Intelligent recommendation method for gymnastics teaching resources based on knowledge graph
By constructing a multi-level knowledge graph of gymnastics movements and a convolutional neural network scoring model, the problem of personalized recommendation of gymnastics teaching resources was solved, achieving highly targeted recommendation of teaching resources and improving learning outcomes.
Patent Information
- Application Number
- CN202511355923.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-22
- Publication Date
- 2026-08-25
- Estimated Expiration
- 2045-09-22
AI Technical Summary
Existing technologies lack personalized recommendations for different athletes in gymnastics teaching resource recommendations, resulting in insufficiently targeted teaching services.
Based on knowledge graphs, single and multi-level knowledge graphs of gymnastics movements are constructed. Combined with convolutional neural networks, frame decomposition and scoring of user-recorded videos are performed to build a scoring model and recommend the most suitable gymnastics movements for users.
It enables personalized gymnastics movement recommendations based on the user's athletic ability, improving the relevance of teaching resources and learning efficiency.
Smart Images
Figure CN121190267B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of resource recommendation technology, specifically a knowledge graph-based intelligent recommendation method for gymnastics teaching resources. Background Technology
[0002] With the increasing abundance of digital teaching resources, the key issue is how to accurately recommend suitable teaching resources to athletes from a vast amount of resources. Existing technologies have emerged that utilize the powerful semantic expression and reasoning capabilities of knowledge graphs to achieve efficient organization and personalized recommendation of gymnastics teaching resources, which can significantly improve teaching effectiveness and learning efficiency. However, existing technologies often only provide simple information services for the utilization of teaching resources, lacking technical solutions to provide subsequent resource recommendations based on the current athletic abilities of different athletes. This results in the failure to achieve more effective and targeted teaching services. To address this, the present invention provides an intelligent recommendation method for gymnastics teaching resources based on knowledge graphs. Summary of the Invention
[0003] The purpose of this invention is to provide an intelligent recommendation method for gymnastics teaching resources based on knowledge graphs.
[0004] The objective of this invention can be achieved through the following technical solution: an intelligent recommendation method for gymnastics teaching resources based on knowledge graphs, comprising the following steps: Step S1: Collect teaching resources for different gymnastic movements, and construct a first knowledge graph for a single gymnastic movement and a second knowledge graph for different gymnastic movements. Step S2: Obtain the user's first recorded video and decompose it into frames to obtain multiple keyframes. Perform action scoring on each keyframe to obtain the first score of the first recorded video. Construct a first scoring model based on the first recorded video of different users and its corresponding keyframes and first scores. Step S3: Obtain the user's set of qualified movements based on the first score, obtain the user's self-selected gymnastics movements by combining the second knowledge graph, and obtain the second recorded video and its second score. Construct a second score model based on the second scores of different self-selected gymnastics movements and the first scores of the corresponding set of qualified movements. Step S4: Use the second scoring model to combine the user's set of qualified movements with the first score to obtain the user's predicted scores for different gymnastic movements, and obtain recommended gymnastic movements.
[0005] Furthermore, the process of collecting teaching resources for different gymnastic movements includes: Gymnastics movements are divided into different major categories, including tumbling, handstands, handsprings, somersaults, horizontal bar, parallel bars, and vault, and each major category contains different subcategories; Obtain teaching resources corresponding to different subcategories of gymnastics movements. The teaching resources include the movement names, text descriptions, teaching points, and corresponding breakdown diagrams for each step of a single subcategory.
[0006] Furthermore, the process of constructing the first knowledge graph for a single gymnastic movement and the second knowledge graph for different gymnastic movements includes: The teaching resources corresponding to each subcategory of gymnastics movements are converted into a unified format and semantics, and the relationships between various entities in the teaching resources are established. Ontology is used to define the various entities and their relationships in the teaching resources to obtain the semantic structure of the knowledge graph. Natural language processing techniques are used to extract key information from text data, reasoning is performed on the key information to fill the gaps in the knowledge graph, and the data structure of the knowledge graph is constructed by combining the obtained semantic structure, and the obtained key information is filled into the knowledge graph. The knowledge graph constructed above is used as the first knowledge graph of the corresponding sub-category of gymnastic movements. Knowledge graph fusion technology is used to merge the first knowledge graphs of various sub-categories belonging to the same major category into the second knowledge graph of the corresponding major category.
[0007] Furthermore, the process of acquiring the user's first recorded video, decomposing it into frames to obtain multiple keyframes, and performing motion scoring on each keyframe to obtain the first score of the first recorded video includes: Set up an input unit and acquire the first recorded video of a user performing a single sub-category of gymnastic movements. Perform frame decomposition processing on the first recorded video to obtain the corresponding video frames. The user selects the corresponding key frames from all the video frames of the first recorded video according to the decomposition steps of the single sub-category. Obtain the key frame interval between the timestamp of each key frame and the start time of the first recorded video. OpenPose is used to extract the coordinates of multiple human key points in a single keyframe and its step breakdown diagram, and the Euclidean distance between two coordinates of the same human key point in the single keyframe and its step breakdown diagram is obtained. The mean Euclidean distance between the two coordinates of each human body key point is used as the similarity between the single keyframe and its step decomposition diagram. The mean similarity between each keyframe belonging to the same first recorded video and its corresponding decomposition step diagram is used as the first score of the gymnastics movement of the corresponding subcategory of the first recorded video.
[0008] Furthermore, the process of constructing a first rating model based on the first recorded video of different users and its corresponding keyframes and first rating includes: Obtain the first recorded videos of gymnastics movements for different subcategories from different users, and obtain each keyframe and its decomposition step diagram from the first recorded videos of different subcategories. Combine the first scores of the first recorded videos to generate a first score set, and divide it into a first training set and a first test set. Construct a first convolutional neural network by using different first recorded videos and corresponding decomposition step diagrams of each subclass in the first training set as input data for the first convolutional neural network, and the first score and each keyframe in the first training set as output data for the first convolutional neural network. Train the first convolutional neural network using the first training set to obtain the initial first convolutional neural network. The initial first convolutional neural network is validated using the first test set, and the initial first convolutional neural network whose output is less than or equal to the preset first test error threshold is used as the first scoring model.
[0009] Furthermore, the process of obtaining the user's set of qualified movements based on the first score, obtaining the user's self-selected gymnastics movements by combining the second knowledge graph, and obtaining the second recorded video and its second score includes: The first recorded video of a single user performing gymnastic movements in different subcategories, along with the corresponding breakdown steps of each subcategory, is input into the first scoring model to obtain the user's first score for each subcategory of gymnastic movements. The first score of a single user for different subcategories of gymnastics movements is compared with a preset pass threshold to obtain the user's pass set. If any gymnastics movement contained in the second knowledge graph intersects with the user's pass set, then the gymnastics movements of each subcategory in the second knowledge graph other than the intersection are taken as the user's candidate gymnastics movements. The user selects one gymnastics move from all available options as their self-selected gymnastics move. The recording unit obtains a second recorded video of the user's self-selected gymnastics move, obtains a second score corresponding to the second recorded video, and binds it to the user's current set of qualified moves.
[0010] Furthermore, the process of constructing a second scoring model based on the second scores of different self-selected gymnastics movements and the first scores of their corresponding sets of qualifying movements includes: A second score set is generated based on the second score of different users for their self-selected gymnastics movements and the first score of each gymnastics movement in the corresponding set of qualified movements, and then divided into a second training set and a second test set. Construct a second convolutional neural network by taking the first score of each gymnastics movement in the set of qualified movements of different users in the second training set and their self-selected gymnastics movements as the input data of the second convolutional neural network, and taking the corresponding second score in the second training set as the output data of the second convolutional neural network. Train the second convolutional neural network using the second training set to obtain the initial second convolutional neural network. The initial second convolutional neural network is validated using the second test set. The initial second convolutional neural network whose output is less than or equal to the preset second test error threshold is used as the second scoring model.
[0011] Furthermore, the process of using a second scoring model to combine the user's set of qualified movements with the first score to obtain the user's predicted scores for different gymnastic movements, and then obtaining recommended gymnastic movements, includes: In subsequent application scenarios, the first score of each gymnastics movement in a single user's current set of qualified movements, as well as the other gymnastics movements, are input into the second scoring model to obtain the user's predicted scores for the other gymnastics movements. The gymnastics movement corresponding to the highest predicted score is then used as the user's recommended gymnastics movement and fed back to the user.
[0012] Compared with the prior art, the beneficial effects of the present invention are: This invention provides users with educational resources by dividing gymnastics movements into different major categories and further subcategories, and constructing knowledge graphs for each category. By acquiring the user's first recorded video and first score for each gymnastics movement, the invention can identify the gymnastics movements the user is proficient in. By acquiring the user's second score for a self-selected gymnastics movement and constructing a second scoring model, the invention can determine the difficulty of learning other gymnastics movements based on the user's proficiency in those movements, and thus provide the easiest gymnastics movements to learn. Based on the user's current athletic ability, the invention can recommend targeted gymnastics movements, which is conducive to forming an effective gymnastics teaching resource recommendation mechanism. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation
[0014] like Figure 1 As shown, the intelligent recommendation method for gymnastics teaching resources based on knowledge graphs includes the following steps: Step S1: Collect teaching resources for different gymnastic movements, and construct a first knowledge graph for a single gymnastic movement and a second knowledge graph for different gymnastic movements. Step S2: Obtain the user's first recorded video and decompose it into frames to obtain multiple keyframes. Perform action scoring on each keyframe to obtain the first score of the first recorded video. Construct a first scoring model based on the first recorded video of different users and its corresponding keyframes and first scores. Step S3: Obtain the user's set of qualified movements based on the first score, obtain the user's self-selected gymnastics movements by combining the second knowledge graph, and obtain the second recorded video and its second score. Construct a second score model based on the second scores of different self-selected gymnastics movements and the first scores of the corresponding set of qualified movements. Step S4: Use the second scoring model to combine the user's set of qualified movements with the first score to obtain the user's predicted scores for different gymnastic movements, and obtain recommended gymnastic movements.
[0015] It should be further explained that, in the specific implementation process, the collection of teaching resources for different gymnastic movements includes: The gymnastic movements refer to individual or combined physical activities performed by athletes using various parts of their bodies in accordance with specific rules and requirements during gymnastics. In embodiments of the present invention, gymnastic movements are divided into different categories, including tumbling, handstands, handsprings, somersaults, horizontal bar, parallel bars, vaulting, etc. Specifically, each major category contains different subcategories. The tumbling category includes forward tumbling and backward tumbling; the handstand category includes headstand and handstand; the tumbling category includes side handstand, forward handstand, and backward handstand; and the somersault category includes forward somersault and backward somersault. The horizontal bar category includes hanging swing, push-off from the ground bar, support back loop, and straddle vault with reverse grab; the parallel bars category includes support swing, straddle sit forward, shoulder stand, hanging arm flexion and extension, and support back swing into handstand; the vault category includes vaulting into kneeling support, straddle vault, bent leg vault, and front handspring vault. The teaching resources refer to the sum of various elements that can provide support and assistance to users in learning gymnastics movements during the teaching process. They include the movement name, text description, teaching points and corresponding breakdown diagrams of each step of a single subcategory. For example, a forward roll can be broken down into steps such as squatting, tucking the head down and rolling, and hugging the legs to form a squatting position. The teaching resources include the movement names, text descriptions, teaching points and diagrams of each step. The same method can be used to obtain teaching resources for different subcategories of gymnastic movements.
[0016] It should be further explained that, in the specific implementation process, the process of constructing the first knowledge graph for a single gymnastic movement and the second knowledge graph for different gymnastic movements includes: The teaching resources (excluding step-by-step diagrams) corresponding to a single category of gymnastic movements are converted into a unified format and semantics to facilitate representation in the knowledge graph. The relationships between various entities in the teaching resources are established, and ontology is used to define the various entities and their relationships in the teaching resources to obtain the semantic structure of the knowledge graph. Natural language processing techniques are used to extract key information from text data, reasoning is performed on the key information to fill the gaps in the knowledge graph, and the data structure of the knowledge graph is constructed by combining the obtained semantic structure, and the obtained key information is filled into the knowledge graph. In the knowledge graph, the teaching resources of this single subcategory of gymnastic movements (excluding the decomposed step diagram) are divided into nodes and edges for separate processing. The nodes are the basic units in the knowledge graph, representing entities in the teaching resources. The edges connect the relationships between nodes and are used to represent the associations between nodes. The edges describe the semantic structure between nodes. Since the knowledge graph constructed above is based on teaching resources for a single subcategory of gymnastic movements, it is used as the first knowledge graph for the corresponding subcategory of gymnastic movements. The same method is used to construct first knowledge graphs for different subcategories of gymnastic movements. Knowledge graph fusion technology is used to merge the first knowledge graphs of various subcategories belonging to the same major category into a second knowledge graph for the corresponding major category. The same method is used to obtain second knowledge graphs for different major categories of gymnastic movements.
[0017] It should be further explained that, in the specific implementation process, the process of acquiring the user's first recorded video, decomposing it into frames to obtain multiple keyframes, and performing action scoring on each keyframe to obtain the first score of the first recorded video includes: An input unit is set up to acquire videos recorded by the user for a single subcategory of gymnastic movements, which are denoted as the first recorded video. The first recorded video is subjected to frame decomposition processing, which means dividing the continuous video into a series of individual segments of fixed time length, and marking the segmented segments as video frames. Each video frame has its corresponding timestamp. The user selects the corresponding video frames from all video frames of the first recorded video according to the decomposition step diagram of the single sub-category, and records them as key frames. The number of key frames is the same as the number of decomposition step diagrams. The time interval between the timestamp of each key frame and the start time of the first recorded video is obtained and recorded as the key frame interval. The motion scoring refers to the process of comparing each keyframe belonging to the same first recorded video with its corresponding decomposition step map to obtain the first score of the first recorded video. OpenPose is used to extract the coordinates of multiple human key points in a single keyframe and its corresponding decomposition step map. OpenPose is an open-source real-time multi-person pose estimation library that can accurately extract the coordinates of 18 / 25 human key points from an image (e.g., (x1, y1) represents the left shoulder, (x2, y2) represents the right shoulder, etc.). The Euclidean distance between the two coordinates of the same human body key point in the single key frame and its step decomposition diagram is obtained respectively. The average value of the Euclidean distance between the two coordinates of each human body key point is used as the similarity between the single key frame and its step decomposition diagram. The average value of the similarity between each key frame belonging to the same first recorded video and its corresponding decomposition step diagram is used as the first score of the gymnastics movement of the sub-category corresponding to the first recorded video.
[0018] It should be further explained that, in the specific implementation process, the process of constructing the first scoring model based on the first recorded video of different users and its corresponding keyframes and first score includes: Obtain the first recorded videos of gymnastics movements for different subcategories from different users, and obtain each keyframe and its decomposition step diagram in the first recorded videos of different subcategories. Combine the first scores corresponding to the first recorded videos to generate the first score set, and divide it into the first training set and the first test set. Construct a first convolutional neural network by using different first recorded videos and corresponding decomposition step diagrams of each subclass in the first training set as input data for the first convolutional neural network, and the first score and each keyframe in the first training set as output data for the first convolutional neural network. Train the first convolutional neural network using the first training set to obtain the initial first convolutional neural network. The initial first convolutional neural network is validated using the first test set. The initial first convolutional neural network whose output is less than or equal to the preset first test error threshold is used as the first scoring model. The first scoring model can output its corresponding first score and each key frame based on the input first recorded video and the decomposition step diagram of the corresponding subclass.
[0019] It should be further explained that, in the specific implementation process, the process of obtaining the user's set of qualified movements based on the first score, obtaining the user's self-selected gymnastics movements by combining the second knowledge graph, and obtaining the second recorded video and its second score includes: The first recorded video of a single user performing gymnastic movements in different subcategories, along with the corresponding breakdown steps of each subcategory, is input into the first scoring model to obtain the user's first score for each subcategory of gymnastic movements. The first score of a single user for different subcategories of gymnastics movements is compared with a preset pass threshold. The gymnastics movements corresponding to the first score that is greater than the pass threshold are included in the user's pass set. If any gymnastics movement contained in the second knowledge graph has an intersection with the user's pass set, then the gymnastics movements of each subcategory in the second knowledge graph other than the intersection are taken as the user's candidate gymnastics movements. The user selects one gymnastics move from all available options as their self-selected gymnastics move. The recording unit then obtains a video of the user's self-selected gymnastics move, which is recorded as the second recorded video. The second score of the second recorded video is obtained using the same method as the first score of the first recorded video, and this score is then linked to the user's current set of qualified moves.
[0020] It should be further explained that, in the specific implementation process, the process of constructing the second scoring model based on the second scores of different self-selected gymnastics movements and the first scores of their corresponding sets of qualifying movements includes: A second score set is generated based on the second score of different users for their self-selected gymnastics movements and the first score of each gymnastics movement in the corresponding set of qualified movements, and then divided into a second training set and a second test set. Construct a second convolutional neural network by taking the first score of each gymnastics movement in the set of qualified movements of different users in the second training set and their self-selected gymnastics movements as the input data of the second convolutional neural network, and taking the second score of the corresponding self-selected gymnastics movements in the second training set as the output data of the second convolutional neural network. Train the second convolutional neural network using the second training set to obtain the initial second convolutional neural network. The initial second convolutional neural network is validated using a second test set. The initial second convolutional neural network whose output is less than or equal to a preset second test error threshold is used as the second scoring model. The second scoring model can evaluate the user's acceptance of other gymnastic movements based on the gymnastic movements the user has mastered.
[0021] It should be further explained that, in the specific implementation process, the process of using the second scoring model to combine the user's set of qualified movements with the first score to obtain the user's predicted scores for different gymnastic movements, and then obtaining recommended gymnastic movements, includes: In subsequent application scenarios, the first score of each gymnastics movement in a single user's current set of qualified movements and the other gymnastics movements are respectively input into the second scoring model to obtain the user's predicted scores for the other gymnastics movements. The predicted score is used to reflect the difficulty level of the user in learning other gymnastic movements. The higher the predicted score, the easier it is to learn, and the lower the predicted score, the more difficult it is to learn. The gymnastic movement corresponding to the highest predicted score is taken as the recommended gymnastic movement for the user and fed back to the user.
[0022] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A knowledge graph-based intelligent recommendation method for gymnastics teaching resources, characterized in that, Includes the following steps: Step S1: Collect teaching resources for different gymnastic movements, and construct a first knowledge graph for a single gymnastic movement and a second knowledge graph for different gymnastic movements. Step S2: Obtain the user's first recorded video and decompose it into frames to obtain multiple keyframes. Perform action scoring on each keyframe to obtain the first score of the first recorded video. Construct a first scoring model based on the first recorded video of different users and its corresponding keyframes and first scores. Step S3: Obtain the user's set of qualified movements based on the first score, obtain the user's self-selected gymnastics movements by combining the second knowledge graph, and obtain the second recorded video and its second score. Construct a second score model based on the second scores of different self-selected gymnastics movements and the first scores of the corresponding set of qualified movements. Step S4: Use the second scoring model to combine the user's set of qualified movements with the first score to obtain the user's predicted score for different gymnastic movements, and obtain recommended gymnastic movements; The process of constructing the first and second knowledge graphs includes: The teaching resources corresponding to each subcategory of gymnastics movements are converted into a unified format and semantics, and the relationships between various entities in the teaching resources are established. Ontology is used to define the various entities and their relationships in the teaching resources to obtain the semantic structure of the knowledge graph. Natural language processing techniques are used to extract key information from text data, reasoning is performed on the key information to fill the gaps in the knowledge graph, and the data structure of the knowledge graph is constructed by combining the obtained semantic structure, and the obtained key information is filled into the knowledge graph. The knowledge graph constructed above is used as the first knowledge graph of the corresponding sub-category of gymnastic movements. Knowledge graph fusion technology is used to merge the first knowledge graphs of various sub-categories belonging to the same major category into the second knowledge graph of the corresponding major category. The process of obtaining the user's set of qualified movements and their self-selected gymnastics movements, and obtaining the second recorded video and its second score, includes: The first recorded video of a single user performing gymnastic movements in different subcategories, along with the corresponding breakdown steps of each subcategory, is input into the first scoring model to obtain the user's first score for each subcategory of gymnastic movements. The first score of a single user for different subcategories of gymnastics movements is compared with a preset pass threshold to obtain the user's pass set. If any gymnastics movement contained in the second knowledge graph intersects with the user's pass set, then the gymnastics movements of each subcategory in the second knowledge graph other than the intersection are taken as the user's candidate gymnastics movements. The user selects one gymnastics move from all available options as their self-selected gymnastics move. The recording unit obtains a second recorded video of the user's self-selected gymnastics move, obtains a second score corresponding to the second recorded video, and binds it to the user's current set of qualified moves. The process of constructing the first rating model includes: Obtain the first recorded videos of gymnastics movements for different subcategories from different users, and obtain each keyframe and its decomposition step diagram from the first recorded videos of different subcategories. Combine the first scores of the first recorded videos to generate a first score set, and divide it into a first training set and a first test set. Construct a first convolutional neural network by using different first recorded videos and corresponding decomposition step diagrams of each subclass in the first training set as input data for the first convolutional neural network, and the first score and each keyframe in the first training set as output data for the first convolutional neural network. Train the first convolutional neural network using the first training set to obtain the initial first convolutional neural network. The initial first convolutional neural network is validated using the first test set, and the initial first convolutional neural network whose output is less than or equal to the preset first test error threshold is used as the first scoring model. The process of constructing the second scoring model includes: A second score set is generated based on the second score of different users for their self-selected gymnastics movements and the first score of each gymnastics movement in the corresponding set of qualified movements, and then divided into a second training set and a second test set. Construct a second convolutional neural network by taking the first score of each gymnastics movement in the set of qualified movements of different users in the second training set and their self-selected gymnastics movements as the input data of the second convolutional neural network, and taking the corresponding second score in the second training set as the output data of the second convolutional neural network. Train the second convolutional neural network using the second training set to obtain the initial second convolutional neural network. The initial second convolutional neural network is validated using the second test set, and the initial second convolutional neural network whose output is less than or equal to the preset second test error threshold is used as the second scoring model. The process of obtaining predicted ratings for different gymnastic movements using a second rating model and then obtaining recommended gymnastic movements includes: In subsequent application scenarios, the first score of each gymnastics movement in a single user's current set of qualified movements and the other gymnastics movements are respectively input into the second scoring model to obtain the user's predicted scores for the other gymnastics movements. The gymnastics movement corresponding to the highest predicted score is taken as the user's recommended gymnastics movement and fed back to the user. The process of obtaining the first recorded video and its first rating includes: Set up an input unit and acquire the first recorded video of a user performing a single sub-category of gymnastic movements. Perform frame decomposition processing on the first recorded video to obtain the corresponding video frames. The user selects the corresponding key frames from all the video frames of the first recorded video according to the decomposition steps of the single sub-category. Obtain the key frame interval between the timestamp of each key frame and the start time of the first recorded video. OpenPose was used to extract the coordinates of multiple human key points in a single keyframe and its step decomposition graph. Obtain the similarity between a single keyframe and its step decomposition diagram, and use the average similarity between each keyframe belonging to the same first recorded video and its corresponding step decomposition diagram as the first score of the gymnastics movement in the corresponding sub-category of the first recorded video.
2. The intelligent recommendation method for gymnastics teaching resources based on knowledge graphs according to claim 1, characterized in that, The process of collecting teaching resources for different gymnastic movements includes: Gymnastics movements are divided into different major categories, including tumbling, handstands, handsprings, somersaults, horizontal bar, parallel bars, and vault, and each major category contains different subcategories; Obtain teaching resources corresponding to different subcategories of gymnastics movements. The teaching resources include the movement names, text descriptions, teaching points, and corresponding breakdown diagrams for each step of a single subcategory.
Citation Information
Patent Citations
Fitness auxiliary teaching method and system
CN115205740A
Group teaching strategy generation method based on smart learning cloud platform
CN120355548A