Multi-center badminton action evaluation method based on trusted execution environment

Abstract

This invention relates to a multi-center badminton motion evaluation method based on a trusted execution environment, comprising the following steps: acquiring a motion segment to be evaluated; extracting motion feature vectors associated with the motion segment locally to obtain motion features to be evaluated; based on the motion features to be evaluated, performing collaborative retrieval with at least one other participating center to obtain a set of candidate motion labels; and based on the set of candidate motion labels and a preset motion evaluation model, obtaining the evaluation result of the motion segment to be evaluated. The technical solution provided by this invention introduces privacy computing and secure aggregation mechanisms without aggregating raw video data or standard / auxiliary databases from various centers, enabling cross-center collaborative retrieval and evaluation, and supporting federated iterative optimization of the global model. This allows for standardized evaluation and deviation localization adapted to different groups and training stages while ensuring the privacy of trainees and the security of database assets.

Description

Technical Field

[0001] This invention relates to the field of data security computing technology, and in particular to a multi-center badminton motion evaluation method based on a trusted execution environment. Background Technology

[0002] Badminton training involves various basic hitting techniques (such as forehand clear, smash, drop shot, lift, net drop, backhand push, etc.), and the standardization of these techniques directly affects hitting stability, power efficiency, and sports safety. Current training methods typically rely on the experience and observation of human coaches, using visual inspection or slow-motion video to correct the trainee's movements. With the development of computer vision and wearable sensor technology, methods such as video-based human pose estimation, target detection / trajectory tracking-based ball trajectory analysis, and sensor-based swing data acquisition have emerged to quantitatively evaluate and guide training movements.

[0003] Existing vision / sensor-based badminton training systems typically include modules for target detection, target tracking, and behavior recognition. Due to the complexity and diversity of human behavior, behavior recognition results are often output in probabilistic form. Some systems further construct motion fragment databases to enable similarity retrieval and comparative analysis between trainees' motion fragments and demonstrated movements in the database. However, the "standardization" of basic badminton movements is not a single, fixed standard. Different age groups, physical qualities and flexibility / strength conditions, training stages, and injury risks will result in differences in the range of motion, joint range of motion, force chain requirements, and training load for individuals. Furthermore, the same movement also presents different technical points and constraints between mass training focused on "sustainable and safe" and competitive training focused on "maximum speed / explosiveness." Therefore, using a single outstanding athlete's movement or a single template as a uniform standard can easily lead to inappropriate training recommendations or even induce injury risks. Simultaneously, trainee videos and sensor data often contain identifiable human characteristics, behavioral spatial information, and movement habit characteristics, constituting highly sensitive personal data; training videos of outstanding athletes or institutions, standard movement databases, and their characteristic representations may also be copyrighted or commercial assets. Existing systems generally lack systematic mechanisms for privacy and asset protection during data collection, uploading, storage, retrieval, comparison, and model optimization, posing risks of personal data leakage and the copying and reverse reconstruction of standard libraries. Summary of the Invention

[0004] Based on the above-mentioned situation of the prior art, the purpose of this embodiment of the invention is to provide a multi-center badminton action evaluation method based on a trusted execution environment, so as to solve the above-mentioned problems existing in the prior art.

[0005] To achieve the above objectives, according to one aspect of the present invention, a multi-center badminton motion evaluation method based on a trusted execution environment is provided, comprising the following steps:

[0006] Obtain the action segment to be evaluated, and extract the action feature vector associated with the action segment to be evaluated locally to obtain the action features to be evaluated;

[0007] Based on the features of the action to be evaluated, a collaborative retrieval with at least one other engagement center is performed to obtain a set of candidate action tags;

[0008] Based on the candidate action label set and the preset action evaluation model, the evaluation result of the action segment to be evaluated is obtained;

[0009] The collaborative retrieval is based on a trusted execution environment, where the similarity of the action features to be evaluated is calculated locally in other participating centers. The candidate action tag set includes candidate action tags and the confidence level corresponding to each candidate action tag. The candidate action tags include standard action tags and erroneous action tags.

[0010] Furthermore, the action feature vector associated with the action segment to be evaluated is extracted locally to obtain the action features to be evaluated, including the following steps:

[0011] Perform key point detection on the image within the human body frame to obtain the two-dimensional or three-dimensional coordinates of the key points of the human skeleton;

[0012] The two-dimensional or three-dimensional coordinates are standardized to obtain the normalized coordinates of the key points.

[0013] Based on the standardized keypoint normalized coordinates, a driving feature sequence is constructed as the action feature to be evaluated.

[0014] The human body frame is the smallest frame that includes the skeleton of the human body to be processed, obtained after processing the action segment to be evaluated.

[0015] Furthermore, the driving feature sequence includes joint angle features, key point velocity and acceleration features, key time point features, and relative position features of the hitting point.

[0016] Furthermore, the method also includes the step of: labeling the action segment to be evaluated with action tags corresponding to the features of the action to be evaluated;

[0017] Based on the characteristics of the action to be evaluated, a collaborative retrieval is performed with at least one other engagement center, including the following steps:

[0018] The action features to be evaluated and their corresponding action tags are encrypted and sent to other participating centers so that they can perform similarity calculations within their local TEEs to obtain a set of candidate action tags.

[0019] Furthermore, participating centers involved in the collaborative retrieval perform similarity calculations within their local TEEs, including the following steps:

[0020] Using the above action feature vector as a query vector, calculate the similarity between the query vector and the standard action feature vector and / or auxiliary action feature vector of the participating center.

[0021] Based on the obtained similarity, select the N most similar feature vectors from the standard action feature vector and / or auxiliary action feature vector;

[0022] The N feature vectors are grouped according to their respective action labels. The similarity is aggregated within each group, and the result of the aggregation calculation is used as the confidence level of the action label.

[0023] The confidence level represents the overall degree of matching between the action to be evaluated and the standard action and the erroneous action; the standard action feature vector is obtained from the standard action database located at the current participant center, the auxiliary action feature vector is obtained from the auxiliary action database located at the current participant center, and N is selected according to the scenario requirements.

[0024] Furthermore, the action evaluation model includes a model input processing layer, a temporal encoding layer, a comparison evaluation layer, and a scoring output layer.

[0025] Furthermore, the model input processing layer is used to receive and preprocess the input feature sequence, which includes action features to be evaluated, standard action features, and erroneous action features.

[0026] Furthermore, the temporal coding layer is used to extract temporal embedding representations that reflect the overall posture changes and force rhythm of the action from the input feature sequence input to the model input processing layer. The temporal embedding representations include the temporal embedding vector to be evaluated, the standard action temporal embedding vector, and the erroneous action temporal embedding vector.

[0027] Furthermore, the comparison evaluation layer is used to compare the timing embedding vector to be evaluated with the standard action timing embedding vector and the erroneous action timing embedding vector to obtain the difference vector and the cosine similarity scalar.

[0028] Furthermore, the scoring output layer is used to obtain the action standardization score, deviation type score, and risk level based on the time-series embedding vector to be evaluated, the standard reference embedding vector, the erroneous reference embedding vector, the difference vector, and the cosine similarity scalar.

[0029] In summary, this invention provides a multi-center badminton motion evaluation method based on a trusted execution environment. The method includes the following steps: acquiring a motion segment to be evaluated; extracting motion feature vectors associated with the motion segment locally to obtain motion features; performing collaborative retrieval with at least one other participating center based on the motion features to obtain a set of candidate motion labels; and obtaining the evaluation result of the motion segment based on the set of candidate motion labels and a preset motion evaluation model. The technical solution provided by this invention introduces privacy computing and secure aggregation mechanisms without aggregating raw video data or standard / auxiliary databases from various centers, enabling cross-center collaborative retrieval and evaluation, and supporting federated iterative optimization of the global model. This allows for standardized evaluation and deviation localization adapted to different groups and training stages while ensuring the privacy of trainees and the security of database assets. Attached Figure Description

[0030] Figure 1 This is a flowchart of a multi-center badminton motion evaluation method based on a trusted execution environment provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0032] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0033] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. An embodiment of the present invention provides a multi-center badminton motion evaluation method based on a Trusted Execution Environment (TEE). This method is implemented based on multi-center badminton motion evaluation. The system includes multiple participating centers, including but not limited to training institutions, schools, hospital rehabilitation centers, sports research units, and enterprise terminal nodes. Each participating center establishes and maintains a standard motion database and an auxiliary motion database locally. Data from each participating center's database is not uploaded or centrally aggregated; trainee data is stored in their respective participating center. Supported by a privacy-preserving computational mechanism, this evaluation system can achieve cross-center motion comparison, evaluation, and iterative optimization. Each participating center deploys data acquisition and evaluation nodes locally and maintains a standard motion database, an auxiliary motion database, and population rules and model parameters locally. Rules and parameters can be stored in encrypted memory or encrypted storage media within a Trusted Execution Environment (TEE) to avoid plaintext exposure. The content stored in each database is a motion feature representation rather than video source files.

[0034] The standard motion database stores feature vectors for standard motion segments, including skeleton keypoint sequences, joint angle and angular velocity sequences, key time points, etc., with labels for the target audience, training phase, and motion type. The auxiliary motion database stores feature vectors for common erroneous motions, including skeleton keypoint sequences, joint angle / angular velocity sequences, key time points, etc., and error type labels. Each participating center can also build vector indexes and other fast retrieval structures in its local database to improve the matching efficiency of local features during the retrieval process. Both the standard motion database and the auxiliary motion database of each participating center are stored locally, and the database content consists of skeleton keypoint sequences and their derived features (not video source files). During cross-center collaborative retrieval, only necessary candidate labels, encrypted intermediate values, and / or aggregate statistics are exchanged, which avoids the leakage and compliance risks associated with centralized storage from an architectural perspective.

[0035] Figure 1 The flowchart of the multi-center badminton motion evaluation method based on a trusted execution environment provided by an embodiment of the present invention is shown below. Figure 1 As shown, the method includes the following steps:

[0036] S202. Obtain the action segment to be evaluated, and extract the action feature vector associated with the action segment locally to obtain the action feature to be evaluated. Simultaneously, based on the action segment to be evaluated, label the action feature corresponding to it. For the same set of action features to be evaluated, one action label or multiple action labels can be labeled. This step is performed locally in the TEE of the participant center to which the participant belongs. Through edge-side feature generation and TEE, multi-center retrieval and comparison inference can be completed without exposing the original video and sensitive action feature plaintext of the trainee. This also reduces the risk of the standard / auxiliary feature library being copied, crawled, or reverse-engineered, achieving dual protection of trainee privacy and database assets.

[0037] S2021. Collect video footage of the movements to be evaluated, and segment the video footage into movement segments to obtain the movement segments to be evaluated. Within the participation center to which the person being evaluated (e.g., the trainee) belongs, capture video footage of the person's badminton movements using local data collection nodes, and segment the video footage into movement segments to obtain the movement segments to be evaluated. Movement segmentation can be performed based on a single rally or a single basic movement.

[0038] S2022. For each action segment, extract the sequence of key points on the human skeleton and calculate the action features to be evaluated. The calculation of the action features to be evaluated can be based on the following steps:

[0039] S2022a. Perform keypoint detection on the image within the human body bounding box to obtain the two-dimensional or three-dimensional coordinates of the keypoints of the human skeleton. The human body bounding box is the smallest bounding box containing the human skeleton to be processed, obtained after processing each captured action segment. Taking two-dimensional coordinates as an example, it can be represented as:

[0040] ;

[0041] Where m represents the keypoint index of the human skeleton, t represents the frame number, and x and y represent the horizontal and vertical coordinates of the keypoint, respectively. Keypoints include at least the left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, left and right ankles, and the center point of the torso. Optional keypoints may also include toes, heels, and related points of the hands. The representation of three-dimensional coordinates is similar to that of two-dimensional coordinates, with the addition of the vertical coordinate (z-axis coordinate) of the keypoints.

[0042] S2022b. Standardize the aforementioned two-dimensional or three-dimensional coordinates to obtain normalized coordinates for keypoints, thereby reducing the impact of camera distance and individual body shape differences. Still using two-dimensional coordinates as an example, let the coordinates of the m-th keypoint in frame t be... With the pelvis as the center, the root node is used. After translation and normalization, intermediate processing coordinates are obtained:

[0043] ;

[0044] Then use shoulder width as a scale factor After scaling and normalization, the normalized coordinates of the key points are obtained:

[0045] .

[0046] S2022c, Based on the above standardized keypoint normalized coordinates A driving feature sequence is constructed as the action features to be evaluated, serving as input to the action evaluation model. The driving feature sequence includes at least joint angle features, key point velocity and acceleration features, key time point features, and relative position features of the impact point. Among these, joint angle features include elbow joint angle, knee joint angle, shoulder joint arm-raising angle, and trunk rotation angle obtained using the angle between the lines connecting the left and right hips. The elbow joint angle can be expressed as:

[0047] ;

[0048] in, Represents the upper arm vector. The forearm vector can be calculated using the following formulas:

[0049] ;

[0050] ;

[0051] in, , and These are the normalized coordinates of the key points on the shoulder, elbow, and wrist in frame t.

[0052] The knee joint angle can be expressed as:

[0053] ;

[0054] in The thigh vector, The lower leg vector can be calculated using the following formulas:

[0055] ；

[0056] ;

[0057] in, , and These are the normalized coordinates of the key points of the hip, knee, and ankle in frame t.

[0058] The shoulder joint elevation angle can be expressed as:

[0059] ;

[0060] in The upper arm vector, The longitudinal reference vector of the torso can be calculated using the following formulas:

[0061] ；

[0062] ;

[0063] in, , These are the normalized coordinates of the key points at the elbow and shoulder in frame t, respectively. The normalized coordinates of the torso reference point in frame t can be taken as the midpoint of the left or right hip, the sternum, or the center point of the hip.

[0064] The trunk rotation angle is expressed as:

[0065] ;

[0066] in The shoulder line vector, The pelvic line vector can be calculated using the following formulas:

[0067] ；

[0068] ;

[0069] in , These are the normalized coordinates of the key points on the left and right shoulders of frame t, respectively. , The coordinates are the normalized coordinates of the left and right hip key points in frame t.

[0070] The velocity and acceleration features of key points are related to the frame rate. Taking wrist key points as an example, they can be calculated using the following formulas:

[0071] ;

[0072] ;

[0073] in, Indicates the speed at key points. This represents the acceleration at key points, and t-1 represents the frame number of the time preceding the t-th frame. Let the normalized coordinates of the wrist key points in frame t be , and the frame rate be fps. Similarly, the velocity and acceleration of key points such as the ankle can be calculated to characterize swing acceleration, start-up return to center, and footwork rhythm.

[0074] Key timing features include the start time of the backswing, the time of impact, and the end time of the follow-through. These can be calculated based on the speed characteristics of key wrist points. The start time of the backswing can be expressed as:

[0075] ;

[0076] The time to strike the ball can be expressed as:

[0077] ;

[0078] The follow-through end time is the latest moment when the wrist speed drops below the threshold and the posture stabilizes.

[0079] The relative position characteristics of the hitting point include horizontal forward / backward offset, vertical vertical offset, and spatial distance offset with the pelvic center as a reference point. Among them, the horizontal forward / backward offset can be expressed as:

[0080] ;

[0081] Vertical height offset can be expressed as:

[0082] ;

[0083] Spatial distance offset can be expressed as:

[0084] ;

[0085] in , These are the normalized coordinates of key points centered on the wrist and pelvis at the time of impact.

[0086] S204. Based on the aforementioned action features to be evaluated and the labeled action tags, a collaborative retrieval is performed with at least one other participating center to obtain a set of candidate action tags. The action features to be evaluated and their corresponding action tags are encrypted and sent to the other participating centers, allowing them to perform similarity calculations within their local TEEs to obtain the candidate action tag set. Conventional encryption methods in the art can be used to encrypt the action features to be evaluated. Upon receiving the action features to be evaluated, the participating centers decrypt them using a pre-defined decryption method. During the collaborative retrieval process, each participating center does not share its database with external entities. The action feature vector to be evaluated is retrieved across domains in encrypted form using privacy computing techniques. The current participating center only outputs candidate action tag information and its confidence level for subsequent evaluation.

[0087] The participating centers involved in collaborative retrieval can perform similarity calculations based on the action features to be evaluated within their local TEEs, using the following steps:

[0088] S2041. Using the above action feature vector as a query vector, calculate the similarity between the query vector and the standard action feature vector and auxiliary action feature vector of other participating centers. The standard action feature vector and auxiliary action feature vector are obtained from the standard action database and auxiliary action database located in the current participating center.

[0089] S2042. Based on the obtained similarity, select the N most similar feature vectors from the standard action feature vector and the auxiliary action feature vector. N can be selected according to the needs of the scenario. In this embodiment of the present invention, N is preferably selected as 30-80. Group the N feature vectors according to their respective action tags to form K groups, where K represents the number of candidate action tags contained in the candidate action tag set. Perform similarity aggregation calculation within each group and use the result of the aggregation calculation as the confidence level of the action tag. The confidence level represents the overall matching degree between the action to be evaluated and the standard actions in the standard action database and the erroneous actions in the auxiliary action database.

[0090] The set of candidate action labels can be represented as:

[0091] ;

[0092] in, This section displays labels for candidate shots, including standard and incorrect shots. Candidate shots include forehand high clear, smash, drop shot, lob, net drop, backhand push, etc. The confidence level corresponding to the candidate action label, and the set of candidate action labels. It includes both standard action labels and incorrect action labels to facilitate subsequent deviation localization.

[0093] The aforementioned collaborative retrieval is preferably implemented through a Trusted Execution Environment (TEE). Each participating center loads its local standard action feature library and auxiliary action feature library within a hardware-protected TEE. Similarity calculations for the action features to be evaluated, whether in encrypted form or transmitted in a secret state, are performed only within the TEE. External operating systems and other processes cannot access the plaintext features, intermediate calculation results, and database parameters within the TEE. Before initiating a cross-center collaborative retrieval, the participating center to which the trainer belongs remotely authenticates the target center's TEE, verifying that it is running an authenticated retrieval program. Then, an encrypted channel is established using a session key, and the action features to be evaluated are sent to the target participating center's TEE memory space. The target participating center performs similarity calculations and candidate tag selection within its TEE, outputting only candidate action tags and their corresponding confidence scores. Anonymized retrieval statistics can also be output. According to some optional embodiments, secure multi-party computation (MPC), homomorphic encryption (HE), and other technologies can be combined to further hide retrieval access patterns or feature distributions, enhancing privacy protection in sensitive scenarios.

[0094] S206. Based on the candidate action label set and the preset action evaluation model, the evaluation result of the action segment to be evaluated is obtained. The action evaluation model provided in this embodiment of the invention reflects the differences in age groups, physical fitness, and training stages. This model is constructed through joint training by multiple participants using a federated learning framework, and is a global model obtained through multi-center federated learning training. Based on the candidate action label set obtained in step S204 above... And N action features, the action features to be evaluated obtained in step S202, and the target population category Training phase Action type Using a motion evaluation model An evaluation is then conducted. In the above technical solution, candidate action labels are first obtained through multi-center collaborative retrieval, and then the action evaluation model is invoked to perform inference and scoring under the conditions of the candidate labels. This can effectively reduce misjudgments caused by action category confusion and scene differences, and improve the accuracy and robustness of action label determination and subsequent deviation localization. According to some optional embodiments, the action evaluation model is based on a neural network architecture, including a model input processing layer, a temporal encoding layer, a contrast evaluation layer, and a scoring output layer.

[0095] The model input processing layer is used to receive the action features to be evaluated and the candidate action label set mentioned above. and The motion features corresponding to each candidate motion label are given by the following: Let the motion feature vector of a single motion segment to be evaluated in frame t be... The motion feature vector is obtained by concatenating multiple numerical features such as joint angles, key point velocities, key point accelerations, and the relative position of the hitting point. The entire motion is represented as a feature sequence. The model input processing layer preprocesses three types of input: the action feature sequence to be evaluated. Standard action feature sequence set and the set of erroneous action feature sequences The sequences are arranged in chronological order. The model input processing layer performs zero-padding and masking on action feature sequences of different lengths to support variable-length sequence inputs.

[0096] The population category g, the training phase l, and the candidate action label set are used. Conditional information is encoded into conditional vectors, which are then input into the subsequent temporal coding layer along with the corresponding action feature sequence. The crowd category and training stage can be obtained through annotation. The temporal coding layer encodes the feature sequence input to the model input processing layer, including modeling and compression in the time dimension, and extracting temporal embedding representations reflecting the overall posture changes and force exertion rhythm of the action. The temporal coding layer includes a parameter-sharing temporal neural network used to process any input feature sequence. Frame-by-frame encoding is performed, and the feature vector of each frame is projected onto the hidden space through a linear mapping to obtain the hidden state sequence. The hidden state sequence is then pooled and attention-weighted converged to obtain the action-level temporal embedding vector z, which represents the features of the entire temporal action. The temporal coding layer shares the same inputs among the three types: the action feature sequence to be evaluated, the standard action feature sequence, and the erroneous action feature sequence. , , Encoding is performed to obtain the corresponding action-level temporal embedding vector: Temporal embedding vector to be evaluated Standard action timing embedding vector and erroneous action timing embedding vector This maps action sequences from different sources to a unified representation space, providing a computational basis for subsequent difference assessment.

[0097] The contrast evaluation layer compares the temporal embedding vector of the action to be evaluated with the temporal embedding vectors of the standard action and the erroneous action, constructing high-level features for action quality evaluation. The aggregation unit in the contrast evaluation layer aggregates the temporal embedding vectors of the standard action to obtain the standard reference embedding vector. The erroneous action timing embedding vectors are aggregated to obtain the erroneous reference embedding vector. The aggregation function Agg can be average pooling, weighted averaging, or attention-based weighted aggregation. This is based on the temporal embedding vector to be evaluated. Standard reference embedding vector and error reference embedding vector Generate a difference vector and a cosine similarity scalar, where the difference vector is represented as the standard deviation vector. and error difference vector :

[0098] ;

[0099] ;

[0100] The cosine similarity scalar is represented as the standard cosine similarity scalar. And the erroneous cosine similarity scalar :

[0101] ;

[0102] ;

[0103] The comparison evaluation layer concatenates the time-series embedding vector to be evaluated, the standard reference embedding vector, the erroneous reference embedding vector, the above-mentioned difference vector, and the cosine similarity scalar in a predetermined order to form a comparison feature vector h, which provides input for the subsequent scoring output layer.

[0104] The scoring output layer outputs action standardization scores, deviation type scores, and risk levels based on the contrast feature vector. The scoring output layer may include a multilayer perceptron, a residual fully connected module, a normalization layer, and a conditional multi-task learning module consisting of nonlinear activation and dropout; it is trained using labeled historical data. The scoring output layer is trained using labeled historical data, with the contrast feature vector h compared to the population category g, training stage l, and candidate action label set. After concatenation or fusion, the input vector is used, and the output includes action standardization score, deviation type score, and risk level. The output in the training data can be obtained through manual standardization during historical data collection. In use, the contrast feature vector h obtained from the contrast evaluation layer is compared with the population category g, training stage l, and candidate action label set. After concatenating or fusing the conditional vectors, they are input into the trained scoring output layer to obtain the action standardization score, deviation type score, and risk level. Among these, the action standardization score... This is used to represent the overall quality of the action, and is a deviation type score vector. Each dimension corresponds to a predefined type of motion deviation score. Predefined deviation types encompass common issues such as improper position (front / back, high / low, near / far) of the hitting point, insufficient torso rotation, abnormal backswing amplitude or direction, abnormal swing acceleration and rhythm, insufficient start and footwork preparation, insufficient positioning and stride coverage, poor landing cushioning, abnormal lower limb alignment, unstable center of gravity control, abnormal racket path or wrist control, poor upper and lower limb coordination rhythm, and significant decrease in motion continuity under fatigue; risk level. This is used to characterize the potential injury risk level of the movement being evaluated in key areas, such as the shoulder, elbow, waist, knee, and ankle. By using population categories and training stages as model conditions or employing population-specific calibration, different assessment scales and output thresholds are applied to the same basic movement under different age groups, fitness levels, and injury risk conditions. Furthermore, the risk level output achieves a "safety-first" error correction sequence, reducing the risk of sports injuries caused by inappropriate training recommendations.

[0105] In summary, this invention relates to a multi-center badminton motion evaluation method based on a trusted execution environment. The method includes the following steps: acquiring a motion segment to be evaluated; extracting motion feature vectors associated with the motion segment locally to obtain motion features; performing collaborative retrieval with at least one other participating center based on the motion features to obtain a candidate motion label set; and obtaining an evaluation result for the motion segment based on the candidate motion label set and a preset motion evaluation model. The technical solution provided by this invention introduces privacy computing and secure aggregation mechanisms without aggregating raw video data or standard / auxiliary databases from various centers, enabling cross-center collaborative retrieval and evaluation, and supporting federated iterative optimization of the global model. This allows for standardized evaluation and deviation localization adapted to different groups and training stages while ensuring the privacy of trainees and the security of database assets.

[0106] It should be understood that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of this invention, technical features of the above embodiments or different embodiments can also be combined, steps can be implemented in any order, and many other variations exist regarding different aspects of one or more embodiments of the invention as described above; for the sake of brevity, they are not provided in the details. The specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A multi-center badminton action evaluation method based on a trusted execution environment, characterized in that, Including the following steps: Obtain the action segment to be evaluated, and extract the action feature vector associated with the action segment to be evaluated locally to obtain the action features to be evaluated; Based on the action features to be evaluated, a collaborative retrieval is performed with at least one other participating center to obtain a set of candidate action tags, including: encrypting the action features to be evaluated and the corresponding action tags and sending them to other participating centers so that other participating centers can perform similarity calculations within their local TEE to obtain a set of candidate action tags; The participating centers involved in collaborative retrieval perform similarity calculations within their local TEEs, including: using the aforementioned action feature vectors as query vectors, calculating the similarity between the query vectors and the participating center's local standard action feature vectors and / or auxiliary action feature vectors; for the obtained similarity, selecting the N most similar feature vectors from the standard action feature vectors and / or auxiliary action feature vectors; grouping the N feature vectors according to their respective action tags, performing aggregation calculations on the similarity within each group, and using the result of the aggregation calculation as the confidence level of the action tag; wherein, the confidence level represents the overall matching degree between the action to be evaluated and the standard action and the erroneous action; the standard action feature vectors are obtained from the standard action database located in the current participating center, the auxiliary action feature vectors are obtained from the auxiliary action database located in the current participating center, and N is selected according to the scenario requirements; Based on the candidate action label set and the preset action evaluation model, the evaluation result of the action segment to be evaluated is obtained; The collaborative retrieval is based on a trusted execution environment, where the similarity of the action features to be evaluated is calculated locally in other participating centers. The candidate action tag set includes candidate action tags and the confidence level corresponding to each candidate action tag. The candidate action tags include standard action tags and erroneous action tags.

2. The method of claim 1, wherein, Extracting the action feature vector associated with the action segment to be evaluated locally to obtain the action features to be evaluated includes the following steps: Perform key point detection on the image within the human body frame to obtain the two-dimensional or three-dimensional coordinates of the key points of the human skeleton; The two-dimensional or three-dimensional coordinates are standardized to obtain the normalized coordinates of the key points. Based on the standardized keypoint normalized coordinates, a driving feature sequence is constructed as the action feature to be evaluated. The human body frame is the smallest frame that includes the skeleton of the human body to be processed, obtained after processing the action segment to be evaluated.

3. The method of claim 2, wherein, The driving feature sequence includes joint angle features, key point velocity and acceleration features, key time point features, and relative position features of the hitting point.

4. The method of claim 1, wherein, The method further includes the step of: labeling the action segment to be evaluated with action tags corresponding to the features of the action to be evaluated.

5. The method of claim 1, wherein, The action evaluation model includes a model input processing layer, a temporal encoding layer, a comparison evaluation layer, and a scoring output layer.

6. The method according to claim 5, characterized in that, The model input processing layer is used to receive and preprocess the input feature sequence, which includes action features to be evaluated, standard action features, and erroneous action features.

7. The method of claim 6, wherein, The temporal coding layer is used to extract temporal embedding representations that reflect the overall posture changes and force rhythm of the action from the input feature sequence input to the model input processing layer. The temporal embedding representations include the temporal embedding vector to be evaluated, the standard action temporal embedding vector, and the erroneous action temporal embedding vector.

8. The method of claim 7, wherein, The comparison evaluation layer is used to compare the timing embedding vector to be evaluated with the standard action timing embedding vector and the erroneous action timing embedding vector to obtain the difference vector and the cosine similarity scalar.

9. The method of claim 8, wherein, The scoring output layer is used to obtain the action standardization score, deviation type score and risk level based on the time-series embedding vector to be evaluated, the standard reference embedding vector, the erroneous reference embedding vector, the difference vector and the cosine similarity scalar.

Images