A sports movement action analysis and correction method and system based on blockchain

Abstract

The embodiment of the application discloses a sports movement action analysis and correction method fused with a blockchain, which comprises the following steps: recording user behavior records by using a Fabric blockchain network in the system, authorizing a cloud database to distribute corresponding follow-up courses to users for follow-up by using an access control mechanism in the Fabric blockchain network, finally giving corresponding scores, using a joint feature real-time system model PostEX for extracting a skeleton graph of a sports video, an action sequence similarity calculation algorithm and an action comparison scoring algorithm. The embodiment of the application also discloses a sports movement action analysis and correction system fused with a blockchain. According to the application, the action recognition and scoring have good practicability and robustness, the calculation amount can be reduced, and the problem of incorrect postures of sports lovers when learning related sports can be solved and corrected.

Description

TECHNICAL FIELD

[0001] The present application relates to image recognition and processing in the field of sports, and in particular to a sports action analysis and correction method and system based on a Fabric blockchain network and joint feature cosine similarity. BACKGROUND

[0002] With the improvement of economic and cultural level, people's spiritual and cultural consumption consciousness is gradually enhanced, and more attention is paid to the improvement of physical quality, and people begin to have higher pursuit of sports.

[0003] However, many sports enthusiasts suffer from the lack of one-on-one coaching from a dedicated coach, and can only learn from online learning resources and learn a series of sports actions, and the wrong actions in the learning process are difficult to be discovered and corrected in time.

[0004] The field of human motion analysis has always been an important field of concern for scholars. The evaluation task of human action similarity is an important part of human motion analysis. With the rapid development of machine learning and deep learning, scholars have proposed many algorithms that can be applied to the evaluation task of human action similarity.

[0005] A fitness action correction method and device, computer equipment and storage medium are disclosed in Chinese Patent Publication No. CN112464918A. The method collects image data of the target user at different angles on a treadmill, fuses the image data of the target user at different angles, obtains a three-dimensional fitness action image of the target user, determines a target fitness action standard image that matches the three-dimensional fitness action image of the target user based on a pre-set fitness action standard image library, and then obtains a score of the three-dimensional fitness action image of the target user according to the key point features of the target fitness action standard image. If the score is lower than the target score, the correction information of the three-dimensional fitness action image of the target user is output to correct the non-standard fitness action of the user in time, and the whole process is user-unaware. However, the three-dimensional fitness action is obtained, which consumes a lot of calculation, and the practicality and robustness of action recognition and scoring are not ideal. In addition, the sports actions in the prior art do not use blockchain technology for storage, and there are short boards in data security and data management. SUMMARY

[0006] The technical problem to be solved by the embodiments of the present application is to provide a sports action analysis and correction method and system integrating blockchain, which can solve and correct the problem of incorrect posture of sports enthusiasts when learning related sports with lower computational overhead, and use blockchain technology to save and follow up the data.

[0007] To solve the above technical problems, the embodiment of the application provides a sports movement action analysis and correction method based on a Fabric block chain network and a joint feature cosine similarity algorithm, which comprises the following steps:

[0008] S1: constructing a standard sports action video sequence teaching material library in a cloud database,

[0009] S2: acquiring a purchase record of a corresponding sports movement action video stored in the Fabric block chain network by the Fabric block chain network, authorizing the cloud database, distributing the corresponding purchased sports teaching video to the user, and using a camera to shoot and save the whole follow-up training process of the user as a to-be-corrected sports action video, and after the user completes the training, the to-be-corrected sports action video is transmitted into the system for processing;

[0010] S3: preprocessing the to-be-corrected sports action video sequence, so that each frame of the video is matched with each frame in the corresponding standard action video sequence in the standard action video material library;

[0011] S4: inputting the preprocessed video sequence, using a joint feature real-time system model PostEX to perform real-time calculation, outputting and extracting limb key point coordinates of each frame of the sports action video sequence, and outputting in the form of a three-dimensional heat map when the user sets to the accurate mode;

[0012] S5: using a feature extraction algorithm of a human body limb angle cosine value to convert the extracted limb key point coordinate value feature into a limb vector value feature, and then into a limb cosine similarity feature value; the action representation after extraction is used to search for a sports action with the closest vector value cosine similarity in the sports action material video library, so as to identify the specific sports action name made by the user, and make the to-be-corrected sports action video correspond to the standard sports action video;

[0013] S6: using a scoring and correction algorithm to compare the to-be-corrected sports action skeleton graph with the standard sports action skeleton graph, providing corresponding scores and action correction suggestions, using an interactive interface to return the matched feedback result to the user, and providing difference details and specific action detailed guidance words;

[0014] S7: encrypting the score record and storing it in the Fabric block chain network in the form of a block, and according to a pre-written smart contract, if the score reaches a certain standard, the block chain network authorizes the relevant user to obtain a corresponding new sports course follow-up video, and continues to follow up and learn.

[0015] The on-chain step of the Fabric block chain network comprises:

[0016] S21: When the to-be-verified block in a node in the Fabric blockchain network has been written full, the verification block is submitted to the Fabric blockchain network and sent to all nodes;

[0017] S22: After receiving the verification block, each node in the Fabric blockchain network verifies the block, including checking whether the block information meets the rules and standards in the network and whether the initiator has sufficient authority and qualification to submit the block;

[0018] S23: The node that passes the verification endorses the to-be-verified block. In the Fabric blockchain network, it is stipulated that at least 50% of the nodes endorse the block to confirm the validity of the block;

[0019] S24: Once the endorsement threshold is reached, the Fabric blockchain network marks the block as confirmed and stores the block in the Fabric blockchain network, while the nodes update the distributed ledger and permanently save it in the Fabric blockchain network.

[0020] The camera includes a monocular camera and an infrared sensor. The monocular camera is used for video recording and saving. The infrared sensor is used for real-time follow-up practice and disassembly of actions.

[0021] The pre-processing of the to-be-corrected motion action video sequence includes: using a sliding window method to obtain a video sequence with the same number of frames as the video library, converting the original RGB frame to a gray-scale representation, using a spatial enhancement method to remove the background, and using an average filter to reduce the generation of random noise.

[0022] The joint feature real-time system model PostEX includes:

[0023] The pre-processed video sequence is first subjected to feature preprocessing of the first 10 layers of the VGG19 network, and then the video sequence is converted into image features F. The image features F are then divided into two branches to predict the key point confidence and affinity vector of each point. S is the confidence network, and L is the affinity vector field network. , After predicting the confidence of each joint point, non-maximum suppression is used to detect the joint points. Then, the affinity vectors between the detected joint points are subjected to line integration to obtain the affinity between the joint points. Then, the human key points and the corresponding affinity are modeled from the perspective of graph theory. Finally, the Hungarian algorithm is used to obtain the final recognition result of the human skeleton.

[0024] In the precise mode, standardization and normalization are also performed in the pretreatment process, and the normalization formula is: . Wherein, and are the mean and variance of the three-dimensional data obtained within 10ms respectively; is the original three-dimensional coordinate vector, is the result after standard action normalization processing; further comprising the steps of standardizing the skeleton scale and data standard processing, the data standard processing step includes: establishing a human body coordinate system, taking the average value of two skeleton points to obtain the standard center point P c (a,b,c,) of the skeleton, a,b,c are the x,y,z coordinate values of the obtained center point, taking the difference between the 25 skeleton points and the center point, thereby obtaining the human skeleton coordinate point P s after center standardization, obtaining the height D of the human body through the Euclidean distance, dividing the x,y and z of the 25 skeleton points after center standardization by the height D, obtaining the scale standardized skeleton coordinate, and analyzing different parts of the body through training the classifier.

[0025] Wherein, the S5 further comprises a motion sequence similarity measurement method based on joint feature real-time system key point extraction, comprising the following steps:

[0026] S51: The 25 key points extracted after the joint feature real-time system are combined two by two to form limb vector features

[0027] S52: Calculate the included angle cosine value between each two limb vectors to obtain 276 included angle cosine values and form a vector, which is converted into a 276-dimensional cosine similarity vector;

[0028] S53: Feature extraction is performed on each photo of the two groups of photos of the obtained practice motion sequence A and standard motion sequence B, and two groups of vectors A n and B n can be obtained respectively, wherein A n represents the vector representation result of the nth picture of the motion sequence A, B n represents the vector representation result of the nth picture of the motion sequence B, and then all A n are spliced to obtain matrix 1, and all B n are spliced to obtain matrix 2;

[0029] S54: Use a feature selection method for matrix to remove redundant matrices.

[0030] Wherein, the S54 further comprises the following steps:

[0031] S541: using G3D dataset and extracting all action sequences in the dataset into a matrix, and then flattening all the matrices into one-dimensional feature vectors;

[0032] S542: calculating the variance of all features, and removing all features with a variance lower than 0.05;

[0033] S543: using recursive feature elimination method to select features;

[0034] S544: saving the features selected by S541-S543 as an action recognition feature subset.

[0035] The S5 further comprises the following steps:

[0036] S55: first extracting matrix 1 and matrix 2 according to the action recognition feature subset, and the results are vector 1 and vector 2, and then calculating the cosine similarity of vector 1 and vector 2;

[0037] S56: the motion video in the video library has been calculated by the cosine similarity algorithm to obtain the corresponding action matrix.

[0038] The S6 specifically comprises the following steps:

[0039] S61: since A n ={a1,a2,a3,...,a n} and B n ={b1,b2,b3,...,b n} are both n*276 matrices, the similarity algorithm of the A matrix and the B matrix is:

[0040]

[0041] wherein, represents the cosine similarity vector of the i-th frame in the matrix A, represents the cosine similarity vector of the j-th frame in the matrix B, represents the similarity size of the i-th frame in the matrix A and the j-th frame in the matrix B;

[0042] S62: according to the above formula, the is calculated, and the specific score is calculated by the following score formula:

[0043]

[0044] wherein, , are mapping parameters converted from similarity to percentage system, and the corresponding score Score obtained is a percentage score, The matrix A represents the similarity between the i-th frame of the matrix A and the j-th frame of the matrix B.

[0045] The application also provides a system using the above method, comprising:

[0046] A video loading module, after a user purchases a corresponding video, preloads the purchased standard action video sequence in the cloud database through Fabric's relevant resource authorization, for the user to learn and imitate;

[0047] An action learning module, a user practices the relevant sports video, generates an action video sequence to be corrected and uploads it;

[0048] A feature extraction module, which pre-processes the action video sequence to be compared and serves as the input of the joint feature real-time system model PostEX, and through the real-time calculation of the model, outputs and extracts the skeleton limb point two-dimensional graph of each frame of the action video sequence, and in the setting of the system precision mode, more traffic and computing power can be consumed to output the action video sequence as a three-dimensional heat map;

[0049] An action recognition module, which uses the feature extraction algorithm of the human body limb angle cosine value to identify the specific action classification of the movement, then extracts the corresponding standard action video from the video library, so that the action video to be corrected corresponds to the standard action video, and through the joint feature real-time system model PostEX, it is converted into a standard video skeleton limb point two-dimensional graph, which can also be output as a three-dimensional heat map in the precision mode;

[0050] A limb correction module, which uses the skeleton graph sequence of the action video to be corrected and the standard action video to calculate the difference between the action done by the athlete and the standard action through an action comparison scoring algorithm, and gives the corresponding score and correction suggestions, and the score record is also recorded in the Fabric blockchain network, according to the pre-written smart contract, if the score reaches a certain standard, the blockchain network can authorize the relevant user to obtain the corresponding new sports course practice video and continue to practice and learn.

[0051] The embodiment of the present application has the beneficial effects that: the present application proposes a multi-module, human-computer interaction, embedded a series of sports movement action recognition and comparison correction algorithm sports movement action analysis and correction system for the demand of existing sports enthusiasts to accurately learn sports movement action, the present application uses a scoring and correction algorithm to compare the corresponding corrected sports movement action and standard sports movement action, calculates the difference between the two action sequences, and gives the corresponding score and specific correction suggestion, which has good practicality and robustness from the use effect, the present application uses the blockchain technology to store the behavior record of the user (including the course purchase record, the sports score record), the sports standard video is all stored in the cloud database, and the access control mechanism in the Fabric blockchain network is used to authorize the cloud database to distribute the corresponding follow-up course to the user for follow-up, and finally give the corresponding score. The score record is also recorded in the Fabric blockchain network, according to the pre-written smart contract, if the score reaches a certain standard, the blockchain network can authorize the relevant user to obtain the corresponding sports course follow-up second video and continue to follow-up and learn. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a method flowchart of the present application;

[0053] Figure 2 is a module classification diagram of the present application;

[0054] Figure 3 is a structure flowchart of the video loading module of the present application;

[0055] Figure 4 is a structure flowchart of the feature extraction module of the present application;

[0056] Figure 5 is a structure flowchart of the action recognition module of the present application;

[0057] Figure 6 is a structure flowchart of the body correction module of the present application. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings.

[0059] As Figure 1As shown, the embodiment of the application is a sports movement analysis and correction method based on Fabric blockchain network and joint feature cosine similarity, the main steps include using blockchain technology to store the behavior records of users (including course purchase records, movement score records), and storing all movement standard videos in the cloud database, and authorizing the cloud database to distribute the corresponding follow-up courses to the users through the access control mechanism in the Fabric blockchain network. A new joint feature real-time system model PostEX is proposed, which is used to extract the skeleton graph of the movement video, a movement sequence similarity calculation algorithm is proposed, which is used to identify specific movements and match corresponding movements from the video library for comparison, and finally a scoring and correction algorithm is used in the movement analysis and correction system to give corresponding scores and correction feedback suggestions. The score record will also be recorded in the Fabric blockchain network, according to the pre-written smart contract, if the score reaches a certain standard, the blockchain network can authorize the relevant user to obtain the corresponding movement course follow-up video and continue to follow up and learn.

[0060] The specific implementation steps are as follows.

[0061] S1: The standard movement video sequence is preloaded in the cloud database for users to learn selectively. The video loading module is divided into two parts, namely standard video loading and user video loading. The standard video loading folder includes various popular sports training projects such as shooting, boxing, golf, tennis, bowling, and driving. Each category has multiple clips for users to choose as standard videos for follow-up. The user video loading folder is the follow-up video saved by the user in the action learning module, which can be loaded into the system to summarize experience.

[0062] S2: Obtain the purchase record of the corresponding sports movement video stored in the Fabric blockchain network by the user, authorize the cloud database by the Fabric blockchain network, and distribute the corresponding purchased movement teaching video to the user. The entire follow-up process of the user will be recorded by a camera and saved as a to-be-corrected movement video. After the user completes the training, the to-be-corrected movement video is transmitted into the system for processing.

[0063] After the user purchases the corresponding video, the Fabric resource authorization can be used to pre-load the purchased standard movement video sequence in the cloud database for the user to learn and imitate.

[0064] The specific Fabric resource authorization process is as follows:

[0065] First, a log channel is set in the Fabric blockchain network, and a distributed ledger is set in the log channel to record user behavior and store log records. The distributed ledger is decentralized, and its copy is stored in all user nodes and updated synchronously. User behavior records are recorded in the Fabric blockchain network in an append-only manner, and encryption technology is used to ensure that once the record is added to the ledger, it cannot be modified again.

[0066] The user needs to register his identity information when logging in to the application for the first time, and returns the user's key pair (sk, pk). Note that only the user knows his private key sk, and other nodes in the Fabric blockchain network can easily get the user's public key pk. When the user needs to log in again, he can directly use the key pair to log in to the application for related sports learning process operation.

[0067] After the user enters the application, he can purchase related sports teaching course videos. The purchase record will be recorded in the local block, and after the block is full, it will be broadcast to the Fabric blockchain network and verified and formally chained. After the chaining process, the block that stores the user behavior record is formally added to the Fabric blockchain network.

[0068] The administrator node in the Fabric blockchain network can query the relevant user behavior record, and according to the behavior record, sign and distribute the specified teaching course to the corresponding user. Among them, the administrator node uses the administrator private key and the user public key to sign the specified teaching course, and the user end uses the user private key to decrypt the signature, thereby obtaining the teaching course index, and downloading the relevant teaching course in the cloud database for learning.

[0069] The chaining process of the to-be-verified block mentioned in the Fabric resource authorization process mainly includes the following steps:

[0070] Step 1: The to-be-verified block in a node in the Fabric blockchain network has been written full, then the verification block is submitted to the Fabric blockchain network and sent to all nodes.

[0071] Step 2: After receiving the verification block, each node in the Fabric blockchain network will verify the block, including checking whether the block information meets the rules and standards in the network, and verifying whether the initiator has sufficient authority and qualification to submit the block.

[0072] Step 3: The nodes that pass the verification endorse the block, that is, they recognize and verify the validity and correctness of the block. In the Fabric blockchain network, it is stipulated that at least 50% of the nodes endorse the block to confirm the validity of the block.

[0073] Step 4: Once the endorsement threshold is reached, the Fabric blockchain network marks the block as "confirmed" and officially stores the block in the Fabric blockchain network, while each node updates the copy of the distributed ledger and permanently stores it in the Fabric blockchain network.

[0074] S3: The pre-processing of the video sequence to be corrected is performed, so that each frame of the video matches each frame of the corresponding standard action video sequence in the standard action video library.

[0075] When the user needs to compare and correct the sports action followed, the follow-up video action sequence A is first sampled in this module (the standard action video sequence in the video library has been pre-sampled). The sampling process is as follows:

[0076] In order to extract the same number of frames from the follow-up video action sequence as the number of frames in the video sequence in the video library, the sampling method used by the present application is the sliding window sampling method. Specifically, x pictures are taken at equal intervals in the follow-up video action sequence (the standard action video sequence in the video library has also been pre-sampled by the sliding window sampling method at equal intervals to take x pictures), thereby obtaining video sequences with the same number of frames.

[0077] This module mainly uses the joint feature real-time system PostEX, which is a human key point recognition network structure. This network structure can extract coordinates of key points including the subject, feet, face, etc., and is suitable for single and multiple people, and can well complete the pose estimation tasks of human action, facial expression, finger movement, etc., and has good robustness. With 25 body parts as key points, the operation time is independent of the number of detected people. The input is mainly a group of pictures, a video sequence or a network camera video stream. The output is the original picture group or video frame + key point display.

[0078] The overall process of feature extraction is as follows:

[0079] First, the input video is preprocessed, and the preprocessing process includes: 1) using a sliding window sampling method to sample the video at equal intervals to obtain a motion-corrected video with the same number of frames as the video library. 2) Because the purpose of this system is to capture human motion, which is independent of the background of the video, the invention eliminates the interference of the video background. The invention converts all original RGB frames to grayscale representation, which reduces video memory and avoids color differences. 3) The method of spatial enhancement is also used to remove the background, so as to achieve the effect of only studying the human motion itself. 4) In the process of converting RGB frames to grayscale frames, sharp transitions may occur due to random noise. In order to eliminate such effects, the invention uses an average filter to reduce the generation of random noise.

[0080] Next, the preprocessed video sequence is first subjected to feature preprocessing of the first 10 layers of the VGG19 network, and then the video sequence is converted into image features F, and then divided into two branches to predict the key point confidence and affinity vector of each point. Where S is the confidence network, and L is the affinity vector field network: , After predicting the confidence of each joint, the NMS (Non-Maximum Suppression) is used to detect the joints. Then the affinity vector between the detected joints is line integrated to obtain the affinity between the joints. Then the human key points and the corresponding affinity can be modeled from the perspective of graph theory. In this way, the problem of human skeleton pose estimation is transformed into a multiple bipartite matching problem, and finally the Hungarian algorithm is used to obtain the final recognition result of the human skeleton.

[0081] In the accurate mode, because the output needs to be a three-dimensional human skeleton graph, and due to the different shooting angles and heights of the characters, the coordinates of the three-dimensional skeleton graph extracted are prone to deviation, so standardization and normalization processing are also performed in the preprocessing process. The normalization formula is: . Wherein, and are the mean and variance of the three-dimensional data obtained within 10ms; is the original three-dimensional coordinate vector, is the result after standard action normalization processing. Because the height and fatness of different people are very different, the bone joint coordinates of the same action sequence are different, so the skeleton scale needs to be standardized. Although there are differences between individuals, the proportions of the sizes of the parts of the human body and the height are almost the same. Therefore, a data standardization method is proposed by using the proportions of the joints of the human body and the height. The idea of three-dimensional skeleton data standardization is to establish a human body coordinate system, and to obtain the standard center point P c(a,b,c,),a,b,c for the acquisition of the center point of the x, y, z coordinate value, take 25 bone points and the difference between the center point, so as to obtain the center standardization after the human skeleton coordinate point P s The height of the human body D is obtained by the Euclidean distance, and the x, y and z of the 25 bone point coordinates after the center standardization are divided by the height D, to obtain the scale standardization of the bone coordinates.

[0082] The real-time computing system uses a machine learning method to assign parts to the depth image of the human body, which is to train a classifier to analyze different parts of the body (such as arms, heads, trunks, hands, etc.). For training the classifier, sample from a large motion capture database to generate depth images of human poses of various shapes and sizes, and a random decision forest classifier is used. The decision tree is trained in advance with some depth images marked with human body parts, and the decision tree is continuously optimized during training. When the decision tree accurately classifies the specified human body part depth image, the trained decision tree obtains the probability of each pixel in each human body part. The next step is to calculate the most likely area for each human body part. Finally, the human joint position predicted by the classifier is calculated to form the final skeleton data.

[0083] S4: input the pre-processed video sequence, use the joint feature real-time system model PostEX for real-time calculation, output and extract the limb key point coordinates of each frame of the motion action video sequence, and when the user sets to the accurate mode, output in the form of three-dimensional heat map.

[0084] Some traditional human similarity measurement algorithms are simple in design, consume less computing resources and operation time, but the effect is not so good. The data set required by using the deep learning method for training is too large, the computing resources consumed are too much, and the training time is too long. Therefore, the motion sequence similarity measurement algorithm based on the key point extraction of the joint feature real-time system.

[0085] The module calculates the similarity of the practice action sequence A and the standard action sequence B, and the specific process is as follows:

[0086] Step 1: The 25 key points extracted by the joint feature real-time system are combined two by two to form limb vector features. The combination rule is: if there is a limb between the joint i (coordinates (x , ) ) and the joint j (coordinates (x , ) ), the limb vector feature is represented as (x - - ); if there is no limb between the joint i and the joint j, the limb vector feature is represented as .

[0087] Step 2: According to the recognition result of the human body action, 24 limbs can be obtained for each human body skeleton, so 24 effective non-zero limb vectors can be obtained by the formula in step 1. The cosine value of the included angle between each two limb vectors is calculated, and 276 cosine values of the included angle can be obtained. The 276 cosine values of the included angle are combined into a vector, so that a picture is converted into a 276-dimensional cosine similarity vector after the above single picture feature extraction.

[0088] Step 3: Then the feature extraction is performed on each picture of the obtained follow-up action sequence A and the standard action sequence B of the two groups of pictures, and two groups of vectors A n and B n can be obtained, wherein A n represents the vector representation result of the nth picture of the action sequence A, and B n represents the vector representation result of the nth picture of the action sequence B. Then, all A n are spliced to obtain matrix 1, and all B n are spliced to obtain matrix 2.

[0089] Step 4: After obtaining the matrix, a problem needs to be considered, that is, the dimension of the matrix obtained after the above steps is 276n. This dimension is very large, and some values of the matrix are redundant in the process of completing the action sequence. It is a difficult problem in the current human body action recognition based on skeleton to find and remove those redundant features. Therefore, the present application designs a method for selecting features of the matrix, and the specific steps are as follows:

[0090] 1. The present application first finds a human body action data set called G3D data set

[0091] 2. All action sequences in the G3D data set are subjected to matrix extraction, and then all the matrices are flattened into one-dimensional feature vectors.

[0092] 3. Calculate the variance of all features. Remove all features with a variance below 0.05. Because these features represent no change in all action sequences, these features are redundant and have no reference value for representing human body action sequences.

[0093] 4. After filtering out the features with low variance, the recursive feature elimination method is used to select the features. The specific implementation steps of the recursive feature elimination method are as follows:

[0094] (1) First, the original feature set D is input into the machine learning algorithm as a whole as a training set, and then the original features are sorted according to certain attributes (for example, the feature_importances_attribute) according to the effect of the machine learning algorithm

[0095] (2) Some features that perform poorly in the machine learning algorithm are removed, and the features that perform well in the machine learning algorithm are retained to form a new feature set D1.

[0096] (3) D1 is input again into the machine learning algorithm, and the same operations as ① and ② are performed, and then the operations of ① and ② are repeatedly repeated until the number of selected features reaches the required number of reduction.

[0097] Here, the present application uses random forest, Lightgbm and xgboost to perform RFE feature selection. Each learner can obtain a feature subset. Then take the intersection of the three feature subsets as the final feature selection output.

[0098] ⑤ Save the features selected from ① to ④ as "action recognition feature subset"

[0099] Step 5: First, extract matrix 1 and matrix 2 according to the "action recognition feature subset", and the result is that matrix 1 and matrix 2 become vector 1 and vector 2. Then calculate the cosine similarity of vector 1 and vector 2, and the result is the similarity of the two action sequences. The higher the similarity value, the more standard the practice action is.

[0100] Step 6: The motion video in the video library has been calculated in advance by the cosine similarity algorithm to obtain the corresponding action matrix. After calculating the practice action matrix, only need to search for the highest similarity standard motion matrix in the video library according to the method of step 5, so that the practice action video is matched with the standard motion video, thereby obtaining the specific name of the practice action.

[0101] S5: Use the feature extraction algorithm of the cosine value of the human body limb angle to convert the extracted limb key point coordinate value feature into a limb vector value feature, and then into a limb cosine similarity feature value; search for the motion action with the closest vector value cosine similarity in the motion action material video library according to the extracted action representation, thereby identifying the specific motion action name made by the user, and making the to-be-corrected motion action video correspond to the standard motion action video.

[0102] Take the to-be-corrected motion matrix A in the action recognition module n and the standard motion matrix B nAs input, the difference between the movement action made by the player and the standard movement action is calculated by a movement comparison scoring algorithm, and the corresponding score and correction suggestions are given. The specific algorithm process is as follows.

[0103] Step 1: Since A n = {a1, a2, a3,..., a n} (a i is the cosine similarity vector of the i-th frame) and B n = {b1, b2, b3,..., b n} (b i is the cosine similarity vector of the i-th frame) are both n*276 matrices (composed of n cosine similarity vectors), the similarity algorithm of the A matrix and the B matrix is:

[0104]

[0105] wherein, represents the cosine similarity vector of the i-th frame in the matrix A, represents the cosine similarity vector of the j-th frame in the matrix B, represents the similarity size of the i-th frame in the matrix A and the j-th frame in the matrix B.

[0106] Step 2: According to the above formula, the calculated is calculated by the following scoring formula:

[0107]

[0108] wherein, , is the mapping parameter converted from similarity to percentage, so the corresponding score score is a percentage score, represents the similarity size of the i-th frame in the matrix A and the j-th frame in the matrix B.

[0109] The user can view the video sequence similarity score column chart to observe the action score of his own body. Then, the human body posture similarity of the learning video image frame is viewed, the similarity of each image frame is observed, and the calculation result is a percentage value. Finally, the similarity of each part of the human body in the video is viewed, the similarity calculation result of the limb on the entire time sequence is generated by selecting each limb part, and more detailed information of the time to be trained and corrected is provided to correct the action in a targeted manner.

[0110] Each time the score record is encrypted into hash data and stored in the Fabric blockchain network in the form of a block, and according to the pre-written smart contract, if the score reaches a certain standard, the administrator node in the blockchain network can authorize the relevant user to obtain the corresponding sports course follow-up video of the second phase and continue to follow up and learn.

[0111] S6: The score and correction algorithm are used to compare the to-be-corrected motion action skeleton graph and the standard motion action skeleton graph, provide corresponding scores and action correction suggestions, use an interactive interface to return the matched feedback result to the user, and provide difference details and specific action detailed guidance words.

[0112] To facilitate the user to use the function interface to realize various functions, the system homepage displays some operation guidance instruction words to the user. The functions designed and realized by the system in the function interface guide the user to use and realize various function effects through buttons, words and the like. After the user performs the corresponding operation according to the system prompt, the system provides the feedback result corresponding to the operation for the user to view.

[0113] S7: The score record is encrypted and stored in the Fabric blockchain network in the form of a block, and according to the pre-written smart contract, if the score reaches a certain standard, the blockchain network authorizes the relevant user to obtain the corresponding new sports course follow-up video and continue to follow up and learn.

[0114] The embodiment of the application also provides a sports motion action analysis and correction method system based on the Fabric blockchain network and the joint feature cosine similarity algorithm using the above method, as shown in Figure 2 , comprising:

[0115] I. Video loading module:

[0116] As shown in Figure 3 , the standard action video sequence is preloaded in the cloud database for the user to learn in a targeted and selective manner. The video loading module is divided into two parts, namely standard video loading and user video loading. The standard video loading folder includes various popular sports training projects, such as shooting, boxing, golf, tennis, bowling, driving, etc. There are multiple clips in each category for the user to choose as a standard video for follow-up practice. The user video loading folder is the follow-up video saved by the user in the action learning module, which can be loaded into the system to summarize experience. After the user purchases the corresponding video, the user can pre-load the purchased standard action video sequence in the cloud database through the Fabric resource authorization, which can be used for the user to learn and imitate.

[0117] The specific Fabric resource authorization process is as follows:

[0118] First, a log channel is set in the Fabric blockchain network, and a distributed ledger is set in the log channel to record user behavior and store log records. The distributed ledger is decentralized, and its copy is stored in all user nodes and updated synchronously. User behavior records are recorded in the Fabric blockchain network in an append-only manner, and encryption technology is used to ensure that once the record is added to the ledger, it cannot be modified again.

[0119] The user needs to register his identity information when logging in to the application for the first time, and returns the user's key pair (sk, pk). Note that only the user knows his private key sk, and other nodes in the Fabric blockchain network can easily get the user's public key pk. When the user needs to log in again, he can directly use the key pair to log in to the application for related sports learning process operation.

[0120] After the user enters the application, he can purchase related sports teaching course videos. The purchase record will be recorded in the local block, and after the block is full, it will be broadcast to the Fabric blockchain network and verified and formally chained. After the chaining process, the block that stores the user behavior record is formally added to the Fabric blockchain network.

[0121] The administrator node in the Fabric blockchain network can query the relevant user behavior record, and according to the behavior record, sign and distribute the specified teaching course to the corresponding user. Among them, the administrator node uses the administrator private key and the user public key to sign the specified teaching course, and the user end uses the user private key to decrypt the signature, thereby obtaining the teaching course index, and downloading the relevant teaching course in the cloud database for learning.

[0122] The chaining process of the to-be-verified block mentioned in the Fabric resource authorization process mainly includes the following steps:

[0123] Step 1: The to-be-verified block in a node in the Fabric blockchain network has been written full, then the verification block is submitted to the Fabric blockchain network and sent to all nodes.

[0124] Step 2: After receiving the verification block, each node in the Fabric blockchain network will verify the block, including checking whether the block information meets the rules and standards in the network, and verifying whether the initiator has sufficient authority and qualification to submit the block.

[0125] Step 3: The nodes that pass the verification endorse the block to be verified, that is, to recognize and verify the validity and correctness of the block. In the Fabric blockchain network, it is stipulated that at least 50% of the nodes endorse the block to confirm the validity of the block.

[0126] Step 4: Once the endorsement threshold is reached, the Fabric blockchain network marks the block as "confirmed" and officially stores the block in the Fabric blockchain network, while each node updates the copy of the distributed ledger and is permanently stored in the Fabric blockchain network.

[0127] II. Action learning module:

[0128] The module allows users to watch the standard video imported by the video loading module and follow the actions of the standard video. When the user starts learning, the user can select the key actions of the learning video, and after all the action learning is completed, the user can follow the complete video. The infrared sensor is used to follow the disassembled action in real time, and the action is corrected in time according to the feedback result. After completing the disassembled action learning, the monocular camera is used for follow-up practice and the video is saved to facilitate the extraction of the skeleton key point by the posture perception module. After the follow-up practice is completed, the follow-up action video sequence is generated and uploaded.

[0129] III. Feature extraction module:

[0130] As shown in Figure 4 When the user needs to compare and correct the followed sports action, the follow-up video action sequence A is first sampled (the standard motion action video sequence in the video library has been pre-sampled). The sampling process is as follows:

[0131] In order to extract the same number of frames from the follow-up video action sequence and the video sequence in the video library, the sampling method used by the present application is the sliding window sampling method. Specifically, x pictures are taken at equal intervals in the follow-up video action sequence (the standard motion action video sequence in the video library has also been pre-sampled by the sliding window sampling method at equal intervals to take x pictures), thereby obtaining video sequences with the same number of frames.

[0132] This module mainly uses the joint feature real-time system PostEX, which is a human key point recognition network structure. This network structure can extract coordinates of key points including the recipient, the foot, the face, etc. It is suitable for single person and multiple persons, and can well complete the posture estimation tasks of human action, facial expression, finger movement, etc. It has good robustness. 25 body parts are used as key points for recognition, and the operation time is independent of the number of detected persons. The input is mainly a group of pictures, a video sequence or a network camera video stream. The output is the original picture group or video frame + key point display.

[0133] The overall process of feature extraction is as follows:

[0134] First, the input video is preprocessed, and the preprocessing process includes: 1) using a sliding window sampling method to sample the video at equal intervals to obtain a motion-corrected video with the same number of frames as the video library. 2) Because the purpose of this system is to capture human motion, it is irrelevant to the background of the video, so it is necessary to exclude the interference of the video background. This application converts all original RGB frames to grayscale representation, which reduces video memory and avoids color differences. 3) Spatial enhancement method is also used to remove the background, so as to achieve the effect of only studying the human motion itself. 4) In addition, during the conversion of RGB frames to grayscale frames, sharp transitions may occur due to random noise. In order to eliminate such effects, an average filter is used to reduce the generation of random noise.

[0135] Next, the preprocessed video sequence is first subjected to feature preprocessing of the first 10 layers of the VGG19 network, and then the video sequence is converted into image features F, and then divided into two branches to predict the key point confidence and affinity vector of each point. Where S is the confidence network, and L is the affinity vector field network: , After predicting the confidence of each joint, NMS (Non-Maximum Suppression) is used to detect the joints. Then the affinity vector between the detected joints is integrated to obtain the affinity between the joints. Then the human key points and the corresponding affinity can be modeled from the perspective of graph theory. In this way, the human skeleton pose estimation problem is transformed into a multiple bipartite matching problem, and finally the Hungarian algorithm is used to obtain the final recognition result of the human skeleton.

[0136] In the accurate mode, because the output needs to be a three-dimensional human skeleton graph, and due to the different shooting angles and heights of the characters, the coordinates of the three-dimensional skeleton graph extracted are prone to deviation, so standardization and normalization processing are also required in the preprocessing process. The normalization formula is: . Wherein, and are the mean and variance of the three-dimensional data obtained within 10ms; is the original three-dimensional coordinate vector, is the result after standard action normalization processing. Due to the great difference in height and fatness of different people, the bone joint coordinates of the same action sequence are different, so it is necessary to standardize the skeleton scale. Although there are differences between individuals, the proportions of the sizes of each part of the human body and the height are almost the same. Therefore, a data standardization method is proposed by using the proportions of each joint and height of the human body. The idea of three-dimensional skeleton data standardization is to establish a human body coordinate system, and to obtain the standard center point P c(a, b, c,), a, b, c are the x, y, z coordinate values of the center point, and the difference between the 25 skeleton points and the center point is obtained, so as to obtain the human skeleton coordinate point P after center standardization s The height D of the human body is obtained by the Euclidean distance, and the x, y and z of the 25 skeleton points after center standardization are divided by the height D to obtain the scale standardized skeleton coordinates.

[0137] The real-time computing system uses a machine learning method to assign parts to the depth image of the human body, which is to train a classifier to analyze different parts of the body (such as arms, heads, trunks, hands, etc.). For training the classifier, sample from a large motion capture database to generate human posture depth images of various shapes and sizes, and a random decision forest classifier is used. The decision tree is trained in advance with some depth images marked with human body parts, and the decision tree is continuously optimized during training. When the decision tree accurately classifies the specified human body part depth image, the trained decision tree obtains the probability of each pixel in each human body part. The next step is to calculate the most likely area for each human body part. Finally, the human joint position predicted by the classifier is calculated to form the final skeleton data.

[0138] Four, action recognition module:

[0139] As shown in Figure 5 , some traditional human similarity measurement algorithms are simple in design, consume less computing resources and computing time, but the effect is not so good. The data set required for training using the deep learning method is too large, the computing resources consumed are too much, and the training time is too long. Therefore, the present application designs an action sequence similarity measurement algorithm based on the key point extraction of the joint feature real-time system.

[0140] The specific process of calculating the similarity of the practice action sequence A and the standard action sequence B in this module is as follows:

[0141] Step 1: The 25 key points extracted after the joint feature real-time system are combined two by two to form limb vector features. The combination rule is: if there is a limb between joint i (coordinates (x , )) and joint j (coordinates (x , )), the limb vector feature is represented as (x - - ); if there is no limb between joint i and joint j, the limb vector feature is represented as .

[0142] Step 2: According to the recognition result of human body action, each human body skeleton can obtain 24 limbs, so the formula in step 1 can obtain 24 effective non-zero limb vectors. The cosine value between each two limb vectors is calculated, and 276 cosine values can be obtained, and the 276 cosine values are combined into a vector, so that a picture is converted into a 276-dimensional cosine similarity vector after the above single picture feature extraction.

[0143] Step 3: Then the feature of each picture of the obtained follow-up action sequence A and the standard action sequence B is extracted, and two groups of vectors A n and B n are obtained, wherein A n represents the vector representation result of the nth picture of the action sequence A, and B n represents the vector representation result of the nth picture of the action sequence B. Then the present application splices all A n to obtain matrix 1, and splices all B n to obtain matrix 2.

[0144] Step 4: After obtaining the matrix, a problem needs to be considered, that is, the dimension of the matrix obtained after the above steps is 276n. This dimension is very large, and some values of the matrix are redundant in the process of completing the action sequence. It is a difficult problem in the current human body action recognition based on skeleton to find and remove those redundant features. Therefore, the present application designs a feature selection method for the matrix, and the specific steps are as follows:

[0145] 1. The present application first finds a human body action data set called G3D data set

[0146] 2. Extract the matrix of all action sequences in the G3D data set, and then flatten all the matrices into one-dimensional feature vectors.

[0147] 3. Calculate the variance of all features. Remove all features with a variance below 0.05. Because these features represent no change in all action sequences, these features are redundant and have no reference value for representing human body action sequences.

[0148] 4. After filtering out the features with low variance, use the recursive feature elimination method to select the features. The specific implementation steps of the recursive feature elimination method are as follows:

[0149] (1) First, the original feature set D is input into the machine learning algorithm as a whole as a training set, and then the original features are sorted according to certain attributes (for example, the feature_importances_attribute) according to the effect of the machine learning algorithm

[0150] (2) Some features that perform poorly in the machine learning algorithm are removed, and the features that perform well in the machine learning algorithm are retained to form a new feature set D1.

[0151] (3) D1 is input into the machine learning algorithm again, and the same operations as ① and ② are performed, and then the operations of ① and ② are repeatedly repeated until the number of selected features reaches the required number of reductions.

[0152] Here, the present application uses random forest, Lightgbm and xgboost to perform RFE feature selection. Each learner can obtain a feature subset. Then take the intersection of the three feature subsets as the final feature selection output.

[0153] ⑤ Save the features selected from ① to ④ as "action recognition feature subset"

[0154] Step 5: First, extract matrix 1 and matrix 2 according to the "action recognition feature subset", and the result is that matrix 1 and matrix 2 become vector 1 and vector 2. Then calculate the cosine similarity of vector 1 and vector 2, and the result is the similarity of the two action sequences. The higher the similarity value, the more standard the practice action is.

[0155] Step 6: The motion video in the video library has been calculated in advance by the cosine similarity algorithm to obtain the corresponding action matrix. After calculating the practice action matrix, only need to search for the highest similarity standard motion matrix in the video library according to the method of step 5, so that the practice action video is matched with the standard motion video, and the specific name of the practice action is obtained.

[0156] Five: limb correction module:

[0157] As shown in Figure 6 , the motion matrix A to be corrected in the action recognition module n and the standard motion matrix B n are input, and the difference between the motion action performed by the athlete and the standard motion action is calculated by an action comparison scoring algorithm, and the corresponding score and correction suggestion are given. The specific algorithm flow is as follows.

[0158] Step 1: Since A n ={a1,a2,a3,...,a n}(a icosine similarity vector of the i-th frame) and B n = {b1, b2, b3,..., bn} (1) n = {b1, b2, b3,..., bn} (1) i cosine similarity vector of the i-th frame) and B

[0159]

[0160] wherein, cosine similarity vector of the i-th frame in matrix A, cosine similarity vector of the j-th frame in matrix B, similarity between the i-th frame in matrix A and the j-th frame in matrix B.

[0161] Step 2: According to the above formula, the is calculated by the following scoring formula:

[0162]

[0163] wherein, , is a mapping parameter converted from similarity to percentage, so the corresponding score Score is a percentage score, similarity between the i-th frame in matrix A and the j-th frame in matrix B.

[0164] The user can view the video sequence similarity score column chart to observe the action score of his / her body. Then, the user can view the human body posture similarity of the image frames of the learning video to observe the similarity of each image frame, and the calculation result is a percentage value. Finally, the user can view the similarity of each part of the human body in the video, select each body part, and generate a curve of the similarity calculation result of the body part in the entire time sequence, so as to provide more detailed information of the time for training and correction, and correct the action in a targeted manner.

[0165] Each score record is encrypted into hash data and stored in the Fabric blockchain network in the form of a block. According to the pre-written smart contract, if the score reaches a certain standard, the administrator node in the blockchain network can authorize the relevant user to obtain the corresponding second-period video of the sports course and continue to follow and learn.

[0166] The module also contains a function interface and feedback prompts. In order to facilitate the user to use the function interface to realize various functions, the system homepage displays some operation guidance instruction words to the user. In the function interface, the functions designed and realized by the system guide the user to use and realize various function effects through buttons, words and other ways. After the user performs the corresponding operation according to the system prompt, the system provides the feedback result corresponding to the operation for the user to view.

[0167] The above only discloses a preferred embodiment of the present application, and of course cannot limit the scope of the rights of the present application, so the equivalent changes made by the claims of the present application still belong to the scope covered by the present application.

Claims

1. A sports movement action analysis and correction method based on a Fabric blockchain network and a joint feature cosine similarity algorithm, characterized in that, The method comprises the following steps: S1: constructing a standard sports action video sequence teaching material library in a cloud database; S2: obtaining a purchase record of a corresponding sports action video stored by a user in a Fabric block chain network, authorizing the cloud database by the Fabric block chain network, distributing a corresponding purchased sports teaching video to the user, and using a camera to shoot and save a to-be-corrected sports action video of the whole follow-up training process of the user, and transmitting the to-be-corrected sports action video into the system for processing after the user completes the training; S3: preprocessing the to-be-corrected sports action video sequence, so that each frame of the video is matched with each frame in the corresponding standard action video sequence in the standard action video material library; S4: inputting the preprocessed video sequence, using a joint feature real-time system model PostEX to perform real-time calculation, outputting and extracting limb key point coordinates of each frame of the sports action video sequence, and outputting in the form of a three-dimensional heat map when the user sets to the accurate mode; S5: using a feature extraction algorithm of a human body limb included angle cosine value to convert the extracted limb key point coordinate value feature into a limb vector value feature, and then into a limb cosine similarity feature value; using the action sequence similarity measurement based on the joint feature real-time system key point extraction to search for a sports action with the closest vector value cosine similarity in the sports action material video library, so as to identify a specific sports action name made by the user, and make the to-be-corrected sports action video correspond to the standard sports action video; The method of the action sequence similarity measurement based on the joint feature real-time system key point extraction comprises the following steps: S51: combining the 25 key points extracted through the joint feature real-time system two by two to form a limb vector feature; S52: calculating the included angle cosine value between each two limb vectors to obtain 276 included angle cosine values, and forming a vector, which is converted into a 276-dimensional cosine similarity vector; S53: Feature extraction is performed on each of the two sets of photos of the obtained follow-up action sequence and the standard action sequence, and a n vector representation result of the nth picture representing the follow-up action sequence, b n a vector representation result of the nth picture representing the standard action sequence, and then all a n are spliced to obtain a matrix A, and all b n are spliced to obtain a matrix B; S54: using a feature selection method for a matrix to remove redundant matrices, which comprises the following steps: S541: using a G3D data set and extracting a matrix for all action sequences in the G3D data set, and then flattening all the matrices into one-dimensional feature vectors; S542: calculating the variance of all features, and removing all features with a variance lower than 0.05; S543: using a recursive feature elimination method to select features, which comprises the following steps: S5431: inputting all features as a training set into a machine learning algorithm, and then arranging the original features according to attributes according to the effect of the machine learning algorithm; S5432: removing some features with poor performance in the machine learning algorithm, and retaining features with good performance in the machine learning algorithm to form a new feature set; S5433: input the new feature set as input into the machine learning algorithm again, and perform the same operation as S5431 and S5432, and then repeatedly perform the operation of S5431 and S5432 in a reciprocating manner until the number of selected features reaches the required number of reduction; S544: save the selected features in S541-S543 as the action recognition feature subset; S6: compare the to-be-corrected motion action skeleton graph with the standard motion action skeleton graph using the scoring and correction algorithm, provide the corresponding score and action correction suggestion, return the matching feedback result to the user using the interactive interface, and provide the difference details and specific action detailed guidance text; S7: encrypt the score record and store it in the Fabric blockchain network in the form of a block, and according to the pre-written smart contract, if the score reaches a certain standard, the blockchain network authorizes the relevant user to obtain the corresponding new period of motion course follow-up video and continue to follow up and learn.

2. The method according to claim 1, wherein, The on-chain step of the Fabric blockchain network comprises: S21: when a to-be-verified block in a node in the Fabric blockchain network has been written full, the verification block is submitted to the Fabric blockchain network and sent to all nodes; S22: after receiving the verification block, each node in the Fabric blockchain network verifies the block, and the verification content includes checking whether the block information conforms to the rules and standards in the network and verifying whether the initiator has sufficient authority and qualification to submit the block; S23: the node that passes the verification endorses the to-be-verified block, and in the Fabric blockchain network, it is stipulated that at least 50% of the nodes endorse the block to confirm the validity of the block; S24: once the endorsement threshold is reached, the Fabric blockchain network marks the block as confirmed, and the block is officially stored in the Fabric blockchain network, and each node updates the distributed ledger copy and is permanently saved in the Fabric blockchain network. 3.The sports action analysis and correction method based on the Fabric blockchain network and the joint feature cosine similarity algorithm of claim 1, wherein, The camera comprises a monocular camera and an infrared sensor, the monocular camera is used for follow-up and video saving, and the infrared sensor is used for real-time follow-up and disassembly action.

4. The method of claim 1, wherein the Fabric-based blockchain network and the joint feature cosine similarity algorithm are used for sports motion analysis and correction. The pre-processing step of the to-be-corrected motion action video sequence comprises: obtaining the same number of video sequences as the number of frames in the video library using the sliding window method, converting all original RGB frames into grayscale representation, removing the background by the spatial enhancement method, and reducing the generation of random noise by using an average filter; the joint feature real-time system model PostEX comprises a feature extraction method: The pre-processed video sequence is first subjected to feature pre-processing of the first 10 layers of the VGG19 network, and then the video sequence is converted into image features F, and then divided into two branches to predict the key point confidence and affinity vector of each point, wherein S is the confidence network, and L is the affinity vector field network: , After predicting the confidence of each joint point, the joint points are detected by non-maximum suppression, and then the affinity vectors between the detected joint points are line integrated to obtain the affinity between the joint points. Then, the human key points and the corresponding affinity are modeled from the perspective of graph theory, and finally the Hungarian algorithm is used to obtain the final recognition result of the human skeleton.

5. The method according to claim 4, wherein, In the precision mode, standardization and normalization are also performed in the preprocessing process, and the normalization formula is: wherein, and are the mean and variance of the three-dimensional data obtained within 10 ms, respectively; A is the original three-dimensional coordinate vector, is the result after standard action normalization; further comprising a step of standardizing the skeleton scale and a data standard processing step, the data standard processing step comprising: establishing a human body coordinate system, obtaining the standard center point P c of the skeleton by taking the average value of the coordinates of two skeleton points, taking the difference between the 25 skeleton points and the center point to obtain the human skeleton coordinate point P s after center standardization, obtaining the height D of the human body by the Euclidean distance, dividing the x, y and z of the 25 skeleton points after center standardization by the height D to obtain the scale-standardized skeleton coordinates, and analyzing different parts of the body by training a classifier. 6.The sports action analysis and correction method based on a Fabric blockchain network and a joint feature cosine similarity algorithm according to claim 1, wherein, The S5 further comprises the following steps: S55: first extract matrix A and matrix B according to the action recognition feature subset, and the result is that matrix A and matrix B become vector a n and vector b n , and then calculate the cosine similarity of vector a n and vector b n ; S56: the motion video in the video library has been calculated to obtain the corresponding action matrix by the cosine similarity algorithm.

7. The method according to claim 6, wherein, The S6 specifically comprises the following steps: S61: Since A={a1,a2,a3,...,a...} n } and B={b1,b2,b3,...,b n Both A and B are n*276 matrices. The similarity algorithm between matrices A and B is as follows: where n represents the frame number, k represents the kth dimension in the 276-dimensional vector, a i represents the cosine similarity vector of the ith frame in matrix A, b j represents the cosine similarity vector of the jth frame in matrix B, d ij represents the similarity size of the ith frame in matrix A and the jth frame in matrix B; S62: d calculated from the above formula ij The specific score is calculated by the following scoring formula: wherein, , is the mapping parameter converted from the similarity to the percentage system, so the corresponding score score is the percentage score, d ij represents the similarity between the i-th frame of matrix A and the j-th frame of matrix B.

8. A system for running the sports motion action analysis and correction method based on the Fabric blockchain network and the joint feature cosine similarity algorithm according to any one of claims 1-7, comprising: Video loading module, after the user purchases the corresponding video, through the Fabric related resource authorization, preloads the purchased standard action video sequence in the cloud database, for the user to learn and imitate; Action learning module, the user follows the relevant sports video, generates the action video sequence to be corrected and uploads; Feature extraction module, pre-processes the motion action video sequence to be compared and takes it as the input of the joint feature real-time system model PostEX, through the real-time calculation of the model, outputs and extracts the skeleton limb point two-dimensional graph of each frame of the motion action video sequence, in the setting of the system precision mode, can consume more traffic and computing power to output the motion action video sequence as a three-dimensional heat map form; Action recognition module, using the feature extraction algorithm of human body limb angle cosine value, identifies the specific action classification of the motion, then extracts the corresponding standard action video from the video library, makes the to-be-corrected motion video correspond to the standard action video, and converts it into a standard video skeleton limb point two-dimensional graph through the joint feature real-time system model PostEX, which can also be output as a three-dimensional heat map in the precision mode; Body correction module, using the skeleton graph sequence of the to-be-corrected motion video and the standard motion video, through an action comparison scoring algorithm, calculates the difference between the user's motion action and the standard motion action, and gives the corresponding score and correction suggestion, the score record will also be recorded in the Fabric blockchain network, according to the pre-written smart contract, if the score reaches a certain standard, the blockchain network can authorize the relevant user to obtain the corresponding new period of motion course follow-up video and continue to follow-up learning.

Images