A posture comparison method and terminal
By identifying the human body frame and key points, calculating the action angle, and comparing the similarity, the problem of insufficient accuracy of posture recognition in the existing technology is solved, and more accurate posture recognition and action evaluation is achieved, reducing guidance costs.
Patent Information
- Application Number
- CN202011074689.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-10-09
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2040-10-09
AI Technical Summary
The existing posture recognition comparison technology is based on the position comparison of key points, and cannot accurately judge the angle of the action, resulting in insufficient accuracy of posture recognition.
By acquiring image data, identifying the human frame and key points, calculating the angle of the action, and comparing it with standard action data or other action data, the accuracy of action angles is performed using dynamic time regularization algorithms and deep learning convolutional neural algorithms. calculate.
It improves the accuracy of posture recognition, can more accurately assist users in adjusting postures, judging the standardization and completion of actions, reduces guidance costs, and adapts to the needs of multiple people and local actions.
Smart Images
Figure CN112200074B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image recognition, and in particular to a posture comparison method and terminal. Background Art
[0002] At present, human visual posture recognition and comparison mainly involves computers obtaining image data of user actions, identifying user postures through different methods, and then comparing them with preset standard action data to determine the standardization of the action. Most existing comparison technologies are based on the comparison of key points. By comparing the positions of key points and calculating the position offset, the accuracy of the action is determined. However, some actions have requirements on the angle of the action. Therefore, the judgment of the accuracy of the action based on key points is not accurate enough, and the accuracy of posture recognition cannot be guaranteed. Summary of the invention
[0003] The technical problem to be solved by the present invention is to provide a posture comparison method and a terminal to improve the accuracy of posture recognition.
[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0005] A posture comparison method comprises the steps of:
[0006] S1, obtaining image data;
[0007] S2, performing posture recognition on the first object in the image data to obtain the human body frame and key points of the first object;
[0008] S3. Calculate the first object motion data according to the human body frame and key points of the first object, compare the first object motion data with other object motion data for similarity to obtain motion similarity, and obtain a posture comparison result according to the motion similarity. The first object motion data and the other object motion data respectively include object motion angles.
[0009] In order to solve the above technical problems, another technical solution adopted by the present invention is:
[0010] A posture comparison terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the following steps when executing the computer program:
[0011] S1, obtaining image data;
[0012] S2, performing posture recognition on the first object in the image data to obtain the human body frame and key points of the first object;
[0013] S3. Calculate the first object motion data according to the human body frame and key points of the first object, compare the first object motion data with other object motion data for similarity to obtain motion similarity, and obtain a posture comparison result according to the motion similarity. The first object motion data and the other object motion data respectively include object motion angles.
[0014] The beneficial effects of the present invention are: a posture comparison method and terminal, the present invention detects the posture of a person in real time, and calculates the angle of the action based on the human body frame and key points. The angle calculation can more accurately assist the user to adjust his posture, and can quickly and effectively help the user judge the standardization and completion of the action, thereby achieving the effect of virtual guidance and greatly reducing the cost of guiding the action. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 A flowchart of a posture comparison method according to an embodiment of the present invention;
[0016] Figure 2 A schematic diagram of a posture comparison method according to an embodiment of the present invention;
[0017] Figure 3 This is a schematic diagram of the structure of a terminal for posture comparison according to an embodiment of the present invention;
[0018] Description of labels:
[0019] 1. A posture comparison terminal; 2. A processor; 3. A memory. DETAILED DESCRIPTION
[0020] In order to explain the technical content, achieved objectives and effects of the present invention in detail, the following is an explanation in combination with the implementation modes and the accompanying drawings.
[0021] Please refer to Figure 1 as well as Figure 2 , a posture comparison method, comprising the steps of:
[0022] S1, obtaining image data;
[0023] S2, performing posture recognition on the first object in the image data to obtain the human body frame and key points of the first object;
[0024] S3. Calculate the first object motion data according to the human body frame and key points of the first object, compare the first object motion data with other object motion data for similarity to obtain motion similarity, and obtain a posture comparison result according to the motion similarity. The first object motion data and the other object motion data respectively include object motion angles.
[0025] From the above description, it can be seen that the beneficial effects of the present invention are: the present invention calculates the angle of the action based on the human body frame and key points, and judges the accuracy of the action by comparing the action angles, thereby improving the accuracy of the action comparison.
[0026] Furthermore, the step S3 of comparing the first object action data with other object action data for similarity specifically includes the following steps:
[0027] Comparing the first object action data with the standard action data of the standard object for similarity;
[0028] Alternatively, the first object action data is compared with other object action data in the same image data for similarity.
[0029] It can be seen from the above description that the similarity comparison of the present invention includes comparison with standard action data of a standard object and action comparison between multiple objects in the same image data, which increases the application scope of the present invention.
[0030] Furthermore, performing similarity comparison between the first object action data and the standard action data of the standard object specifically comprises the following steps:
[0031] Obtain standard action data pre-selected by a user, and compare the first object action data of each action of the first object with the standard action data one by one for similarity, the first object action data and the standard action data including the corresponding maximum angle, minimum angle, average angle, action rate, time of each action and number of times of a group of actions.
[0032] From the above description, it can be seen that when comparing with the standard action data of the standard object, the present invention improves the accuracy of the comparison by comparing the maximum angle, the minimum angle, the average angle, the action rate, the time of each action and the number of times of a group of actions.
[0033] Furthermore, performing similarity comparison between the first object action data and the standard action data of the standard object specifically comprises the following steps:
[0034] Automatically identify important parts for different actions based on the standard action data selected by the user;
[0035] Obtain standard action data pre-selected by a user, and compare the first object local action data corresponding to the important part in the first object action data of each action of the first object with the standard local action data corresponding to the important part in the standard action data for similarity, the first object action data and the standard action data including corresponding maximum angles, minimum angles, average angles, action rates, time for each action, and the number of times for a group of actions.
[0036] From the above description, it can be seen that the comparison with the standard action data of the standard object also includes the comparison of local actions, which can better meet the needs of users and enable users to better correct their actions.
[0037] Furthermore, the comparing the similarity between the first object action data and other object action data in the same image data specifically comprises the following steps:
[0038] Comparing the first object motion angle in the first object motion data with other object motion angles of each other object in the same image data frame by frame;
[0039] The obtaining of the action similarity and obtaining the posture comparison result according to the action similarity specifically comprises the following steps:
[0040] Obtaining a difference value of an action angle between the first object and each other object in each frame, and performing a similarity evaluation on the difference value of the action angle between the first object and each other object in each frame by a preset metric method to obtain action similarity of the objects, wherein the preset metric method includes an average value and / or a variance and / or a maximum similarity;
[0041] Comparing the action angles of each object in the image data one by one and frame by frame in turn until the action similarity of each object in the image data is obtained;
[0042] The action similarities of each object are compared, and objects whose action similarity differences are greater than a preset threshold are marked and displayed.
[0043] It can be seen from the above description that when performing multi-person comparison of unified image data, the present invention will compare the objects in the video frame by frame, accurately identify objects with large movement differences and mark and display them, so that users can make corrections in time and improve the uniformity of movements.
[0044] Furthermore, the similarity of the first object action data and the standard action data of the standard object is compared to obtain the action similarity as follows:
[0045] S4, converting the standard action and the first object action data into a matrix form as the input of DTW, calculating the shortest path value, and determining the similarity of the actions, specifically:
[0046] Standard action data is converted into matrix M k , the first object action data is converted into matrix N j , when k≠j, use DTW to regularize the matrices M and N so that the dimensions or number of sequences between the two templates are the same, and then calculate their distance;
[0047] Calculate the distance between each m and n to form an m-row and n-column matrix p(m,n), and use s(j,k) to represent the distance from p(0,0) to p(j,k). The calculation formula of s(j,k) is as follows:
[0048] s(j,k)=min{[s(j-1,k)+d(j,k)],[s(j,k-1)+d(j,k)],[s(j-1,k-1)+d(j,k)]};
[0049] Based on this, the shortest path value D(M,N) can be calculated. The smaller the shortest path value is, the higher the action similarity is.
[0050] From the above description, it can be seen that the present invention can also use a dynamic time warping algorithm when performing motion comparison, which has higher accuracy, so that the motion comparison of the present invention can adapt to more situations.
[0051] Further, S21, the image data is segmented into each frame of the image, and a preliminary morphological analysis is performed on the first object in the current image data to determine the length of each limb of the first object. Through the deep learning convolutional layer neural algorithm, two key points are first determined, and then the position of the third point is determined according to the angle data of the standard action data. This step is repeated until all key points are drawn, and the angles of the key parts are calculated. The angle calculation formula of the key parts according to the key nodes is as follows:
[0052]
[0053] The determining of the position of the third point is specifically as follows:
[0054] Calculate the linear equation of the straight line where the two determined key points are located, and determine a temporary point according to the linear equation and the actual limb position corresponding to the key point, wherein the length between the temporary point and the adjacent known key point is equal to the limb length corresponding to the actual limb position;
[0055] According to the actual limb position, this temporary point on the straight line is rotated around the adjacent known key point by the angle data of the corresponding standard action data to obtain the coordinates of the third point.
[0056] From the above description, it can be seen that the present invention determines the position of key points based on the convolutional neural algorithm and calculates the action angle, making the recognition and comparison of actions more accurate.
[0057] Furthermore, the convolutional neural network algorithm is as follows:
[0058] a. Convolutional layer:
[0059]
[0060] W1 is the width of the image before convolution; W2 is the width of the Feature Map after convolution; H1 is the height of the image before convolution; H2 is the height of the Feature Map after convolution; P is the amount of padding; S is the stride step size.
[0061] b. Activation function: ReLU function
[0062] φ(x)=max(0,x)
[0063] c. The pooling layer includes mean pooling and maximum pooling.
[0064] d. The fully connected layer consists of multiple neurons, and the calculation formula for each neuron is as follows:
[0065]
[0066] Furthermore, before step S2, the method further includes:
[0067] S11. For the acquired image data, the position information of the key nodes circled by the user is stored in a database for use in model training.
[0068] S12, preprocessing: scaling the collected image data to keep all image data of the same size; then normalizing the data to make it fall into a small numerical range of -1 to +1, or 0 to 1, and then processing the image data to improve the image data quality.
[0069] From the above description, it can be seen that the present invention can also accept the user's annotation of key points before performing posture recognition of the object, so as to better meet the user's needs, and also pre-process the image data to improve the quality of the image data.
[0070] Please refer to Figure 2 as well as Figure 3 , a posture comparison terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor,
[0071] When the processor executes the computer program, the posture comparison method as described above is implemented.
[0072] Please refer to Figure 1 and Figure 2 , Embodiment 1 of the present invention is:
[0073] A posture comparison method comprises the steps of:
[0074] S1, obtaining image data;
[0075] like Figure 2 As shown, the image data can be collected by a camera or obtained through data transmission. In addition to the image data, the user's heart rate, location, fingerprint, account password and other information can also be obtained through smart wearable devices according to functional requirements;
[0076] After acquiring the image data in step S1, the following steps are also included:
[0077] S11. For the acquired image data, the position information of the key nodes circled by the user is stored in the database for model training. Figure 2 As shown, the user uses the annotation tool to mark the joints, bones and other locations, and the system collects the user's annotated information and saves it to the database.
[0078] When analyzing and processing data, Figure 1 As shown, it also includes identity recognition and verification, including fingerprint, face, and account password verification.
[0079] S12, preprocessing: scaling the collected image data to keep all image data of the same size; then normalizing the data to make it fall into a small numerical range of -1 to +1, or 0 to 1, and then processing the image data to improve the image data quality.
[0080] S2, performing posture recognition on the first object in the image data to obtain a human frame and key points of the first object;
[0081] The step S2 of performing posture recognition on the first object in the image data is specifically as follows:
[0082] S21, segment the image data into each frame, perform a preliminary morphological analysis on the first object in the current image data, determine the length of each limb of the first object, first determine two key points through the deep learning convolutional layer neural algorithm, and then determine the position of the third point based on the angle data of the standard action data, repeat this step until all key points are drawn, and calculate the angle of the key parts. The angle calculation formula of the key parts based on the key nodes is as follows:
[0083]
[0084] Among them, the position of the key node is expressed as follows:
[0085] BodyParti:(x i ,y i ,score i ), i is the serial number of the key node, x i is the horizontal coordinate position, y i is the vertical coordinate position, scorei is the credibility of the node at this location.
[0086] The determining of the position of the third point is specifically as follows:
[0087] Calculate the linear equation of the straight line where the two determined key points are located, and determine a temporary point according to the linear equation and the actual limb position corresponding to the key point, wherein the length between the temporary point and the adjacent known key point is equal to the limb length corresponding to the actual limb position;
[0088] According to the actual limb position, this temporary point on the straight line is rotated around the adjacent known key point by the angle data of the corresponding standard action data to obtain the coordinates of the third point.
[0089] Among them, the convolutional neural network algorithm is as follows:
[0090] a. Convolutional layer:
[0091] W2=(W1-F+2P) / S+1
[0092] H2=(H1-F+2P) / S+1
[0093] W1 is the width of the image before convolution; W2 is the width of the Feature Map after convolution; H1 is the height of the image before convolution; H2 is the height of the Feature Map after convolution; P is the amount of padding; S is the stride step size.
[0094] b. Activation function: ReLU function
[0095] φ(x)=max(0,x)
[0096] c. The pooling layer includes mean pooling and maximum pooling.
[0097] d. The fully connected layer consists of multiple neurons, and the calculation formula for each neuron is as follows:
[0098]
[0099] like Figure 1 As shown, when performing gesture recognition on image data, the following content is output in the video:
[0100] (1) Frames per second (FPS)
[0101] (2) Current frame number
[0102] (3) The number of human bodies detected in the current image
[0103] (4) For each human body being identified, mark the key nodes of the human body, including the following parts:
[0104] 1) Overall posture nodes, including but not limited to left eye, right eye, left ear, right ear, nose, neck, left shoulder, right shoulder, left elbow, right elbow, left hand, right hand, left waist, right waist, left knee, right knee, left foot, right foot, etc.
[0105] 2) Facial pose nodes.
[0106] 3) Hand posture node.
[0107] S3. Calculate the first object motion data according to the human body frame and key points of the first object, compare the first object motion data with other object motion data for similarity to obtain motion similarity, and obtain a posture comparison result according to the motion similarity. The first object motion data and the other object motion data respectively include object motion angles.
[0108] The step S3 of comparing the first object action data with other object action data for similarity specifically includes the following steps:
[0109] Comparing the first object action data with the standard action data of the standard object for similarity;
[0110] Alternatively, the first object action data is compared with other object action data in the same image data for similarity.
[0111] Comparing the similarity between the first object action data and the standard action data of the standard object specifically comprises the following steps:
[0112] Obtain standard action data pre-selected by a user, and compare the first object action data of each action of the first object with the standard action data one by one for similarity, the first object action data and the standard action data including the corresponding maximum angle, minimum angle, average angle, action rate, time of each action and number of times of a group of actions.
[0113] The similarity comparison of the action rates is specifically as follows:
[0114] According to the data range of the maximum angle and the minimum angle of the first object action data, the corresponding standard action data is taken. For example, if the standard action angle range is 0°-180°, and the minimum angle of the user action is 5° and the maximum angle is 170°, then the data within the range of 5°-170° is taken. The angle within the maximum angle and the minimum angle range is divided into n equal parts, and the angles after equal division are used as the reference to perform action trend analysis and similarity calculation on each frame of the first object action data and the standard action data.
[0115] The motion trend analysis and similarity calculation for each frame of the first object motion data and the standard motion data are specifically as follows: obtaining the images of the previous and next frames and the motion angle data of the previous and next frames, comparing them with the motion angle corresponding to the current frame, determining whether the trend is increasing or decreasing, obtaining the moment of the current frame, calculating the time intervals between adjacent angles, comparing and quantifying the differences in the time intervals to obtain the similarity of the motion rates.
[0116] Based on the output results obtained above, we can not only get the overall action rate difference, but also the rate difference of the decomposed action. The larger n is, the more detailed the action decomposition is, but at the same time the amount of calculation is greater.
[0117] The comparing the similarity between the first object action data and other object action data in the same image data specifically comprises the following steps:
[0118] Comparing the first object motion angle in the first object motion data with other object motion angles of each other object in the same image data frame by frame;
[0119] The obtaining of the action similarity and obtaining the posture comparison result according to the action similarity specifically comprises the following steps:
[0120] Obtaining a difference value of an action angle between the first object and each other object in each frame, and performing a similarity evaluation on the difference value of the action angle between the first object and each other object in each frame by a preset metric method to obtain action similarity of the objects, wherein the preset metric method includes an average value and / or a variance and / or a maximum similarity;
[0121] Comparing the action angles of each object in the image data one by one and frame by frame in turn until the action similarity of each object in the image data is obtained;
[0122] The action similarities of each object are compared, and objects whose action similarity differences are greater than a preset threshold are marked and displayed.
[0123] Please refer to Figure 1 and Figure 2 , Embodiment 2 of the present invention is:
[0124] A posture comparison method is different from the first embodiment in that the similarity comparison between the first object action data and the standard action data of the standard object specifically includes the following steps:
[0125] Automatically identify important parts for different actions based on the standard action data selected by the user;
[0126] Obtain standard action data pre-selected by a user, and compare the first object local action data corresponding to the important part in the first object action data of each action of the first object with the standard local action data corresponding to the important part in the standard action data for similarity, the first object action data and the standard action data including corresponding maximum angles, minimum angles, average angles, action rates, time for each action, and the number of times for a group of actions.
[0127] Please refer to Figure 1 and Figure 2 , Embodiment 3 of the present invention is:
[0128] A posture comparison method is different from the first embodiment in that the similarity comparison between the first object action data and the standard action data of the standard object is specifically performed as follows:
[0129] S4, converting the standard action and the first object action data into a matrix form as the input of DTW, calculating the shortest path value, and determining the similarity of the actions, specifically:
[0130] Standard action data is converted into matrix M k , the first object action data is converted into matrix N j , when k≠j, use DTW to regularize the matrices M and N so that the dimensions or number of sequences between the two templates are the same, and then calculate their distance;
[0131] Calculate the distance between each m and n to form an m-row and n-column matrix p(m,n), and use s(j,k) to represent the distance from p(0,0) to p(j,k). The calculation formula of s(j,k) is as follows:
[0132] s(j,k)=min{[s(j-1,k)+d(j,k)],[s(j,k-1)+d(j,k)],[s(j-1,k-1)+d(j,k)]};
[0133] Based on this, the shortest path value D(M,N) can be calculated. The smaller the shortest path value is, the higher the action similarity is.
[0134] Please refer to Figure 2 and Figure 3 , Embodiment 4 of the present invention is:
[0135] A posture comparison terminal 1 includes a memory 3, a processor 2, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above posture comparison method is implemented.
[0136] like Figure 2As shown in , if a terminal for posture comparison is divided into modules, such as Figure 2 As shown, it is divided into a data acquisition module, a data annotation module and a data processing module. The data processing module collects user data; the data annotation module provides annotation tools, receives the joint points and bone position information annotated by the user, and saves it to the database; the preprocessing of the data processing module corresponds to the preprocessing of step S12, the posture recognition corresponds to step S2 to perform posture recognition on the first object in the image data, and the analysis and comparison corresponds to step S3 to calculate the first object action data according to the human body frame and key points of the first object, compare the first object action data with other object action data for similarity, obtain action similarity, and obtain the posture comparison result according to the action similarity.
[0137] In summary, the present invention provides a posture comparison method and terminal, which detects the posture of a person in real time and calculates the angle of the action based on the human body frame and key points. The calculation of the action angle can more accurately assist the user to adjust his posture, and can quickly and effectively help the user judge the degree of standardization and completion of the action, thereby achieving the effect of virtual guidance and greatly reducing the cost of guiding the action. The present invention also provides overall action comparison and local action comparison with standard action data and comparison of multiple objects in the same image data, which can adapt to the different usage needs of users and make the guidance effect more outstanding.
[0138] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's specification and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A method for posture comparison, It is characterized in that Includes steps: S1, obtaining image data; S2, performing posture recognition on the first object in the image data to obtain a human frame and key points of the first object; The step S2 of performing posture recognition on the first object in the image data is specifically as follows: S21, segment the image data into each frame image, perform a preliminary morphological analysis on the first object in the current image data, determine the length of each limb of the first object, first determine two key points through the deep learning convolutional neural network algorithm, and then determine the position of the third point according to the angle data of the standard action data, repeat this step until all key points are drawn, and calculate the angle of the key parts. The angle calculation formula of the key parts according to the key nodes is as follows: ， The determining of the position of the third point is specifically as follows: Calculate the linear equation of the straight line where the two determined key points are located, and determine a temporary point according to the linear equation and the actual limb position corresponding to the key point, wherein the length between the temporary point and the adjacent known key point is equal to the limb length corresponding to the actual limb position; According to the actual limb position, the temporary point on the straight line is rotated around the adjacent known key point by the angle data of the corresponding standard action data to obtain the coordinates of the third point; S3. Calculate the first object motion data according to the human body frame and key points of the first object, compare the first object motion data with other object motion data for similarity to obtain motion similarity, and obtain a posture comparison result according to the motion similarity. The first object motion data and the other object motion data respectively include object motion angles.
2. A posture comparison method according to claim 1, It is characterized in that The comparing the similarity between the first object action data and other object action data specifically comprises the following steps: Comparing the first object action data with the standard action data of the standard object for similarity; Alternatively, the first object action data is compared with other object action data in the same image data for similarity.
3. A posture comparison method according to claim 2, It is characterized in that Comparing the similarity between the first object action data and the standard action data of the standard object specifically comprises the following steps: Obtain standard action data pre-selected by a user, and compare the first object action data of each action of the first object with the standard action data one by one for similarity, the first object action data and the standard action data including the corresponding maximum angle, minimum angle, average angle, action rate, time of each action and number of times of a group of actions.
4. A posture comparison method according to claim 2, It is characterized in that Comparing the similarity between the first object action data and the standard action data of the standard object specifically comprises the following steps: Automatically identify important parts for different actions based on the standard action data selected by the user; Obtain standard action data pre-selected by a user, and compare the first object local action data corresponding to the important part in the first object action data of each action of the first object with the standard local action data corresponding to the important part in the standard action data for similarity, the first object action data and the standard action data including corresponding maximum angles, minimum angles, average angles, action rates, time for each action, and the number of times for a group of actions.
5. A posture comparison method according to claim 2, It is characterized in that The comparing the similarity between the first object action data and other object action data in the same image data specifically comprises the following steps: Comparing the first object motion angle in the first object motion data with other object motion angles of each other object in the same image data frame by frame; The obtaining of the action similarity and obtaining the posture comparison result according to the action similarity specifically comprises the following steps: Obtaining a difference value of an action angle between the first object and each other object in each frame, and performing a similarity evaluation on the difference value of the action angle between the first object and each other object in each frame by a preset metric method to obtain action similarity of the objects, wherein the preset metric method includes an average value and / or a variance and / or a maximum similarity; Comparing the action angles of each object in the image data one by one and frame by frame in turn until the action similarity of each object in the image data is obtained; The action similarities of each object are compared, and objects whose action similarity differences are greater than a preset threshold are marked and displayed.
6. A posture comparison method according to claim 2, It is characterized in that The action data of the first object is compared with the standard action data of the standard object for similarity, and the action similarity obtained is specifically: S4, converting the standard action and the first object action data into a matrix form as the input of DTW, calculating the shortest path value, and determining the similarity of the actions, specifically: Standard action data converted into matrix , the first object action data is converted into a matrix ,when When , DTW is used to regularize the matrices M and N so that the dimensions or number of sequences between the two templates are the same, and then the distance is calculated; Calculate the distance between each m and n to form an m-row and n-column matrix p(m,n), and use s(j,k) to represent the distance from p(0,0) to p(j,k). The calculation formula of s(j,k) is as follows: ； Based on this, the shortest path value D(M,N) can be calculated. The smaller the shortest path value is, the higher the action similarity is.
7. A posture comparison method according to claim 1, It is characterized in that The convolutional neural network algorithm is as follows: a. Convolutional layer: is the width of the image before convolution; It is the width of the Feature Map after convolution; is the height of the image before convolution; It is the height of the Feature Map after convolution; is the padding amount; is the stride step length; b. Activation function: ReLU function c. The pooling layer includes mean pooling and maximum pooling; d. The fully connected layer consists of multiple neurons, and the calculation formula for each neuron is as follows: 。 8. A posture comparison method according to claim 1, It is characterized in that Before step S2, the method further includes: S11, for the acquired image data, storing the position information of the key nodes circled by the user into a database for model training; S12, preprocessing: scaling the collected image data to keep all image data of the same size; then normalizing the data to make it fall into a small numerical range of -1 to +1, or 0 to 1, and then processing the image data to improve the image data quality.
9. A posture comparison terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, It is characterized in that When the processor executes the computer program, the method for posture comparison as described in any one of claims 1-8 is implemented.