Table tennis sports skill identification method based on feature fusion
Through feature fusion and principal component analysis method, combined with BP neural network, the shortcomings of existing table tennis movement skills recognition methods are solved, and efficient identification and monitoring of table tennis movement skills are achieved, and the training effect is improved.
Patent Information
- Application Number
- CN202411731609.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-05-02
AI Technical Summary
The existing table tennis sports movement skills recognition methods rely on machine vision technology, which has high computing power, high cost and poor portability, and cannot accurately identify the athlete's movement skills when the movement amplitude is large or visual information is missing.
Using a feature fusion method, the accelerometer and gyroscope data in the wearable device are segmented through the sliding window function, a table tennis movement action feature model is constructed, and feature fusion is performed through principal component analysis method, and finally the BP neural network is used for action skills recognition.
It improves the recognition ability of standard movement skills of table tennis, enhances the monitoring and recognition level of wearable devices for table tennis skills, and assists athletes in targeted improvement of the training effect of movement skills.
Smart Images
Figure CN119917992A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of feature engineering and pattern recognition, and in particular to a table tennis skill recognition method based on feature fusion. Background Art
[0002] Existing methods for identifying sports skills mainly rely on machine vision technology, the core of which is to capture the key movements of the human skeleton through imaging equipment, extract posture sequences, and then judge the type of sportsman's sports skills through extensive software training. Machine vision methods require large computing power, high cost, and poor equipment portability. At the same time, when the movement amplitude is large or visual information is missing, it is impossible to accurately identify the sportsman's sports skills.
[0003] In recent years, smart wearable devices have been widely used in the field of sports. They generally use sensors such as GPS, heart rate monitors, accelerometers, and gyroscopes to monitor vital signs and record basic sports information. They are portable and easy to use. However, the current wearable devices are not yet mature in the recognition and analysis of the action skills of complex sports such as table tennis. Therefore, we can rely on wearable devices to seek to establish a table tennis skill recognition method to meet the needs of athletes for portability of sports tools and refined action analysis.
[0004] The problem of identifying table tennis movement skills can essentially be attributed to the problem of pattern recognition, and artificial neural networks are widely used in the field of pattern recognition due to their powerful induction and reasoning capabilities. However, since the degree of completeness of the features input into the neural network to represent the movement is positively correlated with the recognition accuracy, and simple features such as movement speed, distance, and angle cannot well represent the complex movements of table tennis, it is necessary to establish a suitable feature model to extract complete movement features to improve the recognition accuracy.
[0005] At present, the existing table tennis motion skill recognition methods that can be applied to wearable devices have selected fewer feature dimensions, generally only considering the simple statistical values of three-dimensional synthetic acceleration and angular velocity data, which is also difficult to fully characterize table tennis motions, resulting in low recognition accuracy of motion skills. When the dimension of the established feature model is high, although it can better characterize the motion, the excessive input dimension will also cause the neural network structure to be complex, the amount of calculation is huge, and it is easy to overfit. Therefore, how to reduce the dimension of the neural network input feature while ensuring that the feature model can more comprehensively characterize the table tennis motion is a problem that needs to be solved in current research.
[0006] Feature fusion refers to the combination of features from different layers or branches. The fusion of multiple feature vectors can better extract and utilize information in the data, and can improve the generalization and robustness of the neural network model. In table tennis, the many features extracted from the action data are fused, and the original features are merged into a small number of more discriminative features. Then, a small number of features can be input into the neural network, and the features representing the table tennis movements are continuously iterated, and a certain mapping relationship between action features and action skills is established, thereby realizing the recognition of table tennis action skills. Summary of the invention
[0007] In order to overcome the shortcomings of the prior art, the present invention provides a table tennis skill recognition method based on feature fusion. First, a standard model of six table tennis motion skills is established; then, the X, Y, Z three-axis acceleration, three-axis angular velocity and three-axis attitude angle data recorded by the accelerometer and gyroscope in the wearable device are divided into data segments of specific time lengths through a sliding window function, and the action activation function is used to determine whether a standard motion skill occurs in each segmented data segment; secondly, a table tennis motion feature model is constructed and feature fusion is performed through the principal component analysis method; finally, table tennis motion skills are recognized through a neural network. The present invention effectively improves the recognition ability of standard table tennis motion skills, is conducive to improving the monitoring and recognition level of wearable devices for table tennis skills, and better assists athletes in targeted training for improving table tennis motion skills.
[0008] The present invention provides a table tennis skill recognition method based on feature fusion, and the adopted technical solution comprises the following steps:
[0009] Step 1: Establish a standard action skill model for table tennis movements;
[0010] Step 2: Segment the three-axis acceleration, angular velocity and attitude angle data in motion into data segments through a sliding window function, and determine whether a standard table tennis action skill occurs in each data segment through an action activation function;
[0011] Step 3, constructing a table tennis motion feature model according to the table tennis motion features;
[0012] Step 4: fuse the action features through principal component analysis to determine the input features for action skill recognition;
[0013] Step 5: Identify table tennis movement skill types through BP neural network.
[0014] Further, in step 1, the process of establishing the standard action skill model is as follows:
[0015] S1.1, determine that there are six standard table tennis skills, which are: forehand attack M1, backhand attack M2, forehand rubbing M3, backhand rubbing M4, forehand slice M5 and backhand slice M6;
[0016] S1.2, establish a standard action skill database: use wearable devices to collect six standard action skills of professional athletes in a complete table tennis cycle. A complete table tennis cycle refers to the swing process before the athlete's racket touches the table tennis ball and the process of pulling back the racket after the racket touches the table tennis ball, and establish a standard action skill database;
[0017] S1.3, for each standard action skill, nine types of data are collected when the athlete completes a swing, including three-axis acceleration, three-axis angular velocity and three-axis posture angle, for a total of 54 sets of standard data.
[0018] Further, in step 2, the judgment process is as follows:
[0019] S2.1, construct sliding window function:
[0020] Statistical analysis of the time required to complete each standard action skill of table tennis. The time required to complete a standard action skill is 1s to 2s. When the athlete completes multiple action skills continuously, in order to facilitate subsequent identification, the multiple sets of motion data collected need to be segmented. Therefore, a sliding window function needs to be constructed.
[0021] When the sensor of the wearable device collects the three-axis motion data of a standard motion skill, n sets of data are generated, and the sliding window function is:
[0022]
[0023] Where t is the position index of the sliding window function. When there are D groups of data in the data set collected by the wearable device, t takes an integer from 0 to 2D / n. At this time, the sliding window function is The group data is the moving interval, n groups of data are a data segment, and n is the number of data groups in a data segment; the three-axis acceleration, three-axis angular velocity and three-axis attitude angle data collected by the wearable device are segmented by a sliding window function;
[0024] S2.2, construct action skill activation function:
[0025] Due to the instability of the athlete's movements, a standard action skill may not occur in one of the n data segments segmented by the sliding window function. Therefore, the action skill activation function is constructed using the three-axis acceleration and three-axis angular velocity data in the motion data as indicators to determine whether the table tennis player has completed a standard action skill in a data segment; the action activation function is:
[0026]
[0027] Among them, a ix 、a iy 、a iz They are the X, Y, and Z coordinates of the i-th group of motion data collected. Axis direction The acceleration value, ω ix ,ω iy ,ω iz They are the X, Y, and Z coordinates of the i-th group of motion data. Axis direction When the activation function value is greater than the action activation threshold 95, it is considered that a standard action skill has occurred, otherwise, it is considered that no standard action skill has occurred;
[0028] S2.3, when the action skill activation function determines that a standard action skill has occurred in a data segment, the data segment is considered valid and the action skill type identification process continues. When the action skill activation function determines that no standard action skill has occurred in a data segment, the data segment is considered invalid and the t in the sliding window function is changed to t+1, that is, the data sequence number in the data segment is moved backward. The newly formed data segment is judged again using the action skill activation function until all the data sets collected by the wearable device are judged.
[0029] Further, in step 3, the method for establishing the table tennis motion feature model is as follows:
[0030] For the data segments that are judged as valid by the action skill activation function in step 2, it is necessary to extract multiple features of the action data in the data segments for the classification of action skills. The completeness of the extracted action features in representing the movement will greatly affect the recognition accuracy of the action skills, so it is necessary to establish a suitable feature model to extract the action features.
[0031] S3.1, determine 15 types of characteristic indicators in the table tennis motion data: the 15 types of characteristic indicators are the average value, variance, maximum value, minimum value, peak-to-valley value, mean square value, root mean square, correlation coefficient, peak factor, impulse factor, margin factor, kurtosis factor, waveform factor, skewness and kurtosis of the table tennis motion data; the table tennis motion data used to calculate the motion characteristics are the motion acceleration, angular velocity and attitude angle of the XYZ three axes and the composite direction;
[0032] S3.2, based on the 15 types of characteristic indicators in the table tennis motion data, a table tennis motion characteristic model is constructed:
[0033] S3.2.1. Determine the average value of table tennis motion data: the average value is denoted as F 1 ,but where si =(a ix ,a iy ,a iz ,a i ,ω ix ,ω iy ,ω iz ,ω i ,θ ix ,θ iy ,θ iz ,θ i ), a i ,ω i ,θ i Respectively represent the resultant acceleration, angular velocity and attitude angle, θ ix ,θ iy ,θ iz Respectively represent the X, Y, and Z axis attitude angles;
[0034] S3.2.2. Determine the variance of table tennis motion data: The variance is denoted by F 2 ,but
[0035] S3.2.3. Determine the maximum value of table tennis motion data: the maximum value is recorded as F 3 , then F 3 =max(s i ), max(s i ) means taking the maximum value in a data segment for each component of si;
[0036] S3.2.4. Determine the minimum value of table tennis motion data: the minimum value is recorded as F 4 , then F 4 =min(s i ), min(s i ) represents the i Each component of takes the minimum value in a data segment;
[0037] S3.2.5. Determine the peak and valley values of table tennis motion data: The peak and valley values are recorded as F 5 , then F 5 =max(s i )-min(s i );
[0038] S3.2.6. Determine the mean square value of table tennis motion data: The mean square value is denoted as F 6 ,but
[0039] S3.2.7. Determine the root mean square of table tennis motion data: The root mean square is denoted as F 7 ,but
[0040] S3.2.8. Determine the correlation coefficient of table tennis motion data: The correlation coefficient is denoted by F 8 ,but
[0041] S3.2.9. Determine the peak factor of table tennis motion data: The peak factor is denoted by F 9 ,but
[0043] S3.2.10. Determine the impulse factor of table tennis motion data: The impulse factor is denoted by F 10 ,but
[0045] S3.2.11. Determine the margin factor for table tennis motion data: The margin factor is denoted by F 11 ,but
[0047] S3.2.12. Determine the kurtosis factor of table tennis motion data: The kurtosis factor is denoted by F 12 ,but
[0049] S3.2.13. Determine the waveform factor of table tennis motion data: The waveform factor is denoted by F 13 ,but
[0051] S3.2.14. Determine the skewness of table tennis motion data: The skewness is denoted by F 14 ,but
[0053] S3.2.15. Determine the kurtosis of table tennis motion data: The kurtosis is denoted by F 15 ,but
[0055] In a valid data segment containing n sets of data, according to F 1 To F 15 These 15 types of feature models extract 15×12 180-dimensional motion features from table tennis motion data for the recognition of table tennis motion skills.
[0056] Further, in step 4, the feature fusion method is as follows:
[0057] S4.1, based on principal component analysis and action feature data, the decentralized matrix A is obtained pq ,but
[0059] where a pqrepresents the qth feature of the pth valid data segment, r is the number of valid data segments, and j is the feature dimension;
[0060] S4.2, according to the decentralization matrix A pq , calculate the covariance matrix C, then
[0061] S4.3, perform eigendecomposition on the covariance matrix and find the eigenvalue λ of the covariance matrix k and the corresponding eigenvector v k , then Cv k =λ k v k ,
[0062] where k = 1, 2, ..., j, λ k represents the kth eigenvalue, v k represents the kth eigenvector;
[0063] S4.4, the contribution rate u of the qth feature to the total features q ,but
[0064] S4.5, for u q Sort them in descending order and calculate the cumulative contribution rate U q ,
[0065]
[0066] When U q When >δ, the first q features are taken as the principal components of the action features, and the first q features are new features for table tennis action skill recognition, where δ is the information retention rate.
[0067] Further, in step 5, the table tennis movement skill recognition is performed by using a BP neural network model, wherein the BP neural network model has four layers: an input layer, an output layer, and two hidden layers. The number of neurons in the input layer of the BP neural network model is the new feature dimension after feature fusion, the number of neurons in the output layer is 6, and the hidden layer is 2 layers, with a total of 120 neurons;
[0068] Based on the new features constructed by the principal component analysis method in step 4, a BP neural network model is built, and the 4-layer BP neural network is trained to obtain the functional relationship between the input features and the output action type, so as to realize the recognition of table tennis movement skills.
[0069] The beneficial effects of the present invention are as follows: based on the feature fusion method in feature engineering, 15 types of feature models are constructed, and the principal component analysis method is used for feature fusion, thereby establishing a table tennis movement skill recognition method, which can accurately and effectively recognize the various table tennis movement skills of the athletes, and assist the athletes to carry out targeted training to improve the table tennis movement skills; at the same time, the method improves the accuracy of table tennis movement skill recognition using a neural network model, and can improve the monitoring and recognition level of table tennis movement skills by wearable devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1 It is a schematic diagram of the process of the present invention;
[0071] Figure 2 It is a model diagram of the six standard movements and skills of table tennis;
[0072] Figure 3 It is a schematic diagram of the sliding window function dividing the three-axis acceleration, three-axis angular velocity and three-axis attitude angle data into data segments;
[0073] Figure 4 It is a schematic diagram of the structure of a 4-layer BP neural network constructed by the present invention;
[0074] Figure 5 This is a comparison chart of recognition accuracy under different feature dimensions. DETAILED DESCRIPTION
[0075] The present invention is further described below in conjunction with the accompanying drawings and embodiments. The present invention includes but is not limited to the following embodiments.
[0076] In order to accurately identify the characteristics of table tennis skills and distinguish the types of action skills, four aspects of work need to be completed: first, a standard model of table tennis action skills needs to be established; second, the problem of processing continuous motion data in actual sports needs to be solved. Effective motion information processing can reduce the difficulty of subsequent feature fusion and action skill identification; third, it is necessary to use the motion data measured by sensors in wearable devices to establish a suitable table tennis action feature model. The feature model needs to accurately and comprehensively reflect the information contained in different table tennis action skills; third, it is necessary to use a reasonable feature engineering method to fuse multi-dimensional action features, determine the input of the neural network model, and then realize the identification of table tennis skills through neural network training.
[0077] The present invention provides a table tennis skill recognition method based on feature fusion, the process is as follows: Figure 1As shown. First, a standard model of six table tennis motion skills is established; then, the X, Y, Z three-axis acceleration, three-axis angular velocity and three-axis attitude angle data recorded by the accelerometer and gyroscope in the wearable device are divided into data segments of specific time lengths through a sliding window function, and the action activation function is used to determine whether a standard motion skill occurs in each segmented data segment; secondly, a table tennis motion feature model is constructed and feature fusion is performed through the principal component analysis method; finally, the table tennis motion skills are identified through a neural network. The present invention effectively improves the recognition ability of table tennis standard motion skills, is conducive to improving the monitoring and recognition level of table tennis skills by wearable devices, and better assists athletes to carry out targeted training on table tennis motion skills.
[0078] The present invention provides a table tennis skill recognition method based on feature fusion, and the adopted technical solution comprises the following steps:
[0079] Step 1: Establish a standard action skill model for table tennis movements;
[0080] Step 2: Segment the three-axis acceleration, angular velocity and attitude angle data in motion into data segments through a sliding window function, and determine whether a standard table tennis action skill occurs in each data segment through an action activation function;
[0081] Step 3, constructing a table tennis motion feature model according to the table tennis motion features;
[0082] Step 4: fuse the action features through principal component analysis to determine the input features for action skill recognition;
[0083] Step 5: Identify table tennis movement skill types through BP neural network.
[0084] Further, in step 1, the process of establishing the standard action skill model is as follows:
[0085] S1.1, determine that there are six standard table tennis skills, which are: forehand attack M1, backhand attack M2, forehand rubbing M3, backhand rubbing M4, forehand slice M5 and backhand slice M6; Figure 2 shown.
[0086] S1.2, establish a standard action skill database: use wearable devices to collect six standard action skills of professional athletes in a complete table tennis cycle. A complete table tennis cycle refers to the swing process before the athlete's racket touches the table tennis ball and the process of pulling back the racket after the racket touches the table tennis ball, and establish a standard action skill database;
[0087] S1.3, for each standard action skill, nine types of data are collected when the athlete completes a swing, including three-axis acceleration, three-axis angular velocity and three-axis posture angle, for a total of 54 sets of standard data.
[0088] Further, in step 2, the judgment process is as follows:
[0089] S2.1, construct sliding window function:
[0090] Statistical analysis of the time required to complete each standard action skill of table tennis. The time required to complete a standard action skill is 1s to 2s. When the athlete completes multiple action skills continuously, in order to facilitate subsequent identification, the multiple sets of motion data collected need to be segmented. Therefore, a sliding window function needs to be constructed.
[0091] When the athlete completes multiple action skills continuously, in order to facilitate subsequent recognition, it is necessary to segment the multiple sets of motion data collected. Therefore, a sliding window function is constructed. Where t is the position index of the sliding window function. When there are D groups of data in the data set collected by the wearable device, t takes an integer from 0 to 2D / n. The sliding window function uses 50 groups as the moving interval and 100 groups of data as a data segment to segment the collected three-axis acceleration, angular velocity and attitude angle data.
[0092] When the sensor of the wearable device collects the three-axis motion data of a standard motion skill, n sets of data are generated, and the sliding window function is:
[0093]
[0094] Where t is the position index of the sliding window function. When there are D groups of data in the data set collected by the wearable device, t takes an integer from 1 to 2D / n. At this time, the sliding window function is The group data is the moving interval, n groups of data are a data segment, and n is the number of data groups in a data segment; the three-axis acceleration, three-axis angular velocity and three-axis attitude angle data collected by the wearable device are segmented by a sliding window function;
[0095] S2.2, construct action skill activation function:
[0096] Due to the instability of the athlete's movements, a standard action skill may not occur in one of the n data segments segmented by the sliding window function. Therefore, the action skill activation function is constructed using the three-axis acceleration and three-axis angular velocity data in the motion data as indicators to determine whether the table tennis player has completed a standard action skill in a data segment; the action activation function is:
[0097]
[0098] Among them, aix 、a iy 、a iz They are the X, Y, and Z coordinates of the i-th group of motion data collected. Axis direction The acceleration value, ω ix ,ω iy ,ω iz They are the X, Y, and Z coordinates of the i-th group of motion data. Axis direction The angular velocity value of the direction. When the value of the activation function is greater than the action activation threshold of 95, it is considered that a standard action skill has occurred. Otherwise, it is considered that no standard action skill has occurred.
[0099] S2.3, when the action skill activation function determines that a standard action skill has occurred in a data segment, the data segment is considered valid and the action skill type identification process continues. When the action skill activation function determines that no standard action skill has occurred in a data segment, the data segment is considered invalid and the t in the sliding window function is changed to t+1, that is, the data sequence number in the data segment is moved backward. The newly formed data segment is judged again using the action skill activation function until all the data sets collected by the wearable device are judged. Figure 3 Shown
[0100] Further, in step 3, the method for establishing the table tennis motion feature model is as follows:
[0101] For the data segments that are judged as valid by the action skill activation function in step 2, it is necessary to extract multiple features of the action data in the data segments for the classification of action skills. The completeness of the extracted action features in representing the movement will greatly affect the recognition accuracy of the action skills, so it is necessary to establish a suitable feature model to extract the action features.
[0102] S3.1, determine 15 types of characteristic indicators in the table tennis motion data: the 15 types of characteristic indicators are the average value, variance, maximum value, minimum value, peak-to-valley value, mean square value, root mean square, correlation coefficient, peak factor, impulse factor, margin factor, kurtosis factor, waveform factor, skewness and kurtosis of the table tennis motion data; the table tennis motion data used to calculate the motion characteristics are the motion acceleration, angular velocity and attitude angle of the XYZ three axes and the composite direction;
[0103] S3.2, based on the 15 types of characteristic indicators in the table tennis motion data, a table tennis motion characteristic model is constructed:
[0104] S3.2.1. Determine the average value of table tennis motion data: the average value is denoted as F 1 ,but where s i =(a ix ,aiy ,a iz ,a i ,ω ix ,ω iy ,ω iz ,ω i ,θ ix ,θ iy ,θ iz ,θ i ), a i ,ω i ,θ i Respectively represent the resultant acceleration, angular velocity and attitude angle, θ ix ,θ iy ,θ iz Respectively represent the X, Y, and Z axis attitude angles;
[0105] S3.2.2. Determine the variance of table tennis motion data: The variance is denoted by F 2 ,but
[0106] S3.2.3. Determine the maximum value of table tennis motion data: the maximum value is recorded as F 3 , then F 3 =max(s i ), max(s i ) represents the i Each component of takes the maximum value in a data segment;
[0107] S3.2.4. Determine the minimum value of table tennis motion data: the minimum value is recorded as F 4 , then F 4 =min(s i ), min(s i ) represents the i Each component of takes the minimum value in a data segment;
[0108] S3.2.5. Determine the peak and valley values of table tennis motion data: The peak and valley values are recorded as F 5 , then F 5 =max(s i )-min(s i );
[0109] S3.2.6. Determine the mean square value of table tennis motion data: The mean square value is denoted as F 6 ,but
[0110] S3.2.7. Determine the root mean square of table tennis motion data: The root mean square is denoted as F 7 ,but
[0111] S3.2.8. Determine the correlation coefficient of table tennis motion data: The correlation coefficient is denoted by F 8 ,but
[0112] S3.2.9. Determine the peak factor of table tennis motion data: The peak factor is denoted by F 9 ,but
[0114] S3.2.10. Determine the impulse factor of table tennis motion data: The impulse factor is denoted by F 10 ,but
[0116] S3.2.11. Determine the margin factor for table tennis motion data: The margin factor is denoted by F 11 ,but
[0118] S3.2.12. Determine the kurtosis factor of table tennis motion data: The kurtosis factor is denoted by F 12 ,but
[0120] S3.2.13. Determine the waveform factor of table tennis motion data: The waveform factor is denoted by F 13 ,but
[0122] S3.2.14. Determine the skewness of table tennis motion data: The skewness is denoted by F 14 ,but
[0124] S3.2.15. Determine the kurtosis of table tennis motion data: The kurtosis is denoted by F 15 ,but
[0126] In a valid data segment containing 100 sets of data, according to F 1 To F 15 These 15 types of feature models extract 15×12 180-dimensional motion features from table tennis motion data for the recognition of table tennis motion skills.
[0127] Further, in step 4, the feature fusion method is as follows:
[0128] S4.1, based on principal component analysis and action feature data, the decentralized matrix A is obtained pq ,but
[0130] where a pq represents the qth feature of the pth valid data segment, r is the number of valid data segments, and j is the feature dimension;
[0131] S4.2, according to the decentralization matrix A pq , calculate the covariance matrix C, then
[0132] S4.3, perform eigendecomposition on the covariance matrix and find the eigenvalue λ of the covariance matrix k and the corresponding eigenvector v k , then Cv k =λ k v k ,
[0133] where k = 1, 2, ..., j, λ k represents the kth eigenvalue, v k represents the kth eigenvector;
[0134] S4.4, the contribution rate u of the qth feature to the total features q ,but
[0135] S4.5, for u q Sort them in descending order and calculate the cumulative contribution rate U q ,
[0136]
[0137] When U q When >δ, the first q features are taken as the principal components of the action features, and the first q features are new features for table tennis action skill recognition, where δ is the information retention rate.
[0138] Furthermore, when the number of valid data segments r=300, the feature dimension j=180, and the information retention rate δ=95%, the new feature dimension constructed by the principal component analysis method is 23 dimensions.
[0139] The number of input layer neurons obtained by feature fusion is 23, the number of output layer neurons is set to 6, the training method adopts stochastic gradient descent, the hidden layer has 2 layers, the number of hidden layer neurons is 120, and the learning rate is 0.01. The BP neural network structure constructed is as follows: Figure 4 shown.
[0140] Further, in step 5, the table tennis movement skill recognition is performed by using a BP neural network model, wherein the BP neural network model has four layers: an input layer, an output layer, and two hidden layers. The number of neurons in the input layer of the BP neural network model is the new feature dimension after feature fusion, the number of neurons in the output layer is 6, and the hidden layer is 2 layers, with a total of 120 neurons;
[0141] Based on the action features constructed by the principal component analysis method in step 4, a BP neural network model is built, and the 4-layer BP neural network is trained to obtain the functional relationship between the input features and the output action type, so as to realize the recognition of table tennis action skills.
[0142] Different numbers of feature types were selected and divided into group 1, group 2, and group 3. The features contained in each group were used to train the BP neural network. The accuracy of the action skill recognition of the training set and the test set in each group was counted. The results are as follows: Figure 5 As shown;
[0143] Among them, the feature types included in group one include the mean, variance, maximum, minimum and peak-to-valley values of the acceleration, angular velocity and attitude angle of the motion action in the three directions of X, Y, Z and the combined direction; the feature types included in group two include the mean, variance, maximum, minimum, peak-to-valley value, mean square value, root mean square and correlation coefficient of the acceleration, angular velocity and attitude angle of the motion action in the three directions of X, Y, Z and the combined direction; group three includes all 15 types of motion features included in step three;
[0144] from Figure 5 It can be seen that the accuracy of neural network recognition of table tennis skills using the 15-category feature model constructed by the present invention is 98.33%, which is higher than the recognition accuracy of Group 1 and Group 2. This shows that the feature model constructed by the present invention and the feature fusion method used are necessary and effective.
[0145] It is understandable that those skilled in the art can make equivalent substitutions or changes based on the technical solution and inventive concept of the present invention, and all these changes or substitutions should fall within the protection scope of the claims attached to the present invention.
Claims
1. A table tennis skill recognition method based on feature fusion, characterized in that: The table tennis skill recognition method comprises the following steps: Step 1: Establish a standard action skill model for table tennis movements; Step 2: Segment the three-axis acceleration, angular velocity and attitude angle data in motion into data segments through a sliding window function, and determine whether a standard table tennis action skill occurs in each data segment through an action activation function; Step 3, constructing a table tennis motion feature model according to the table tennis motion features; Step 4: fuse the action features through principal component analysis to determine the input features for action skill recognition; Step five, identify the table tennis movement skill types through BP neural network.
2. A table tennis skill recognition method based on feature fusion according to claim 1, characterized in that: In step 1, the process of establishing the standard action skill model is as follows: S1.1, determine that there are six standard table tennis skills, which are: forehand attack M1, backhand attack M2, forehand rubbing M3, backhand rubbing M4, forehand slice M5 and backhand slice M6; S1.2, establish a standard action skill database: use wearable devices to collect six standard action skills of professional athletes in a complete table tennis cycle. A complete table tennis cycle refers to the swing process before the athlete's racket touches the table tennis ball and the process of pulling back the racket after the racket touches the table tennis ball, and establish a standard action skill database; S1.3, for each standard action skill, nine types of data are collected when the athlete completes a swing, including three-axis acceleration, three-axis angular velocity and three-axis posture angle, for a total of 54 sets of standard data.
3. A table tennis skill recognition method based on feature fusion according to claim 1, characterized in that: In step 2, the judgment process is as follows: S2.1, construct sliding window function: Statistical analysis of the time required to complete each standard action skill of table tennis. The time required to complete a standard action skill is 1s to 2s. When the athlete completes multiple action skills continuously, in order to facilitate subsequent identification, the multiple sets of motion data collected need to be segmented. Therefore, a sliding window function needs to be constructed. When the sensor of the wearable device collects the three-axis motion data of a standard motion skill, n sets of data are generated, and the sliding window function is: Where t is the position index of the sliding window function. When there are D groups of data in the data set collected by the wearable device, t takes an integer from 0 to 2D / n. At this time, the sliding window function is The group data is the moving interval, n groups of data are a data segment, and n is the number of data groups in a data segment; the three-axis acceleration, three-axis angular velocity and three-axis attitude angle data collected by the wearable device are segmented by a sliding window function; S2.2, construct action skill activation function: Due to the instability of the athlete's movements, a standard action skill may not occur in one of the n data segments segmented by the sliding window function. Therefore, the action skill activation function is constructed using the three-axis acceleration and three-axis angular velocity data in the motion data as indicators to determine whether the table tennis player has completed a standard action skill in a data segment; the action activation function is: Among them, a ix 、a iy 、a iz They are the X, Y, and Z coordinates of the i-th group of motion data collected. Axis direction The acceleration value, ω ix ,ω iy ,ω iz They are the X, Y, and Z coordinates of the i-th group of motion data. Axis direction When the activation function value is greater than the action activation threshold 95, it is considered that a standard action skill has occurred, otherwise, it is considered that no standard action skill has occurred; S2.3, when the action skill activation function determines that a standard action skill has occurred in a data segment, the data segment is considered valid and the action skill type identification process continues. When the action skill activation function determines that no standard action skill has occurred in a data segment, the data segment is considered invalid and the t in the sliding window function is changed to t+1, that is, the data sequence number in the data segment is moved backward. The newly formed data segment is judged again using the action skill activation function until all the data sets collected by the wearable device are judged.
4. A table tennis skill recognition method based on feature fusion according to claim 1, characterized in that: In step 3, the method for establishing the table tennis motion feature model is as follows: For the data segments that are judged as valid by the action skill activation function in step 2, it is necessary to extract multiple features of the action data in the data segments for the classification of action skills. The completeness of the extracted action features in representing the movement will greatly affect the recognition accuracy of the action skills, so it is necessary to establish a suitable feature model to extract the action features. S3.1, determine 15 types of characteristic indicators in the table tennis motion data: the 15 types of characteristic indicators are the average value, variance, maximum value, minimum value, peak-to-valley value, mean square value, root mean square, correlation coefficient, peak factor, impulse factor, margin factor, kurtosis factor, waveform factor, skewness and kurtosis of the table tennis motion data; the table tennis motion data used to calculate the motion characteristics are the motion acceleration, angular velocity and attitude angle of the XYZ three axes and the composite direction; S3.2, construct a table tennis motion feature model based on 15 types of feature indicators in table tennis motion data.
5. A table tennis skill recognition method based on feature fusion according to claim 4, characterized in that: In S3.2, the method for constructing the table tennis motion feature model is as follows: S3.2.
1. Determine the average value of table tennis motion data: the average value is recorded as F1, then where s i =(a ix ,a iy ,a iz ,a i ,ω ix ,ω iy ,ω iz ,ω i ,θ ix ,θ iy ,θ iz ,θ i ), a i ,ω i ,θ i Respectively represent the resultant acceleration, angular velocity and attitude angle, θ ix ,θ iy ,θ iz Respectively represent the X, Y, and Z axis attitude angles; S3.2.
2. Determine the variance of table tennis motion data: The variance is denoted as F2, then S3.2.
3. Determine the maximum value of table tennis motion data: the maximum value is recorded as F3, then F3 = max(s i ), max(s i ) represents the i Each component of takes the maximum value in a data segment; S3.2.
4. Determine the minimum value of table tennis motion data: the minimum value is recorded as F4, then F4 = min (s i ), min(s i ) represents the i Each component of takes the minimum value in a data segment; S3.2.
5. Determine the peak-to-valley value of table tennis motion data: The peak-to-valley value is recorded as F5, then F5 = max(s i )-min(s i ); S3.2.
6. Determine the mean square value of table tennis motion data: The mean square value is denoted as F6, then S3.2.
7. Determine the root mean square of table tennis motion data: The root mean square is denoted as F7, then S3.2.
8. Determine the correlation coefficient of table tennis motion data: The correlation coefficient is denoted as F8, then S3.2.
9. Determine the peak factor of table tennis motion data: The peak factor is denoted as F9, then S3.2.
10. Determine the impulse factor of table tennis motion data: The impulse factor is denoted by F 10 ,but S3.2.
11. Determine the margin factor for table tennis motion data: The margin factor is denoted by F 11 ,but S3.2.
12. Determine the kurtosis factor of table tennis motion data: The kurtosis factor is denoted by F 12 ,but S3.2.
13. Determine the waveform factor of table tennis motion data: The waveform factor is denoted by F 13 ,but S3.2.
14. Determine the skewness of table tennis motion data: The skewness is denoted by F 14 ,but S3.2.
15. Determine the kurtosis of table tennis motion data: The kurtosis is denoted by F 15 ,but In a valid data segment containing n sets of data, according to F1 to F 15 These 15 types of feature models extract 15×12 180-dimensional motion features from table tennis motion data for the recognition of table tennis motion skills.
6. A table tennis skill recognition method based on feature fusion according to claim 1, characterized in that: In step 4, the feature fusion method is as follows: S4.1, based on principal component analysis and action feature data, the decentralized matrix A is obtained pq ,but where a pq represents the qth feature of the pth valid data segment, r is the number of valid data segments, and j is the feature dimension; S4.2, according to the decentralization matrix A pq , calculate the covariance matrix C, then S4.3, perform eigendecomposition on the covariance matrix and find the eigenvalue λ of the covariance matrix k and the corresponding eigenvector v k , then Cv k =λ k v k , where k = 1, 2, ..., j, λ k represents the kth eigenvalue, v k represents the kth eigenvector; S4.4, the contribution rate u of the qth feature to the total features q ,but S4.5, for u q Sort them in descending order and calculate the cumulative contribution rate U q , When U q When >δ, the first q features are used as the principal components of action features for table tennis action skill recognition, where δ is the information retention rate.
7. A table tennis skill recognition method based on feature fusion according to claim 1, characterized in that: In step 5, the table tennis movement skill recognition is performed by using a BP neural network model, wherein the BP neural network model has four layers, namely, an input layer, an output layer, and two hidden layers. The number of neurons in the input layer of the BP neural network model is the new feature dimension after feature fusion, the number of neurons in the output layer is 6, and the hidden layer is 2 layers, with a total of 120 neurons. Based on the new features constructed by the principal component analysis method in step 4, a BP neural network model is built, and the 4-layer BP neural network is trained to obtain the functional relationship between the input features and the output action type, so as to realize the recognition of table tennis movement skills.
8. An electronic device, characterized in that: include: one or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured to execute the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes, and the program codes can be called by a processor to execute the method according to any one of claims 1 to 6.