Badminton racket holding action intelligent recognition and scoring method based on cross-feature interaction

Abstract

The application discloses a badminton grip action intelligent recognition and scoring method based on cross-feature interaction, continuous frame grip data collection is collected through an intelligent sensor, and preprocessing is carried out; after preprocessing, the FT-Transformer module of cross-feature interaction is input, the correlation between feature points is automatically learned, and single-frame cross-feature space features are extracted; then, through the lightweight multiple graph TCN module, the hidden correlation of the manual timing logic and the automatic learning is combined to capture the timing trend; finally, the spatial and timing double features are spliced by the double-task output module, 10 types of action classification and action standard degree scoring are realized, and the scene requirements of badminton individual training, professional teaching and the like can be effectively met.

Description

Technical Field

[0001] This invention belongs to the field of action recognition technology, specifically a method for intelligent recognition and scoring of badminton racket grip actions based on cross-feature interaction. Background Technology

[0002] Badminton, a globally popular high-speed ball sport, has seen a continuous increase in participants thanks to the promotion of national fitness policies. The grip is the core foundation of badminton technique, and its standardization directly affects power generation efficiency and the risk of injury, making it a crucial element for both professional training and amateur improvement. However, the industry currently faces a significant supply-demand imbalance. Professional coaching resources are limited, and most amateur enthusiasts lack systematic guidance. While the smart sports equipment market is growing rapidly, badminton smart equipment struggles to meet user needs in terms of motion recognition and the accuracy of training guidance.

[0003] Invention CN 114272584 A discloses a racket grip posture correction glove that identifies grip posture using pressure sensors, but it only identifies grip posture statically. Research in sports biomechanics shows that professional athletes' grip movements contain finely coordinated feature patterns and strict temporal characteristics. The lack of these micro-features is a major cause of technical bottlenecks for amateur players. Although smart sensors and edge computing technologies have been applied to smart badminton racket products, the accuracy of motion recognition and training guidance still needs improvement.

[0004] Existing technologies for badminton grip motion recognition and scoring have several shortcomings. Technical bottlenecks exist in areas such as motion recognition accuracy, scoring logic construction, and embedded device adaptation. Furthermore, existing models do not adapt to the "motion characteristics" of badminton grip feature data, making it difficult to meet the high-precision, low-latency, and interpretable requirements of badminton scenarios. Therefore, there is an urgent need to research and improve methods to enhance the performance of intelligent badminton grip motion recognition and scoring methods based on cross-feature interaction. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses a method for intelligent recognition and scoring of badminton racket grip movements based on cross-feature interaction.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A method for intelligent recognition and scoring of badminton racket grip actions based on cross-feature interaction includes the following steps:

[0008] Step 1: Arrange an array of pressure sensors on the racket handle. Define the pressure signals sensed by all pressure sensors at the same time as a frame of data. The pressure data of a single pressure sensor corresponds to a feature point.

[0009] Step 2: Collect continuous frame grip feature data at a preset frequency;

[0010] Step 3: Perform 3-frame moving average denoising, normalization, and invalid frame processing on the continuous frame grip feature data to obtain the preprocessed continuous frame grip feature data and construct the training dataset.

[0011] Step 4: Construct a grip action recognition and scoring model; the total loss function of the grip action recognition and scoring model. for:

[0012] ;

[0013] in, Cross-entropy loss for action classification, The mean squared error loss of the score, All weights for the grip action recognition and scoring model Regularization; where They are respectively and The weights;

[0014] Step 5: Train the action recognition and scoring model using the training dataset, with the total loss function... The model parameters are optimized with the goal of minimizing the loss until the loss converges and stabilizes, thus obtaining the trained grip action recognition and scoring model.

[0015] Step Six: Action Classification and Scoring Output. The model takes each frame of pressure data generated by the trainee holding the racket handle, preprocessed in Step Three, and inputs it into the trained model. The model outputs the corresponding racket grip action classification results and standard score.

[0016] Further improvements include a grip action recognition and scoring model comprising a cross-feature interaction FT-Transformer module, a lightweight multi-graph TCN module, and a dual-task output module;

[0017] The cross-feature interaction FT-Transformer module extracts cross features from single-frame grip feature data and outputs a single-frame cross-feature space feature vector.

[0018] The lightweight multi-graph TCN module performs temporal modeling on cross-feature space features across 10 consecutive frames to construct hand-crafted and automatically learned graphs. The hand-crafted and automatically learned graphs extract features through independent graph convolution paths, and feature fusion is achieved by element-wise addition. The fused features are input into two layers of causal convolution, and after layer normalization, residual connections, and global average pooling, the temporal feature vector is finally obtained. The causal convolution kernel has 3 kernels, and the dilation rates are 1 and 2, respectively.

[0019] The global average pooling method is as follows:

[0020] The fused features are then processed through causal convolution, followed by layer normalization and residual connection for output. Then, global average pooling is used to aggregate the mean across the time series dimensions, using the following formula:

[0021] ;

[0022] Represents time series feature vectors

[0023] Indicates the first The first batch, the first Frame, First The characteristic values ​​of each channel;

[0024] The single-frame cross-feature space feature vector and temporal feature vector are input into the dual-task output module, which outputs an action classification and a standard score. The standard score is calculated as follows:

[0025] ;

[0026] Total score = Spatial score + Temporal score;

[0027] Indicates spatial fraction; Indicates the spatial weight of the current action; Represents time series fractions; This indicates the temporal weight of the current action.

[0028] Further improvements include the following data processing procedure for the cross-feature interaction FT-Transformer module:

[0029] S11. Initialization of feature point embedding vectors:

[0030] The array-type pressure sensor is a 4×10 matrix. To capture the relative position information of the sensors in two-dimensional space, the pressure value of each pressure sensor is first obtained. :

[0031] ;

[0032] in The first in the matrix line, number Column elements, Number the sensor. The calculation formula is: ;

[0033] A spatial location embedding layer is introduced at the input stage. This layer is used to handle coordinate normalization, location encoding, feature concatenation and mapping, and spatial offset coefficients. for:

[0034] ;

[0035] in For action For sensors The degree of importance;

[0036] The method to obtain it is as follows:

[0037] The pressure values ​​of each pressure sensor were collected from professional athletes during various movements, and then the following calculations were performed:

[0038] ;

[0039] in Indicates action The set of indices for the downforce sensors; This represents the variance of the pressure values. Indicates the first The maximum variance of observations from all pressure sensors under the same action type. Indicates the importance of features. Indicates the first The greatest importance of all pressure sensors under similar actions; Indicates the first Type of action, first The sample, the first The first frame Pressure values ​​from each pressure sensor; Indicates the number of samples. Indicates the total number of frames; Indicates the first The first type of action The average pressure value of the entire sample and the entire frame of each sensor; Indicates the first The pressure sensor corresponds to the first Weight coefficients for action types;

[0040] Then, the motion is calculated based on the spatial offset coefficient and the coordinate normalization value. Below, the row index is The column index is Pressure sensor position coding The location coding formula is:

[0041] ;

[0042] in, Indicates row index Normalization; Indicates column index Normalized value ;

[0043] The pressure value, coordinate normalization, and position code of the pressure sensor are used for feature concatenation; the feature concatenation formula is as follows:

[0044] ;

[0045] in, This represents the normalized value of the sensor. Indicates the spatial offset coefficient. Indicates the first Pressure values ​​from each sensor;

[0046] S12: To improve the generalization ability of the grip action recognition and scoring model, a linear layer is used. The formula for the linear layer is:

[0047] ;

[0048] in, It is a learnable linear mapping matrix. For the corresponding bias term, This is an element-wise nonlinear activation function, used to introduce nonlinear expressive power; Indicates nonlinear results; Indicates matrix transpose;

[0049] Action after the linear layer The overall embedding matrix below :

[0050] ;

[0051] in, Indicates the current action Below is a high-dimensional set of embedded features for all pressure sensor inputs;

[0052] S13: Acquire Action Each pressure sensor Contribution Finally, the 12 pressure sensors with the highest contribution were designated as the primary force-generating areas, and the remaining pressure sensors as secondary force-generating areas. Focusing on the core force signals corresponding to the primary force-generating areas of the grip, the main force-generating areas of the sensors are:

[0053] ;

[0054] in Indicates action The index set of the main force sensors;

[0055] Subsequently, cosine similarity and weighted similarity between features were added to strengthen the model features.

[0056] The cosine similarity is as follows:

[0057] ;

[0058] in, Number the pressure sensor. They represent the first The overall embedding matrix of the sensors; Indicates the first Cosine similarity of the overall embedding matrix of each sensor;

[0059] Different force application zones are assigned different weight parameters, and the dynamic intensity of any two pressure sensors is given for different actions. :

[0060] ;

[0061] Indicates action Downward pressure sensor The degree of contribution; Indicates "and"; Indicates "or";

[0062] Combining feature similarity with dynamic weights, the weighted similarity is defined as:

[0063] ;

[0064] in Representation in the current action Below, pressure sensor and The overall interaction strength between them;

[0065] Finally, based on the force application zone to which the pressure sensor belongs, different association weights are assigned, and the weighted similarity is used as the message passing coefficient to perform neighborhood fusion on the features of each sensor to obtain aggregated features:

[0066] ;

[0067] in Indicates the first Aggregation characteristics of a pressure sensor; Indicates the first An integrated matrix of pressure sensors Indicates the first The overall embedded matrix of pressure sensors other than the individual pressure sensors Indicates the first Pressure sensors other than the individual pressure sensor;

[0068] S14: Set up 8 multi-head attention modules and initialize them using a fully connected layer. The matrix is ​​scaled using a dot product to eliminate the influence of matrix dimension on aggregated features. The scaling dot product formula is:

[0069] ;

[0070] in, For feature dimensions; These represent the query vector matrix, the key vector matrix, and the value vector matrix, respectively. Represents the normalized exponential function, This indicates a multi-head attention module;

[0071] Introducing a feedforward neural network module to aggregate features Perform sensor-by-sensor augmentation mapping:

[0072] ;

[0073] in, Indicates the first Aggregated characteristics of individual sensors , It is a learnable linear transformation matrix. For the corresponding bias term; Represents a non-linear activation function;

[0074] Obtain the output of the feedforward neural network module :

[0075] ;

[0076] This represents the feedforward neural network module. for A set; The output of the feedforward neural network module represents the first... The pressure sensor is at the first eigenvalues ​​of dimension;

[0077] Extract the global correlation features of all sensors within a single frame, and perform global average pooling on the enhanced feature output to output a single-frame cross-feature space feature vector. :

[0078] .

[0079] Further improvements are made to the methods for constructing manual and automatically learned graphs as follows:

[0080] Constructing a manual adjacency matrix :

[0081] ;

[0082] , For batch indexing, and All are frame indexes; Indicates batch size, Indicates the number of frames in the timing window;

[0083] Generate handmade images :

[0084] ;

[0085] Indicates the first The first batch, the first Feature vectors of all feature channels of the frame Map the 10-dimensional main power generation area features to dimension, The dimension index representing the output features of the hand-drawn graph; The input dimension represents the features of the main power generation area. Indicates the dimension of the output feature;

[0086] The automatic learning graph construction method is as follows:

[0087] Generate automatic adjacency matrix :

[0088] ;

[0089] Indicates the first The first batch Frame and the The characteristic Euclidean distance of the frame, For Euclidean distance, For the first The first batch, the first Frame, First eigenvalues ​​of dimension For the first The first batch, the first Frame, First eigenvalues ​​of dimension;

[0090] Generate automatic learning graph :

[0091] ; and Sharing core parameters, the 64-dimensional complete feature is mapped to 32 dimensions, and the output is... , Indicates the first The first batch, the first The 64-dimensional complete feature vector of the frame.

[0092] In a further improvement, in step one, the array-type pressure sensors are arranged in a 4×10 matrix on the racket grip, covering the index-middle finger force area, the thumb auxiliary area, the ring-little finger support area, etc. After conditioning the original signal, digital-to-analog conversion is performed, and the obtained digital voltage value corresponds to the real-time pressure of a single sensor unit; in step two, the preset frequency is 50Hz.

[0093] In a further improvement, the preprocessing method in step three is as follows: First, perform 3-frame moving average noise reduction:

[0094] ;

[0095] in, For the first Frame number Line number Column feature data, for Feature data after smoothing and denoising;

[0096] Then, Min-Max normalization is performed to shrink the feature data to the range [0,1].

[0097] ;

[0098] in, The minimum value of the feature data in the entire dataset. The maximum value of the feature data in the entire dataset. For the normalized first Line number Column characteristic data;

[0099] Secondly, for single-frame data, abnormal frames with pressure values ​​outside the range of [0, 100kPa] are removed. If the number of missing feature points is greater than 3, the frame is discarded directly; if the number of missing feature points is ≤ 3, the frame is filled with the average of the corresponding positions of the adjacent valid frames before and after it.

[0100] ;

[0101] in Indicates the first Frame, First Line number The filled characteristic values ​​of the column sensor.

[0102] Further improvements include exporting the trained grip action recognition and scoring model into a format supported by TensorRT, using a calibration quantization mode, completing quantization through a toolchain, and verifying the effect based on embedded deployment metrics after quantization.

[0103] In a further improvement, the dual-task output module constructs a single fully connected layer to horizontally concatenate cross-feature space feature vectors and temporal feature vectors to obtain a fused feature vector. The 1×96 dimensional fused feature vector is mapped to a 1×10 dimensional output. :

[0104] ;

[0105] in , These are learnable parameters;

[0106] Then, Softmax activation is performed on the output of the fully connected layer, transforming it into a probability distribution for each action category. :

[0107] ;

[0108] in The index corresponding to the maximum value is the predicted action category;

[0109] Next, set the confidence threshold. The final action category is determined based on the probability distribution:

[0110] ;

[0111] in This represents the maximum predicted probability among the 10 action categories. This represents the confidence threshold; the model only outputs a class when it has at least 50% confidence. The index represents the category with the highest probability. The unknown action indicates the output result when the model's confidence in all categories is below the threshold. express The maximum value.

[0112] Further improvements involve first defining spatial correlation strength and temporal fluctuation coefficients, and then using cosine similarity to calculate spatial components.

[0113]

[0114] in The 96-dimensional fusion feature template for standard movements was obtained through training with data from professional athletes. This is the 96-dimensional fusion feature of the current action, i.e., a fusion vector of spatial and temporal features. Cosine similarity;

[0115] The time series fraction calculation uses the reciprocal of the deviation from the standard time series template:

[0116] ;

[0117] in The 32-dimensional temporal features of standard actions, The temporal characteristics of the current action, Indicates the absolute deviation of the time series characteristics;

[0118] Then, the spatial and temporal scores are dynamically weighted to obtain the spatial and temporal weights, and finally, the total score is calculated.

[0119] ;

[0120] in Spatial weights, For time series weights.

[0121] A system for running the above-described intelligent recognition and scoring method for badminton grip movements based on cross-feature interaction.

[0122] The advantages of this invention are as follows:

[0123] The present invention has a grip action recognition and classification accuracy of ≥95%, high recognition precision, scores based on action sequence, deviation from professional coach scores of ≤5 points, clear and accurate scoring logic, and fast response speed, which can effectively meet the needs of badminton individual training, professional teaching and other scenarios. Attached Figure Description

[0124] Figure 1 This is a single frame of data collected by the pressure sensor array.

[0125] Figure 2 Flowchart for a lightweight multi-graph TCN architecture.

[0126] Figure 3 This is a flowchart of the data processing for the FT-Transformer module, which interacts across features. Detailed Implementation

[0127] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0128] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0129] This invention provides a badminton grip action recognition and scoring system based on cross-feature interaction FT-Transformer and lightweight TCN, including the following steps:

[0130] 1. Raw data collection

[0131] Equipment: It adopts a piezoelectric pressure sensor, which is arranged in 4 rows × 10 columns to fit the racket grip, covering key areas such as the thumb auxiliary area, the index-middle finger force area, and the ring-little finger support area; the piezoelectric pressure sensor adopts the Freescale MMA8451Q piezoelectric sensor, with an accuracy of 0.1kPa and a response time of ≤1ms. The sensor array is attached to the inside of the racket grip with waterproof adhesive to ensure no displacement and no data loss during the acquisition process.

[0132] Parameters: Sampling frequency 50Hz (1 frame is collected every 20ms), 10 frames are collected for a single action (corresponding to a 0.2-second force exertion process);

[0133] Action coverage: Ten types of hitting actions were collected (as shown in Table 1), and 20 sets of data were collected from each of the 60 subjects (including professional athletes, amateur enthusiasts, and beginners).

[0134] Table 1 Ten Categories of Shooting Actions

[0135]

[0136] 2. Data Cleaning

[0137] Invalid frame rejection: If the number of missing sensors in a single frame is greater than 3 (due to poor sensor contact, etc.), the frame is discarded directly;

[0138] Outlier filtering: Remove abnormal frames with pressure values ​​outside the range of [0, 100 kPa] (caused by sensor noise or accidental touch).

[0139] Missing value imputation: If the number of missing sensors in a single frame is ≤3, the average pressure value of the corresponding sensors in adjacent valid frames is used for imputation, using the formula:

[0140] in Indicates the first Frame, First Line number The filled characteristic values ​​of the column sensor;

[0141] 3. Data Preprocessing

[0142] 3-frame moving average denoising: For the temporal stress data of each sensor, the noise is smoothed using the average of the current frame and the frames before and after it. The formula is as follows:

[0143] in, Indicates the first Frame, First Line number Characteristic values ​​of the column sensor;

[0144] Min-Max Normalization: Scales the pressure value to the [0,1] interval to eliminate individual differences in force intensity. The formula is:

[0145]

[0146] 4. Feature Alignment and Action Labeling

[0147] Sensor coordinate mapping: Flatten the 4×10 two-dimensional pressure matrix into a 40-dimensional one-dimensional eigenvector, indexing... Compared with the original coordinates The correspondence is ( For line numbers, (for column number);

[0148] Time-series slicing: Extract 10 frames of data containing the complete power generation process from the continuously acquired data stream and form a 10×40 time-series feature matrix;

[0149] Action classification labeling: Label each 10×40 time series data with the unique hot codes of 10 types of hitting actions (e.g., the unique hot code for a forehand high clear is [0,1,0,...]).

[0150] Standardization rating label: Three national-level badminton coaches score the standardization of force application for each set of movements (score range 0-100 points), and the average of the three scores is taken as the standardization rating label.

[0151] 4.1. Core Indicator: Sensor Contribution Calculation

[0152] For each type of action The overall contribution of 40 sensors is calculated to reflect the criticality of the sensors in this action.

[0153] 4.2. Method 1: Pressure variance analysis (reflecting the intensity of force fluctuations)

[0154] The sensor pressure value in the force application zone fluctuates drastically with the action cycle, while the fluctuation is gradual in the non-force application zone, as shown in the following formula:

[0155]

[0156] Indicates the first Type of action, first One sample ( =1200, 60 people × 20 groups), the frame( ), No. Pressure values ​​from each sensor (unit: kPa); Indicates the first The first type of action The average pressure value of the entire sample and the entire frame of each sensor; This indicates the pressure variance (the larger the value, the more drastic the fluctuation, and the more likely it is to be a stress area).

[0157] Method 2: Feature importance score (reflecting the correlation with action labels)

[0158] The pressure-action classification model is trained using sensors. The absolute values ​​of the feature coefficients output by the model represent the importance of the sensor for action classification, as shown in the following formula:

[0159]

[0160] In the model, the first The sensor corresponds to the first Weight coefficients for action types; Feature importance (the larger the value → the better the sensor pressure value can distinguish this action from other actions → the more likely it is to be the force application zone)

[0161] Method 3: Overall Contribution (integrating two indicators to avoid bias from a single indicator)

[0162] right and After normalization, the mean value is taken to obtain the overall contribution (range [0,1]):

[0163]

[0164] Indicates the first The maximum variance of all sensors under the same action type; Indicates the first The greatest importance of all sensors under similar actions; This indicates the overall contribution (the higher the value, the more critical the sensor is in this action).

[0165] 4.3. Grading of Intensity Zones (Automatic K-means Clustering)

[0166] For each type of action, 40 sensors, according to natural breakpoints ( Automatically divides the area into two regions, eliminating the need for manual thresholding:

[0167] Key areas of focus: (The top 30% of sensors by contribution, the core focus);

[0168] Secondary focus areas: (Sensors contribute 70% of the power, assisting in the exertion of force)

[0169] 5. Dataset partitioning

[0170] The data was divided into a training set (70%), a validation set (20%), and a test set (10%) in a ratio of 7:2:1 to ensure a balanced distribution of the 10 types of hitting actions across the three datasets.

[0171] 6. Construct a cross-feature interaction FT-Transformer module to extract cross-feature space features from a single frame:

[0172] The specific implementation includes four sub-steps:

[0173] (1) Initialization of feature point embedding vector: The array pressure sensor is a 4×10 matrix. In order to capture the relative position information of the sensor in two-dimensional space, the position code is calculated by spatial offset coefficient, etc., and a spatial position embedding layer is introduced in the input stage. The spatial position embedding layer mainly processes coordinate normalization, position encoding and feature splicing and mapping. Then, the position code is calculated according to the spatial offset coefficient and coordinate normalization value, and the sensor's observation value, coordinate normalization and position code are spliced ​​together.

[0174] (2) To improve the generalization ability of the model, a linear layer is used. A linear layer can both increase the data dimensionality and enhance non-linear expressive power; after passing through a linear layer, the action can be obtained. The overall embedding matrix below ;

[0175] (3) Sensor overall contribution through K-means clustering The distance metric is Euclidean distance, and the top 30% of sensors in terms of contribution are selected as the main power generation area. The remaining sensors represent secondary force application areas and other regions, focusing on the core force application signals of gripping the racket. Subsequently, cosine similarity and weighted similarity between features were added to strengthen the model's features. Finally, different association weights were assigned based on the force application area to which the sensor belongs, and the weighted similarity was used as a message passing coefficient to perform neighborhood fusion on the features of each sensor.

[0176] (4) Feature extraction from the FT-Transformer encoder: Construct a 2-layer Transformer encoder, each layer containing:

[0177] Multi-head attention mechanism: Fixed to 8 heads, responsible for calculating the global collaborative relationships of high-dimensional embedded features from 40 sensors. The core computation is initialized through a fully connected layer. (Query vector) (key vector) (Value vector), attention weights are calculated using scaled dot product; K-means clustering is used to filter the associations in the main focus area.

[0178] Feedforward Network (FFN): It adopts a three-stage approach of dimensionality increase → activation → dimensionality reduction. The input and output dimensions are both 64, and the intermediate dimension is increased to 128, that is, 64-dimensional → 128-dimensional → 64-dimensional. The computation logic adopts independent nonlinear transformation of the associated features on a sensor-by-sensor basis.

[0179] Layer Normalization (LayerNorm): By standardizing the feature distribution, it eliminates the difference in feature magnitude between the main and secondary regions, normalizes the features of attention output and FFN output to a range with a mean of approximately 0 and a variance of approximately 1, ensures the stability of feature propagation in each layer, and alleviates the gradient vanishing problem in deep training.

[0180] Complete structure and flow of FT-Transformer with cross-feature interaction:

[0181] Step 1: Input layer (derived from preprocessed sensor data)

[0182] Input data: 4×10 sensor single-frame pressure matrix (physical layout: rows 0-3, columns 0-9).

[0183] Format: [1, 4, 10] (single sample, 4 rows, 10 columns).

[0184] Data meaning: Each element Represents "the Line number The normalized pressure value of the sensor (range [0,1]).

[0185] Step 2: Sensor Index Mapping

[0186] Operation: Convert the 2D matrix into a 1D index (to facilitate subsequent embedding calculations).

[0187] Mapping rule: The Line number Column → Index .

[0188] Output: 1D index sequence of 40 sensors .

[0189] Step 3: Spatial Location Embedding Layer (corresponding to Improvement Point 1)

[0190] Core operation: Add physical location coding to each sensor to enhance spatial correlation modeling capabilities.

[0191] Sub-step 3.1: Single sensor feature stitching

[0192] Input: Pressure value: ;coordinate Offset coefficient

[0193] Empty (from overall contribution:) ).

[0194] After splicing: , ,

[0195] Sub-step 3.2: Extending the Dimension of a Linear Map

[0196] Output embedding matrix It contains motion space information, and the embedded features of the force zone sensor have higher distinguishability, providing a precise spatial basis for the subsequent AFC layer.

[0197] Step 4: Core operation of the Approximate Feature Interaction (AFC) layer: Only calculate the interaction of strongly correlated feature pairs to reduce complexity (O(n²) → O(n)).

[0198] Sub-step 4.1: Calculate cosine similarity (formula as above)

[0199] formula: .

[0200] Meaning: To measure the first The and the first Orientational consistency of sensor features (range [-1, 1]).

[0201] Sub-step 4.2: Filtering strongly correlated pairs (dynamic correlation weights) (Weighted)

[0202] formula: .

[0203] Sub-step 4.3: Aggregate strongly correlated features

[0204] Meaning: When aggregating, prioritize retaining the strong correlation features of the main force area, and set the correlation degree of irrelevant areas <0.1 directly to 0.

[0205] Output: Sparse Incidence Matrix .

[0206] By preserving strong correlations in the power zone and reducing noise, the low similarity analysis in the subsequent scoring module becomes more accurate.

[0207] Sub-step 4.4: Feedforward Network (FFN) Enhances Nonlinearity

[0208] operate: .

[0209] 7. Construct a lightweight multi-graph TCN module to extract temporal features from consecutive frames:

[0210] The specific implementation includes three sub-steps, as follows: Figure 2 As shown:

[0211] (1) Construction of multigraph structure:

[0212] Hand-drawn diagram: An adjacency matrix is ​​constructed based on the biomechanical priors of badminton racket grip force, and the matrix calculation relies on the correlation strength between frames;

[0213] Automatic learning graph: By calculating the Euclidean distance between frames and taking the minimum value of 30% of the distances of all frame pairs within a batch as a threshold, strongly correlated frame pairs are selected to construct an adjacency matrix;

[0214] (2) Lightweight causal convolution and feature compression: Two depthwise separable causal convolutions are fused to the feature input. The first layer has a kernel of 3 and a dilation rate of 1 to capture short-range temporal correlations between two adjacent frames. The second layer has a kernel of 3 and a dilation rate of 2 to capture long-range temporal correlations between frames. After each convolution, layer normalization and residual connections are added to avoid gradient vanishing. Finally, the feature vector is compressed into a temporal feature vector by global average pooling.

[0215] (3) Temporal feature compression: The 10×32-dimensional temporal feature matrix output by TCN is compressed into a 1×32-dimensional temporal feature vector by global average pooling. The TCN module uses INT8 quantization preprocessing, which reduces the memory usage of parameters and is compatible with ARM Cortex-M4 chips to meet the requirements of real-time training guidance.

[0216] Specifically, the steps include the following:

[0217] Step 1: Dual-graph input and adjacency matrix construction

[0218] Input: Manual drawing path: Automatic graph path:

[0219] Adjacency matrix construction:

[0220] Manual graph adjacency matrix

[0221]

[0222] Automatic graph adjacency matrix

[0223] Calculate the Euclidean distance between frames:

[0224]

[0225] Dynamic threshold filtering for strong correlations:

[0226]

[0227] Step 2: Shared-weight graph convolution operation

[0228] Manually drawn path diagram (aggregating "prior associations of the main force areas"):

[0229]

[0230] ( Map the 10-dimensional main power generation area features to Dimension; Output )

[0231] express().

[0232] Automatic graph path (aggregating "hidden associations driven by full-area data"):

[0233]

[0234] ( and Sharing core parameters, the 64-dimensional complete feature is mapped to 32 dimensions, and the output is... )

[0235] The hand-drawn graph and the automatically learned graph each extract features through independent graph convolution paths, and feature fusion is completed by element-wise addition. The fused features are input into two layers of causal convolution, and after layer normalization, residual connection and global average pooling, the temporal feature vector is finally obtained. The convolution kernel of the causal convolution is 3, and the dilation rate is 1 and 2 respectively.

[0236] The global average pooling method is as follows:

[0237] The fused features are then processed through causal convolution, followed by layer normalization and residual connection for output. Then, global average pooling is used to aggregate the mean across the time series dimensions, using the following formula:

[0238]

[0239] Represents time series feature vectors

[0240] Indicates the first The first batch, the first Frame, First The characteristic values ​​of each channel.

[0241] Step 3: Dual-path feature fusion

[0242] The convolutional features of hand-drawn and automated graphs are added element-wise to integrate "prior knowledge" and "data-driven knowledge":

[0243]

[0244] (output) (This provides input for subsequent lightweight temporal convolutional layers).

[0245] Comparison between traditional TCN and lightweight multi-TCN:

[0246] Table 2 Comparison of Traditional TCN and Lightweight Multigraph TCN

[0247]

[0248] 8. Dual-output layer output:

[0249] The dual-task output module enables action classification and standardization scoring;

[0250] The specific implementation includes three sub-steps:

[0251] (1) Feature concatenation: The 64-dimensional spatial feature vector output by the FT-Transformer module and the 32-dimensional temporal feature vector output by the TCN module are concatenated horizontally to form a 96-dimensional fused feature vector;

[0252] First, the fused features are processed through causal convolution, then normalized and connected to residuals before being output. Then, global average pooling is used to aggregate the mean across the time series dimensions, using the following formula:

[0253]

[0254] Represents the temporal feature vector after pooling

[0255] Indicates the first The first batch, the first Frame, First The characteristic values ​​of each channel.

[0256] (2) Action classification:

[0257] Let the maximum value of the predicted probability distribution be... The threshold is The final classification result is:

[0258]

[0259] in This represents the maximum predicted probability among the 10 action categories. This represents the confidence threshold; the model only outputs a class when it has at least 50% confidence. The index of the class with the highest probability is indicated; the unknown action indicates that this result will be output when the model's confidence in all classes is below the threshold.

[0260] (3) Standard score:

[0261] Scoring logic: Total score = Spatial score + Temporal score

[0262] The standard score is calculated as follows:

[0263]

[0264] in Indicates spatial fraction; Indicates the spatial weight of the current action; Represents time series fractions; Indicates the temporal weight of the current action;

[0265] Spatial score: Cosine similarity between fused features and standard action feature templates × 50;

[0266] Timing score: the reciprocal of the deviation between the timing characteristics and the "standard timing template" × 50;

[0267] Dynamic weights are divided into spatial weights and temporal weights, both based on and The normalized weights ensure that the sum of the two is 1, adapting to the characteristics of each type of action:

[0268] Spatial weights:

[0269] Time series weights:

[0270] (4) Total loss function: Define the classification loss weights Change rate of loss in classification tasks Average difficulty level of the movements Rating loss weights Rate of change of scoring task loss Average spatial correlation strength Regularization weights Total number of model parameters Then, the classification loss, rating loss, and weight regularization are combined:

[0271]

[0272] in, Cross-entropy loss for action classification, The weights represent the cross-entropy loss for action classification. The mean squared error loss of the score, This represents the weight of the mean squared error loss in the scoring. For weight regularization, This represents the weight of the weight regularization term. For all model parameters.

[0273] 9. Model quantization and embedded deployment;

[0274] The specific implementation includes three sub-steps:

[0275] (1) Model quantization: The trained FP32 precision model was quantized to INT8 precision using TensorRT tool, which reduced the parameter memory usage by 75% (from 400KB to 100KB) and the accuracy loss was ≤2%;

[0276] (2) Model conversion: The quantized model is converted into binary firmware (.bin format) supported by the ARM Cortex-M4 chip through the TVM framework, and the compilation optimization option is set to "lightweight priority";

[0277] (3) Firmware burning and updating:

[0278] Burning: The .bin firmware is burned to the chip Flash (1MB) via the JTAG debugging interface, and the model startup time is ≤500ms;

[0279] As an improvement of the present invention, in step S6, the embedded chip adopts an ARM Cortex-M4 core (80MHz), and is equipped with 128KB RAM (for model operation) and 1MB Flash (for firmware storage). It can run continuously for ≥8 hours on a single charge (3.7V lithium battery) to meet the needs of outdoor training.

[0280] The present invention also provides a badminton grip action recognition and scoring system for implementing the above method, comprising:

[0281] (1) Feature data acquisition module: It consists of a sensor array and a data acquisition circuit. It is responsible for continuously acquiring grip feature data, conditioning the original signal, and performing digital-to-analog conversion. The resulting digital voltage value corresponds to the real-time pressure of a single sensor unit.

[0282] (2) Data preprocessing module: Implements 3-frame moving average denoising, Min-Max normalization, and invalid data removal;

[0283] (3) Cross-feature interaction FT-Transformer module: including spatial location encoding construction, linear layer dimension mapping, cosine similarity weighted association calculation, multi-head attention and FFN encoder;

[0284] (4) Lightweight multi-graph TCN module: includes manual and automatic graph construction, multi-layer lightweight causal convolution and global average pooling to extract temporal features;

[0285] (5) Dual-task output module: realizes action classification, standard score and conclusion generation;

[0286] (6) Embedded deployment module: responsible for model quantization and conversion;

[0287] (7) User interaction module: It consists of a mobile APP and racket vibration feedback. The APP displays the action category, score and conclusion.

[0288] The model recognition data of this invention is as follows:

[0289] Table 3. Action Recognition Accuracy and Scores of the Model of the Invention

[0290]

[0291] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent recognition and scoring of badminton racket grip actions based on cross-feature interaction, characterized in that, Includes the following steps: Step 1: Arrange an array of pressure sensors on the racket handle. Define the pressure signals sensed by all pressure sensors at the same time as a frame of data. The pressure data of a single pressure sensor corresponds to a feature point. Step 2: Collect continuous frame grip feature data at a preset frequency; Step 3: Perform 3-frame moving average denoising, normalization, and invalid frame processing on the continuous frame grip feature data to obtain the preprocessed continuous frame grip feature data and construct the training dataset. Step 4: Construct a grip action recognition and scoring model; the total loss function of the grip action recognition and scoring model. for: in, L cls Cross-entropy loss for action classification, L score The mean squared error loss of the score, L1Reg(W) All weights for the grip action recognition and scoring model L1 Regularization; where λ cls , λ score、 λ reg They are respectively L cls 、L score and L1Reg(W) The weights; Step 5: Train the action recognition and scoring model using the training dataset, with the total loss function... The model parameters are optimized with the goal of minimizing the loss until the loss converges and stabilizes, thus obtaining the trained grip action recognition and scoring model. Step Six: Action Classification and Scoring Output. The model takes each frame of pressure data generated by the trainee holding the racket handle, preprocessed in Step Three, and inputs it into the trained model. The model outputs the corresponding racket grip action classification results and standard score. The grip action recognition and scoring model includes a cross-feature interaction FT-Transformer module, a lightweight multi-graph TCN module, and a dual-task output module; The cross-feature interaction FT-Transformer module extracts cross features from single-frame grip feature data and outputs a single-frame cross-feature space feature vector. The lightweight multi-graph TCN module performs temporal modeling on cross-feature space features across 10 consecutive frames to construct hand-crafted and automatically learned graphs. The hand-crafted and automatically learned graphs extract features through independent graph convolution paths, and feature fusion is achieved by element-wise addition. The fused features are input into two layers of causal convolution, and after layer normalization, residual connections, and global average pooling, the temporal feature vector is finally obtained. The causal convolution kernel has 3 kernels, and the dilation rates are 1 and 2, respectively. The global average pooling method is as follows: The fused features are then processed through causal convolution, followed by layer normalization and residual connection for output. Then, global average pooling is used to aggregate the mean across the time series dimensions, using the following formula: Represents time series feature vectors This represents the feature value of the b-th batch, the t-th frame, and the c-th channel; The single-frame cross-feature space feature vector and temporal feature vector are input into the dual-task output module, which outputs an action classification and a standard score. The standard score is calculated as follows: Total score = Spatial score + Temporal score; Indicates spatial fraction; Indicates the spatial weight of the current action; Represents time series fractions; This indicates the temporal weight of the current action.

2. The method for intelligent recognition and scoring of badminton racket grip actions based on cross-feature interaction as described in claim 1, characterized in that, The data processing procedure of the cross-feature interaction FT-Transformer module is as follows: S11. Initialization of feature point embedding vectors: The array-type pressure sensor is a 4×10 matrix. To capture the relative position information of the sensors in two-dimensional space, the pressure value of each pressure sensor is first obtained. : in Let be the element in the i-th row and j-th column of the matrix, and s be the sensor number. The formula for calculating s is: ; A spatial location embedding layer is introduced at the input stage. This layer is used to handle coordinate normalization, location encoding, feature concatenation and mapping, and spatial offset coefficients. for: in For action For sensors The degree of importance; The method to obtain it is as follows: The pressure values ​​of each pressure sensor were collected from professional athletes during various movements, and then the following calculations were performed: in Indicates action a The set of indices for the downforce sensors; This represents the variance of the pressure values. Indicates the importance of features. Indicates the first The greatest importance of all pressure sensors under similar actions; Indicates the first Type of action, first The sample, the first The first frame Pressure values ​​from each pressure sensor; N represents the number of samples, and T represents the total number of frames; Indicates the first The first type of action The average pressure value of the entire sample and the entire frame of each sensor; This represents the weighting coefficient of the s-th pressure sensor corresponding to the a-th action; Then, the motion is calculated based on the spatial offset coefficient and the coordinate normalization value. a Below, the row index is i, and the column index is... j Pressure sensor position coding The location coding formula is: in, This indicates the normalization of row index i; Indicates column index j Normalized value ; The pressure value, coordinate normalization, and position code of the pressure sensor are used for feature concatenation; the feature concatenation formula is as follows: in, This represents the normalized value of the sensor. Indicates the spatial offset coefficient. Indicates the first Pressure values ​​from each sensor; S12: To improve the generalization ability of the grip action recognition and scoring model, a linear layer is used. The formula for the linear layer is: in, It is a learnable linear mapping matrix. For the corresponding bias term, This is an element-wise nonlinear activation function, used to introduce nonlinear expressive power; Indicates a nonlinear result; T denotes matrix transpose; Action after the linear layer The overall embedding matrix E: in, Indicates the current action Below is a high-dimensional set of embedded features for all pressure sensor inputs; S13: Acquire Action a Each pressure sensor s Contribution Finally, the 12 pressure sensors with the highest contribution were designated as the primary force-generating areas, and the remaining pressure sensors as secondary force-generating areas. Focusing on the core force signals corresponding to the primary force-generating areas of the grip, the main force-generating areas of the sensors are: in This represents the set of indices of the main force sensors used in action a; Subsequently, cosine similarity and weighted similarity between features were added to strengthen the model features. The cosine similarity is as follows: Where s1 and s2 are the pressure sensor numbers. They represent the first , The overall embedding matrix of the sensors; Indicates the first , Cosine similarity of the overall embedding matrix of each sensor; Different force application zones are assigned different weight parameters, and the dynamic intensity of any two pressure sensors is given for different actions. : Indicates action a Downward pressure sensor The degree of contribution; Indicates "and"; Indicates "or"; Combining feature similarity with dynamic weights, the weighted similarity is defined as: in Characterizing the pressure sensor under the current action a and The overall interaction strength between them; Finally, based on the force application zone to which the pressure sensor belongs, different association weights are assigned, and the weighted similarity is used as the message passing coefficient to perform neighborhood fusion on the features of each sensor to obtain aggregated features: in Indicates the first Aggregation characteristics of a pressure sensor; Indicates the first s An integrated matrix of pressure sensors Indicates the first The overall embedded matrix of pressure sensors other than the individual pressure sensors Indicates the first Pressure sensors other than the individual pressure sensor; S14: Set up 8 multi-head attention modules, initialize the Q, K, and V matrices using fully connected layers, and perform scaling dot products to eliminate the influence of matrix dimensions on aggregated features. The scaling dot product formula is: in, The feature dimension is represented by Q, K, and V, which represent the query vector matrix, key vector matrix, and value vector matrix, respectively. Represents the normalized exponential function, This indicates a multi-head attention module; Introducing a feedforward neural network module to aggregate features Perform sensor-by-sensor augmentation mapping: in, Indicates the first Aggregated characteristics of individual sensors , It is a learnable linear transformation matrix. For the corresponding bias term; Represents a non-linear activation function; Obtain the output of the feedforward neural network module : ； This represents the feedforward neural network module. for A set; The output of the feedforward neural network module represents the feature value of the s-th pressure sensor in the d-th dimension. Extract the global correlation features of all sensors within a single frame, and perform global average pooling on the enhanced feature output to output a single-frame cross-feature space feature vector. : 。 3. The intelligent recognition and scoring method for badminton racket grip actions based on cross-feature interaction as described in claim 2, characterized in that, The methods for constructing manual and automatically learned graphs are as follows: Constructing a manual adjacency matrix : ； , For batch indexing, Both are frame indices; B represents the batch size, and T represents the number of frames in the timing window; Generate handmade images : This indicates the b-th batch, the... Feature vectors of all feature channels of the frame Map the 10-dimensional main power generation area features to dimension, The dimension index representing the output features of the hand-drawn graph; The input dimension represents the features of the main power generation area. Indicates the dimension of the output feature; The automatic learning graph construction method is as follows: Generate automatic adjacency matrix （ ）； Let represent the characteristic Euclidean distance between the t-th frame and the τ-th frame in the b-th batch. For Euclidean distance, Let b be the feature value of the b-th batch, t-th frame, and d-th dimension. For the bth batch, the The frame, the d-th dimension's eigenvalue; Generate automatic learning graph : ; and Sharing core parameters, the 64-dimensional complete feature is mapped to 32 dimensions, and the output is... , This indicates the b-th batch, the... The 64-dimensional complete feature vector of the frame.

4. The method for intelligent recognition and scoring of badminton grip movements based on cross-feature interaction as described in claim 1, characterized in that, In step one, the array-type pressure sensors are arranged in a 4×10 matrix on the racket grip, covering the index-middle finger force area, the thumb auxiliary area, the ring-little finger support area, etc. After the original signal is conditioned, it is converted from digital to analog, and the obtained digital voltage value corresponds to the real-time pressure of a single sensor unit; in step two, the preset frequency is 50Hz.

5. The intelligent recognition and scoring method for badminton racket grip actions based on cross-feature interaction as described in claim 1, characterized in that, In step three, the preprocessing method is as follows: First, perform 3-frame moving average noise reduction: in, For the feature data of the i-th row and j-th column of the t-th frame, for Feature data after smoothing and denoising; Then, Min-Max normalization is performed to shrink the feature data to the range [0,1]. in, The minimum value of the feature data in the entire dataset. The maximum value of the feature data in the entire dataset. The feature data in the i-th row and j-th column after normalization; Secondly, for single-frame data, abnormal frames with pressure values ​​outside the range of [0, 100kPa] are removed. If the number of missing feature points is greater than 3, the frame is discarded directly; if the number of missing feature points is ≤ 3, the frame is filled with the average of the corresponding positions of the adjacent valid frames before and after it. in This represents the filled feature value of the sensor in the t-th frame, row i, and column j.

6. The intelligent recognition and scoring method for badminton racket grip actions based on cross-feature interaction as described in claim 1, characterized in that, The trained grip action recognition and scoring model is exported to a format supported by TensorRT. Quantization is performed using the calibration quantization mode and a toolchain. After quantization, the effect is verified based on embedded deployment metrics.

7. The intelligent recognition and scoring method for badminton racket grip actions based on cross-feature interaction as described in claim 1, characterized in that, The dual-task output module constructs a fully connected layer that horizontally concatenates cross-feature space feature vectors and temporal feature vectors to obtain a fused feature vector. The 1×96 dimensional fused feature vector is mapped to a 1×10 dimensional output. : in , These are learnable parameters; Then, Softmax activation is performed on the output of the fully connected layer, transforming it into a probability distribution for each action category. : in The index corresponding to the maximum value is the predicted action category; Next, set the confidence threshold. The final action category is determined based on the probability distribution: in This represents the maximum predicted probability among the 10 action categories. This represents the confidence threshold; the model only outputs a class when it has at least 50% confidence. The index representing the category with the highest probability. This indicates that the output is given when the model's confidence level for all categories is below the threshold. express The maximum value.

8. The method for intelligent recognition and scoring of badminton racket grip actions based on cross-feature interaction as described in claim 1, characterized in that, First, define the spatial correlation strength and temporal fluctuation coefficient. The spatial fraction is calculated using the cosine similarity method. in The 96-dimensional fusion feature template for standard movements was obtained through training with data from professional athletes. This is the 96-dimensional fusion feature of the current action, i.e., a fusion vector of spatial and temporal features. Cosine similarity; The time series fraction calculation uses the reciprocal of the deviation from the standard time series template: in The 32-dimensional temporal features of standard actions, The temporal characteristics of the current action, Indicates the absolute deviation of the time series characteristics; Then, the spatial and temporal scores are dynamically weighted to obtain the spatial and temporal weights, and finally, the total score is calculated. in Spatial weights, For time series weights.

9. A badminton racket grip action intelligent recognition and scoring system based on cross-feature interaction, characterized in that, The system is used to run the intelligent recognition and scoring method for badminton racket grip actions based on cross-feature interaction as described in any one of claims 1-8.

Images