training weakness action segments, the types of actions the athlete's shoulder can withstand, and the athlete's maximum shoulder range of motion, determine multiple simulated video information of the match.

[0056] In some embodiments, a gated loop unit can be used to determine multiple simulated match video information for this game.

[0057] The gated loop unit possesses temporal correlation modeling and multi-source information fusion capabilities. By resetting the gate, the gated loop unit can uncover the inherent correlation between the technical characteristics, frequency of occurrence, and threat scores of various opponent action segments. Simultaneously, it can correlate the execution characteristics of athletes' training strengths and weaknesses, thereby filtering out high-probability action confrontation scenarios during the match. The update gate can memorize the physiological boundaries of the types of movements and maximum range of motion that athletes' shoulders can withstand, and perform action adaptation and intensity adjustment on the selected confrontation scenarios. Thus, the gated loop unit can simulate a confrontation process that conforms to the technical characteristics of both sides and the physiological tolerance of the athletes, ultimately generating multiple simulated confrontation video messages that closely resemble actual match situations.

[0058] Step S3: Determine the distribution map of shoulder muscle soreness based on the shoulder muscle information of the athlete after the pre-competition training.

[0059] In some embodiments, a muscle state analysis model can be used to determine the distribution map of shoulder muscle soreness. The muscle state analysis model is a Transformer model. The input to the muscle state analysis model is the shoulder muscle information of the athlete after pre-competition training, and the output of the muscle state analysis model is the distribution map of shoulder muscle soreness.

[0060] The Transformer model consists of an encoder and a decoder. The encoder learns representations from the input sequence, incorporating self-attention and a feed-forward network. The decoder, building upon the encoder, introduces an additional multi-head attention mechanism to decode the encoder output and generate the target sequence. The Transformer model can process sequence data in parallel and focuses on the relationships between different positions in the input data.

[0061] The shoulder muscle soreness distribution map is a spatial image output by a muscle condition analysis model that visually displays the degree of soreness in different areas of an athlete's shoulder. The map uses color depth and numerical labels to differentiate the physiological state of soreness in different muscle regions.

[0062] The strength of electromyographic (EMG) signals in the shoulder muscles after pre-competition training is directly related to the degree of muscle fatigue. High-frequency, high-amplitude EMG signals often correspond to muscle tension, while changes in the signal spectrum reflect the accumulation of metabolic products. These microscopic physiological electrical signal data provide a basis for the model to convert from the EMG signal domain to a distribution map of shoulder muscle soreness.

[0063] The Transformer model's self-attention mechanism enables comprehensive feature extraction of shoulder muscle information after an athlete's pre-competition training, capturing the correlation and distribution patterns between electrical signal data from different muscle regions. The model focuses on the spatial correlation between signals from different acquisition points and the trend of signal changes over time. The encoder maps discrete electromyographic signals into high-dimensional feature vectors, while the decoder transforms these feature vectors into a two-dimensional distribution representation of the corresponding shoulder anatomical structure. Through multi-head attention, the Transformer model accurately locates muscle bundles with abnormal signals and calculates specific soreness values ​​based on signal intensity, ultimately generating a shoulder muscle soreness distribution map reflecting the overall physiological state of the shoulder.

[0064] Step S4: Based on the multiple simulated match video information of this competition, use the stretching information processing model to determine the multiple shoulder stretching information of each simulated match video.

[0065] The stretching information processing model is a recurrent neural network. The input to the stretching information processing model is multiple simulated match videos from this competition, and the output of the stretching information model is multiple shoulder stretching data from each simulated match video.

[0066] Recurrent Neural Networks (RNNs) are neural network models specifically designed for processing sequential data. In an RNN, the output of the hidden layers is fed back to the input and participates in the computation of the next time step. Through this recurrent structure, RNNs can remember information from previous time steps, thus capturing temporal dependencies in sequential data. RNNs are suitable for processing sequential information such as video frame sequences.

[0067] The shoulder stretching information in the simulated competition video is detailed information about each stretching movement of the athlete's shoulder during the simulated competition, output by a stretching information processing model. Each stretching information includes the stretching angle, stretching speed, stretching duration, starting position of the stretching movement, ending position of the stretching movement, muscle force exertion during the stretch, and the continuity level of the stretching movement.

[0068] The way muscles exert force during stretching refers to the specific form of muscle contraction or relaxation when an athlete performs a stretching movement in the shoulder, such as active stretching, passive stretching, and explosive stretching.

[0069] The stretching motion continuity level is a numerical indicator of the smoothness of the connection between multiple stretching movements in an athlete's shoulder. The stretching motion continuity level can reflect whether the transition between adjacent stretching movements is smooth.

[0070] Each of the multiple simulated combat videos in this competition fully presents the simulated combat scenario, including various shoulder movement sequences of the athletes. As visual input, the simulated combat videos can provide the model with the time and space information needed to determine the type of movement.

[0071] Recurrent neural networks (RNNs) can scan and analyze every frame of each simulated sparring video. Through the memory function of the recurrent units, the RNN can track the positional changes of key points on the athlete's shoulder joint across consecutive frames. When the shoulder joint angle exceeds the threshold for normal, smooth movement or when a large swing occurs, the RNN can identify and record the movement characteristics for that time period. By integrating contextual information from preceding and following frames, the RNN can calculate the start and end points of the movement, thereby quantifying the specific parameters of the stretch and ultimately determining the multiple shoulder stretches in each simulated sparring video.

[0072] Step S5: Based on the shoulder muscle soreness distribution map and the shoulder stretching information of each sparring simulation video, determine the shoulder strain probability distribution map of each sparring simulation video.

[0073] In some embodiments, Figure 2 This is a flowchart illustrating a method for determining the probability distribution of shoulder strain in each simulated combat video, as provided in an embodiment of the present invention. The determination of the probability distribution of shoulder strain in each simulated combat video includes steps S51 to S53:

[0074] Step S51: Based on the shoulder stretching information of each simulated hand-to-hand combat video, cluster K clusters are obtained.

[0075] The clustering method used is K-means clustering, an unsupervised learning algorithm that divides an input dataset into K discrete clusters based on the similarity of data features. The value of K can be preset manually. K-means clustering uses the distance between data points and the centers of each cluster as a similarity metric, iteratively updating the cluster center positions until the similarity of data points within a cluster is maximized and the similarity between data points between clusters is minimized, ultimately yielding a stable clustering result.

[0076] The K clusters are K sets formed by grouping the shoulder stretching information from each simulated combat video according to similarity. Each cluster contains shoulder stretching information that is highly similar in core features such as stretching angle, stretching force, and stretching speed, and represents a type of stretching movement with similar characteristics. There are significant differences in the core features of shoulder stretching information between different clusters, and different clusters correspond to different types of stretching movement patterns.

[0077] The process of using the K-means clustering algorithm to cluster the shoulder stretching information from each simulated combat video into K clusters is as follows: First, K data samples are randomly selected from the shoulder stretching information of each simulated combat video as initial cluster centers. Then, the similarity distance between all remaining shoulder stretching information samples and each initial cluster center is calculated, and each sample is assigned to the cluster containing the nearest cluster center. Next, based on all sample data within each cluster, the center position of each cluster is recalculated, and the cluster centers are updated. This process of sample assignment and cluster center update is repeated until the cluster center positions no longer change significantly or a preset number of iterations is reached. The clustering process then terminates, resulting in K stable clusters.

[0078] Clustering into K clusters allows for the categorization and organization of massive amounts of scattered shoulder stretching information based on feature similarity, reducing data dimensionality and analytical complexity. Stretching information within the same cluster exhibits uniform characteristics, enabling the centralized classification of similar stretching movements and resulting in more regular data presentation. Clear feature differences exist between different clusters, allowing for the differentiation of different stretching movement patterns and avoiding feature confusion caused by the chaotic distribution of various stretching information. This provides a clear and orderly data foundation for further analysis.

[0079] Step S52: Determine multiple high-risk stretching information based on the K clusters.

[0080] In some embodiments, a hazardous tensile analysis model can be used to determine multiple high-risk tensile information. The hazardous tensile analysis model is a Transformer model. The input to the hazardous tensile analysis model is the K clusters, and the output of the hazardous tensile analysis model is multiple high-risk tensile information.

[0081] The high-risk stretching information includes detailed descriptions of extreme or improper stretching movements identified by the hazardous stretching analysis model as potentially causing injury to athletes. This information includes the type of high-risk movement, the frequency of occurrence of the high-risk movement, the direction of force applied to the high-risk movement, and the duration of muscle stretching associated with the high-risk movement.

[0082] High-risk movement types include excessive shoulder abduction stretching, abrupt shoulder supination stretching, shoulder flexion stretching beyond the physiological range, explosive stretching under weight-bearing conditions, and asymmetrical shoulder twisting stretching.

[0083] The frequency of high-risk movements refers to the number of times each type of high-risk stretching movement occurs in the combat scenario corresponding to a single combat simulation video.

[0084] The direction of force corresponding to high-risk movements refers to the specific direction in which the force acts on the shoulder muscles during high-risk stretching movements, such as longitudinal traction, lateral tearing, and oblique twisting.

[0085] The duration of muscle stretching in high-risk movements refers to the length of time that a single high-risk stretching movement continuously acts on the shoulder muscles.

[0086] K clusters organize the originally chaotic stretching data into statistically significant categories, giving the data a structured distribution of characteristics. Different clusters represent movement patterns of varying intensities, including anomalous data groups that deviate from normal physiological ranges. These clustered datasets highlight the characteristics of extreme movements, thus providing a clear classification basis for models to identify potential risks.

[0087] The Transformer model, through a global attention mechanism, can deeply analyze the internal features of each of K clusters. It can compare parameters such as the stretching amplitude and duration within a cluster with known sports injury biomechanics models. The encoder extracts cluster center features and identifies stretching patterns that exceed the safe tolerance range of human joints. The decoder then transforms these marked dangerous patterns into specific textual or numerical descriptions, thereby filtering and identifying multiple high-risk stretching patterns.

[0088] Step S53: Based on the multiple high-risk stretching information and the shoulder muscle soreness distribution map, determine the shoulder strain probability distribution map for each combat simulation video.

[0089] In some embodiments, a strain probability determination model can be used to determine the shoulder strain probability distribution map for each sparring simulation video. The strain probability determination model is a Transformer model. The inputs to the strain probability determination model are the multiple high-risk stretch information and the shoulder muscle soreness distribution map, and the output of the strain probability determination model is the shoulder strain probability distribution map for each sparring simulation video.

[0090] The shoulder strain probability distribution map for each simulated match is a spatial distribution image of the probability of strains occurring in different parts of the shoulder during the simulated match, output by the strain probability determination model.

[0091] Multiple high-risk stretching profiles clearly identify the characteristics of high-risk stretching movements. These characteristics directly point to key movement factors that may lead to shoulder strain, thus providing a concrete basis for assessing the risk of shoulder muscle injury from different high-risk stretching movements. The shoulder muscle soreness distribution map shows the baseline soreness state of different areas of the shoulder; the degree of soreness directly reflects muscle tolerance.

[0092] The Transformer model can simultaneously process data from multiple high-risk stretching events and shoulder muscle soreness distribution maps. Through its self-attention mechanism, the Transformer model captures the correlation between risk characteristics in high-risk stretching events and the soreness status of different regions in the shoulder muscle soreness distribution map. The model can analyze the load impact of each high-risk stretching action on different areas of the shoulder, and then, combined with the baseline soreness status of the corresponding regions in the shoulder muscle soreness distribution map, it can quantify the probability of strain in each region under the simulated combat scenario. The model can then map this probability back to the anatomical structure of the shoulder, ultimately generating a shoulder strain probability distribution map for each simulated combat video that reflects dynamic risk.

[0093] Step S6: Determine the shoulder strain risk value for this match based on the shoulder strain probability distribution map of each simulated match video using a graph neural network.

[0094] In some embodiments, Figure 3 This invention provides a flowchart illustrating the process for determining the shoulder strain risk value in a competition, comprising steps S61-S62:

[0095] Step S61: Construct a muscle strain analysis map. The muscle strain analysis map includes multiple simulated combat nodes and multiple edges between the simulated combat nodes. The node features of each simulated combat node include information from each simulated combat video, a probability distribution map of shoulder strain in each simulated combat video, shoulder muscle information of the athlete after pre-competition training, and a distribution map of shoulder muscle soreness. The edges between simulated combat nodes represent the credibility difference between simulated combat videos.

[0096] A muscle strain analysis graph is a graphical data structure that describes the relationships and overall risk characteristics between different simulated match scenarios. The muscle strain analysis graph includes multiple simulated match nodes and multiple edges between these nodes, organizing all simulation data and risk assessment results in a structured graph format.

[0097] Multiple engagement simulation nodes are the core components of the injury analysis graph. Each node corresponds to an engagement simulation scenario, and the edges between engagement simulation nodes represent the confidence difference between the engagement simulation videos associated with the corresponding two engagement simulation nodes.

[0098] Step S62: Process the strain analysis map based on graph neural network to obtain the shoulder strain risk value for this competition.

[0099] Graph Neural Networks (GNNs) are deep learning models that can run directly on a graph. GNNs update node feature representations through a neighborhood aggregation mechanism, or message passing mechanism. Each node in the graph can collect and aggregate information from its neighbors, thus fusing local structure and global contextual information. GNNs can process data in non-Euclidean space and effectively capture dependencies and topological features between nodes. The input to the GNN is the strain analysis graph, and the output is the shoulder strain risk value for this competition.

[0100] The shoulder strain risk value for this competition was determined by a graph neural network and is a numerical indicator used to quantitatively assess the degree of risk of shoulder strain for our athletes during this competition.

[0101] The shoulder strain risk value for this competition is a value within a specific range. The higher the risk value, the greater the likelihood that the athlete will strain their shoulder muscles during the competition, and the closer the exercise load on the shoulder muscles is to the critical state of injury.

[0102] The shoulder strain analysis atlas can graphically represent all information related to the shoulder strain risk in this competition, and can intuitively present multiple simulated competition scenarios and their relationships. Each simulated competition node represents the features required for strain risk analysis in the corresponding scenario, including information from each simulated competition video, a shoulder strain probability distribution map for each simulated competition video, shoulder muscle information after the athlete's pre-competition training, and a shoulder muscle soreness distribution map. The edges between nodes are based on the confidence difference between the simulated competition videos associated with the corresponding two simulated competition nodes. This structured representation method facilitates graph neural networks in understanding and processing the complex relationships between scenarios.

[0103] Through the synergistic support of node features and edges, graph neural networks can more accurately capture the differences in risk characteristics across different simulated competition scenarios, as well as the mutual influence caused by the difference in credibility between scenarios. The shoulder strain probability distribution map in the node features reflects the risk distribution in different areas of the shoulder under the corresponding scenario, while the credibility difference on the edges reflects the reliability correlation of risk data in different scenarios. The construction of the strain analysis map can effectively organize scattered risk-related data and scenario relationships, providing a clear data structure for graph neural network processing. This structured organization also reduces the complexity of data processing, thereby improving the training and inference efficiency of the model. Simultaneously, with the structured node and edge information provided by the strain analysis map, the graph neural network model can process multi-scenario risk data more efficiently, avoiding the data sparsity or dimensional redundancy problems that may occur in traditional risk assessment methods. Relying on the structured advantages of the graph, graph neural networks are more targeted in processing complex relationships between scenarios. By comprehensively analyzing the features and relationships of all simulated competition nodes, they can more accurately and comprehensively integrate multi-scenario risk information. Ultimately, the graph neural network can accurately calculate the risk value that objectively reflects the probability of a shoulder strain in this competition.

[0104] Graph neural networks (Graph Neural Networks) can perform deep processing of shoulder strain analysis graphs. By aggregating feature information from each simulated match node, Graph Neural Networks can uncover the correlations between various data within each node. Furthermore, they can utilize the confidence difference information represented by edges between nodes to weigh the weights of different simulated match nodes in risk assessment. Through a multi-layered network structure, Graph Neural Networks extract and fuse feature information and correlations layer by layer from the graph, comprehensively considering the strain risk characteristics under all simulated match scenarios and the confidence weights of each scenario. Ultimately, they can calculate a quantitative value that comprehensively reflects the degree of shoulder strain risk in this match—the shoulder strain risk value for this match.

[0105] Step S7: If the shoulder strain risk value of this competition is greater than the preset threshold, a risk warning signal is sent to the control terminal.

[0106] Once the shoulder strain risk value for this competition is determined, the risk value is assessed. If the shoulder strain risk value for this competition is greater than a preset threshold, a risk warning signal is sent to the control terminal to remind the athlete to participate with caution. Figure 4 This is a schematic diagram of a control terminal provided in an embodiment of the present invention.

[0107] Based on the same inventive concept Figure 5 This is a schematic diagram of a sports event risk prediction system based on image vision, provided in an embodiment of the present invention. The sports event risk prediction system based on image vision includes:

[0108] The acquisition module 81 is used to acquire historical head-to-head videos of the athlete's opponents in the competition, pre-competition training videos of the athlete, and shoulder muscle information of the athlete after the pre-competition training.

[0109] The simulation module 82 is used to generate multiple simulated match video information for this match based on the athlete's historical head-to-head videos against opponents, the athlete's pre-match training videos, and the athlete's shoulder muscle information after the pre-match training.

[0110] The soreness distribution map determination module 83 is used to determine the soreness distribution map of the shoulder muscles based on the shoulder muscle information of the athlete after the pre-competition training.

[0111] The stretching information determination module 84 is used to determine the multiple shoulder stretching information of each simulated match video based on the multiple simulated match video information of this competition using the stretching information processing model.

[0112] The strain probability distribution map determination module 85 is used to determine the shoulder strain probability distribution map for each sparring simulation video based on the shoulder muscle soreness distribution map and the shoulder multiple stretching information of each sparring simulation video.

[0113] The shoulder strain risk determination module 86 is used to determine the shoulder strain risk value for this match based on the shoulder strain probability distribution map of each simulated match video using a graph neural network.

[0114] The early warning module 87 is used to send a risk warning signal to the control terminal if the shoulder strain risk value in this competition is greater than a preset threshold.

[0115] It should be noted that, in order to simplify the descriptions disclosed herein and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of embodiments of this specification may sometimes combine multiple features into a single embodiment, drawing, or description thereof. However, this method of disclosure does not imply that the subject matter of this specification requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of a single embodiment disclosed above.

[0116] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A method for predicting the risk of sports events based on image vision, characterized in that, include: Obtain historical head-to-head videos of athletes' opponents, pre-match training videos of athletes, and shoulder muscle information of athletes after pre-match training; Based on the athlete's historical head-to-head videos against opponents, the athlete's pre-match training videos, and the athlete's shoulder muscle information after pre-match training, multiple simulated head-to-head video information for this match is generated. A distribution map of shoulder muscle soreness was determined based on the shoulder muscle information of the athletes after their pre-competition training. Based on the simulated video information of multiple matches in this competition, the stretching information processing model was used to determine the multiple shoulder stretching information of each simulated match video. Based on the shoulder muscle soreness distribution map and the shoulder stretching information from each sparring simulation video, a shoulder strain probability distribution map for each sparring simulation video is determined. This determination includes: K clusters were obtained by clustering the shoulder stretching information from each simulated engagement video. Multiple high-risk stretching information is determined based on the K clusters; Based on the multiple high-risk stretching information and the shoulder muscle soreness distribution map, a shoulder strain probability distribution map is determined for each combat simulation video; The shoulder strain risk value for this match is determined based on the shoulder strain probability distribution map of each simulated match video using a graph neural network. This determination includes: A muscle strain analysis atlas is constructed, which includes multiple simulated competition nodes and multiple edges between the simulated competition nodes. The node features of each simulated competition node include information from each simulated competition video, a probability distribution map of shoulder strain in each simulated competition video, shoulder muscle information of the athlete after pre-competition training, and a distribution map of shoulder muscle soreness. The edges between the simulated competition nodes represent the credibility difference between the simulated competition videos. The shoulder strain risk value for this competition was obtained by processing the strain analysis map using a graph neural network. If the risk of shoulder strain in this competition exceeds a preset threshold, a risk warning signal will be sent to the control terminal.

2. The sports event risk prediction method based on image vision as described in claim 1, characterized in that, The information for each simulated match in this competition includes the simulated match video itself and the credibility of the simulated match video.

3. A sports event risk prediction system based on image vision, characterized in that, include: The acquisition module is used to acquire historical head-to-head videos of athletes' opponents in competitions, pre-competition training videos of athletes, and shoulder muscle information of athletes after pre-competition training. The simulation module is used to generate multiple simulated match videos for this competition based on the athlete's historical head-to-head videos against opponents, the athlete's pre-match training videos, and the athlete's shoulder muscle information after the pre-match training. The soreness distribution map determination module is used to determine the soreness distribution map of the shoulder muscles based on the shoulder muscle information of the athlete after the pre-competition training. The stretching information determination module is used to determine the multiple shoulder stretching information in each simulated match video based on the multiple match simulation video information of this competition using the stretching information processing model; The strain probability distribution map determination module is used to determine the shoulder strain probability distribution map for each sparring simulation video based on the shoulder muscle soreness distribution map and the shoulder multiple stretching information of each sparring simulation video. The strain probability distribution map determination module is also used for: K clusters were obtained by clustering the shoulder stretching information from each simulated engagement video. Multiple high-risk stretching information is determined based on the K clusters; Based on the multiple high-risk stretching information and the shoulder muscle soreness distribution map, a shoulder strain probability distribution map is determined for each combat simulation video; The shoulder strain risk determination module is used to determine the shoulder strain risk value for this match based on the shoulder strain probability distribution map of each simulated match video using a graph neural network. The shoulder strain risk determination module is also used for: A muscle strain analysis atlas is constructed, which includes multiple simulated competition nodes and multiple edges between the simulated competition nodes. The node features of each simulated competition node include information from each simulated competition video, a probability distribution map of shoulder strain in each simulated competition video, shoulder muscle information of the athlete after pre-competition training, and a distribution map of shoulder muscle soreness. The edges between the simulated competition nodes represent the credibility difference between the simulated competition videos. The shoulder strain risk value for this competition was obtained by processing the strain analysis map using a graph neural network. The early warning module is used to send a risk warning signal to the control terminal if the shoulder strain risk value in this competition is greater than a preset threshold.

4. The image-based sports event risk prediction system as described in claim 3, characterized in that, The information for each simulated match in this competition includes the simulated match video itself and the credibility of the simulated match video.

5. An electronic device, characterized in that, include: processor; Memory; And a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the image vision-based sports event risk prediction method as described in any one of claims 1 to 2.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the image-based sports event risk prediction method as described in any one of claims 1 to 2.

Images