An action evaluation method based on deep learning

Abstract

The present invention discloses a motion evaluation method based on deep learning, which relates to the field of data processing technology and includes the following steps: S1, using a three-axis accelerometer to obtain a posture data set of a user within a time period; S2, using the posture data set of the user within the time period, calculating the original posture weights of the joint points to obtain a fused posture set; S3, using the fused posture set to generate the user's motion posture evaluation result. In terms of data processing, the present invention fully fuses the posture data set to improve the reliability of data processing; and the classifier can obtain accurate evaluation results based on the fused posture set, improve the accuracy of the evaluation results, and provide users with more valuable reference action guidance.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an action evaluation method based on deep learning. Background Art

[0002] With the advancement of science and technology and the improvement of human living standards, people's demand for various sports is growing. During exercise, people are not only satisfied with the simple amount of exercise, but are more concerned about the similarity between their own movements and template movements. Most of the current sports training requires on-site guidance and teaching, and the training cost is too high. It is often difficult for people to get the coach's movement guidance at any time. In addition, even if learning is done by watching teaching videos, there will be a large difference between people's movements during practice and template movements due to differences in comprehension ability and the lack of targeted teaching. Therefore, people's demand for action evaluation guidance is increasing. In the prior art, the user's real-time action is usually compared with the template action at a certain moment in a static manner, and it is not possible to combine the actions within a certain time period for a comprehensive comparison, and the accuracy of the comparison result is low. Summary of the invention

[0003] In order to solve the above problems, the present invention proposes an action evaluation method based on deep learning.

[0004] The technical solution of the present invention is: a motion evaluation method based on deep learning comprises the following steps:

[0005] S1, using a three-axis accelerometer to obtain a user's posture data set within a time period;

[0006] S2, using the user's posture data set within the time period, calculate the original posture weights of the joint points to obtain a fused posture set;

[0007] S3. Generate the user's action posture evaluation result by using the fused posture set.

[0008] Furthermore, in S1, the posture data set includes the acceleration of each joint point of the user in the X-axis, the acceleration of the Y-axis, and the acceleration of the Z-axis in the time period.

[0009] Furthermore, S2 includes the following sub-steps:

[0010] S21, determining the posture components of the joint points at each moment in the time period according to the acceleration of each joint point of the user on the X-axis, the acceleration on the Y-axis, and the acceleration on the Z-axis in the time period;

[0011] S22, constructing a posture change curve for each joint point with the time period as the horizontal axis and the posture component as the vertical axis;

[0012] S23, determining the original posture weight of the joint point according to the posture change curve of the joint point;

[0013] S24. According to the original posture weights of the joint points, the posture data sets of the joint points within the time period are fused to generate fused posture values ​​of the joint points, thereby obtaining a fused posture set.

[0014] The beneficial effect of the above further scheme is that: in the present invention, each joint point has accelerations in different directions in three-dimensional space, and the accelerations of the joint points in different directions are characterized by curves, and the characteristics of the accelerations in different directions are determined to form the original posture weights. The acceleration data of the X-axis, Y-axis and Z-axis are fused through the original posture weights, and the overall motion state of the user's joint points in three-dimensional space can be comprehensively considered, so as to obtain more accurate and comprehensive motion information. This fusion process can reduce the errors and uncertainties that may be caused by single-axis acceleration data, and improve the accuracy and reliability of data processing.

[0015] Further, in S21, the joint point is The weight of the posture at each moment The calculation formula is:

[0016] ;

[0017] In the formula, Indicates the joint point The acceleration on the X axis at a given moment, Indicates the joint point The acceleration on the Y axis at that moment, Indicates the joint point The acceleration on the Z axis at that moment.

[0018] Furthermore, in S23, the number of zero-crossing points in the posture change curve is used as the original posture weight of the joint point.

[0019] Furthermore, in S24, the fusion posture value of the joint point The calculation formula is:

[0020] ;

[0021] In the formula, Represents the original posture weight of the joint point, Indicates the joint point The acceleration on the X axis at a given moment, Indicates the joint point The acceleration on the Y axis at that moment, Indicates the joint point The acceleration on the Z axis at a given moment is, The total number of moments in the time period.

[0022] Furthermore, S3 includes the following sub-steps:

[0023] S31, inputting the fused posture set into the autoencoder to obtain effective posture feature representation;

[0024] S32: Input the effective posture feature representation into the classifier to generate the user's action posture evaluation result.

[0025] The beneficial effect of the above further scheme is: in the present invention, the autoencoder is usually composed of two parts: an encoder and a decoder. The encoder compresses the input data (the fused posture set) into a low-dimensional representation (encoding), and the decoder attempts to reconstruct the original input from the encoding. The present invention can use an unsupervised learning method to train the autoencoder so that it can learn the effective feature representation of the data. The effective posture feature representation obtained by the autoencoder in the present invention can be a vector. The classifier can be adjusted and optimized according to the fused posture set to complete the action evaluation and improve the accuracy of the evaluation.

[0026] Furthermore, in S32, the classifier The expression is:

[0027] ;

[0028] In the formula, represents the F2 norm operation, represents the effective posture feature representation, represents the penalty slack variable, represents the exponential function, represents the sum of all elements in the fusion posture set, Represents the mean of all elements in the fused pose set.

[0029] The beneficial effects of the present invention are as follows: the present invention proposes a motion evaluation method based on deep learning, which depicts the posture change curve of the acceleration data of the user in a three-dimensional space over a period of time, and determines the fused posture set of the user's joints over the entire time; then uses a classifier involving deep learning technology to classify and evaluate the fused posture set to obtain a final evaluation result; in terms of data processing, the present invention fully integrates the posture data set to improve the reliability of data processing; and the classifier can obtain accurate evaluation results based on the fused posture set, improve the accuracy of the evaluation results, and provide users with more valuable reference action guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Flowchart of the deep learning-based action evaluation method. DETAILED DESCRIPTION

[0031] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0032] like Figure 1 As shown, the present invention provides an action evaluation method based on deep learning, comprising the following steps:

[0033] S1, using a three-axis accelerometer to obtain a user's posture data set within a time period;

[0034] S2, using the user's posture data set within the time period, calculate the original posture weights of the joint points to obtain a fused posture set;

[0035] S3. Generate the user's action posture evaluation result by using the fused posture set.

[0036] In the embodiment of the present invention, in S1, the posture data set includes the acceleration of each joint point of the user in the X-axis, the acceleration of the Y-axis and the acceleration of the Z-axis in the time period.

[0037] In this embodiment of the present invention, S2 includes the following sub-steps:

[0038] S21, determining the posture components of the joint points at each moment in the time period according to the acceleration of each joint point of the user on the X-axis, the acceleration on the Y-axis, and the acceleration on the Z-axis in the time period;

[0039] S22, constructing a posture change curve for each joint point with the time period as the horizontal axis and the posture component as the vertical axis;

[0040] S23, determining the original posture weight of the joint point according to the posture change curve of the joint point;

[0041] S24. According to the original posture weights of the joint points, the posture data sets of the joint points within the time period are fused to generate fused posture values ​​of the joint points, thereby obtaining a fused posture set.

[0042] In the present invention, each joint point has accelerations in different directions in three-dimensional space, and the accelerations of the joint points in different directions are characterized by curves, and the characteristics of the accelerations in different directions are determined to form the original posture weights. The acceleration data of the X-axis, Y-axis and Z-axis are fused through the original posture weights, and the overall motion state of the user's joint points in three-dimensional space can be comprehensively considered, so as to obtain more accurate and comprehensive motion information. This fusion process can reduce the errors and uncertainties that may be caused by single-axis acceleration data, and improve the accuracy and reliability of data processing.

[0043] In the embodiment of the present invention, in S21, the joint point is The weight of the posture at each moment The calculation formula is:

[0044] ;

[0045] In the formula, Indicates the joint point The acceleration on the X axis at a given moment, Indicates the joint point The acceleration on the Y axis at that moment, Indicates the joint point The acceleration on the Z axis at that moment.

[0046] In the embodiment of the present invention, in S23, the number of zero-crossing points in the posture change curve is used as the original posture weight of the joint point.

[0047] In the embodiment of the present invention, in S24, the fusion posture value of the joint point The calculation formula is:

[0048] ;

[0049] In the formula, Represents the original posture weight of the joint point, Indicates the joint point The acceleration on the X axis at a given moment, Indicates the joint point The acceleration on the Y axis at that moment, Indicates the joint point The acceleration on the Z axis at a given moment is, The total number of moments in the time period.

[0050] In this embodiment of the present invention, S3 includes the following sub-steps:

[0051] S31, inputting the fused posture set into the autoencoder to obtain effective posture feature representation;

[0052] S32: Input the effective posture feature representation into the classifier to generate the user's action posture evaluation result.

[0053] In the present invention, the autoencoder is usually composed of two parts: an encoder and a decoder. The encoder compresses the input data (fusion posture set) into a low-dimensional representation (encoding), and the decoder attempts to reconstruct the original input from the encoding. The present invention can use an unsupervised learning method to train the autoencoder so that it can learn the effective feature representation of the data. The effective posture feature representation obtained by the autoencoder in the present invention can be a vector. The classifier can be adjusted and optimized according to the fusion posture set to complete the action evaluation and improve the accuracy of the evaluation.

[0054] In the embodiment of the present invention, in S32, the classifier The expression is:

[0055] ;

[0056] In the formula, represents the F2 norm operation, represents the effective posture feature representation, represents the penalty slack variable, represents the exponential function, represents the sum of all elements in the fusion posture set, Represents the mean of all elements in the fused pose set.

[0057] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A motion evaluation method based on deep learning, characterized in that: The following steps are involved: S1, using a three-axis accelerometer to obtain a user's posture data set within a time period; S2, using the user's posture data set within the time period, calculate the original posture weights of the joint points to obtain a fused posture set; S3, using the fused posture set to generate the user's action posture evaluation result; The S2 comprises the following sub-steps: S21, determining the posture components of the joint points at each moment in the time period according to the acceleration of each joint point of the user on the X-axis, the acceleration on the Y-axis, and the acceleration on the Z-axis in the time period; S22, constructing a posture change curve for each joint point with the time period as the horizontal axis and the posture component as the vertical axis; S23, determining the original posture weight of the joint point according to the posture change curve of the joint point; S24, fusing the posture data sets of the joint points within the time period according to the original posture weights of the joint points, generating fused posture values ​​of the joint points, and obtaining a fused posture set; In S24, the fusion posture value of the joint point The calculation formula is: ； In the formula, Represents the original posture weight of the joint point, Indicates the joint point The acceleration on the X axis at a given moment, Indicates the joint point The acceleration on the Y axis at that moment, Indicates the joint point The acceleration on the Z axis at a given moment is, The total number of moments representing the time period; The S3 comprises the following sub-steps: S31, inputting the fused posture set into the autoencoder to obtain effective posture feature representation; S32, inputting the effective posture feature representation into the classifier to generate the user's action posture evaluation result; In S32, the classifier The expression is: ； In the formula, represents the F2 norm operation, represents the effective posture feature representation, represents the penalty slack variable, represents the exponential function, represents the sum of all elements in the fusion posture set, Represents the mean of all elements in the fused pose set.

2. The action evaluation method based on deep learning according to claim 1, characterized in that: In S1, the posture data set includes the acceleration of each joint point of the user in the X-axis, the acceleration of the Y-axis and the acceleration of the Z-axis in the time period.

3. The action evaluation method based on deep learning according to claim 1, characterized in that: In S21, the joint point is within the time period The weight of the posture at each moment The calculation formula is: ； In the formula, Indicates the joint point The acceleration on the X axis at a given moment, Indicates the joint point The acceleration on the Y axis at that moment, Indicates the joint point The acceleration on the Z axis at that moment.

4. The action evaluation method based on deep learning according to claim 1, characterized in that: In S23, the number of zero-crossing points in the posture change curve is used as the original posture weight of the joint point.

Images