Training action recognition and correction method and system based on image point position comparison

Abstract

The invention belongs to the technical field of training action recognition, and provides a training action recognition and correction method and system based on image point position comparison, and the method comprises the following steps: collecting an image sequence of a trainer during action through an image obtaining device, extracting a training action key point position, and comparing the training action key point position with a standard action point position; and obtaining an error point location. According to the method, the related correction action group query library is constructed based on data of multiple trainers by carrying out correction training on the first training action to be corrected and combining the comparative analysis of the subsequent secondary training action set to be corrected, so that personalized correction suggestion combinations can be provided for a single trainee, the conditions of different trainers are integrated, and the correctness of the training actions is improved. When the trainee has action deviation, the trainee can obtain targeted correction suggestions from the query library, so that invalid or wrong training modes are reduced, correct actions can be mastered more quickly, and the time for reaching standard actions is shortened.

Description

Technical Field

[0001] The present invention belongs to the technical field of training action recognition, and specifically relates to a training action recognition and correction method and system based on image point position comparison. Background Art

[0002] In sports training, the standardization and accuracy of actions play a decisive role in training effects and the growth of athletes. However, in existing training, generally, the method of manually recognizing training actions and correcting them is used, which highly depends on the experience and subjective judgment of professionals, with low efficiency. In large-scale training scenarios, due to limited energy of professionals, it is difficult to conduct detailed and comprehensive analysis and guidance on the actions of each trainer. Moreover, there are differences in judgment criteria among different personnel, resulting in lack of consistency and accuracy in action evaluation results. On the other hand, some technical means using sensors for monitoring, although able to obtain some basic data, have obvious limitations in the layout and functions of sensors, and cannot comprehensively and accurately capture the key details of complex actions, and the acquisition of important information such as action postures and the coordination relationships of various parts is incomplete, making the evaluation of training actions insufficiently in-depth and difficult to discover deep-seated action deviation problems, ultimately affecting the improvement of training effects. Summary of the Invention

[0003] In order to make up for the deficiencies of the prior art and solve at least one of the technical problems proposed in the background art.

[0004] The technical solution adopted by the present invention to solve its technical problems is as follows: including the following steps: Collect an image sequence of a trainer performing an action through an image acquisition device, extract the key points of the training action and compare them with the key points of the standard action to obtain error points. Evaluate the severity of the training action by analyzing and calculating the error data of the error points of the trainer during the training cycle, and construct a set of training actions to be corrected. Conduct correction training on the training actions to be corrected, obtain relevant correction groups, and construct a query library for relevant correction action groups. Based on the constructed query library for relevant correction action groups, provide correction action selection suggestions for the trainer when the action deviation occurs again, and obtain a combination of correction suggestions.

[0005] As a further solution of the present invention: the process of obtaining the error points is as follows: Obtain the number of training actions of the sports event that the trainer is performing, and the images of each training action. Extract the key point coordinates of the training action based on the image of each action, and compare them with the key point coordinates of the corresponding standard action in the standard action image library. Calculate the deviation value between the key point coordinates of the training action and the key point coordinates of the standard action using the Mahalanobis distance; Compare the deviation value with the deviation threshold. If the deviation value is greater than the deviation threshold, mark this key point as an error point.

[0006] As a further solution of the present invention: the process of obtaining the set of training actions to be corrected is as follows: Analyze the error points of each training action to obtain the error characterization value and the ratio of the number of error occurrences; Multiply the error characterization value by the ratio of the number of error occurrences to obtain the action deviation evaluation coefficient; Mark the training actions with a value greater than the action deviation evaluation coefficient as training actions to be corrected; Obtain all the training actions to be corrected and construct a set of training actions to be corrected.

[0007] As a further solution of the present invention: the process of obtaining the error characterization value and the ratio of the number of error occurrences is as follows: Based on any single training, obtain the error points of each training action, and count the number of error points and the number of key points of each training action; Calculate the ratio of the number of error points to the number of key points of the training action to obtain the ratio of error points; Calculate the difference between the deviation value of each error point and the deviation threshold respectively, calculate the ratio of the obtained difference to the deviation threshold to obtain the error degree value, sum and average the error degree values of all error points to obtain the average error degree; Perform a weighted sum calculation on the ratio of error points and the average error degree to obtain the error characterization value; Count the number of times the error points appear in each training action, and calculate the ratio with the total number of trainings to obtain the ratio of the number of error occurrences.

[0008] As a further solution of the present invention: the process of constructing the relevant correction action group query library is as follows: Mark the set of training actions to be corrected as the first set of training actions to be corrected, and arrange them in order; Take the first training action to be corrected in the sorted order as the first correction training action and perform the first correction training; After the first correction training is completed, the trainer performs another training cycle of action training and constructs a second set of training actions to be corrected; Compare the training actions to be corrected in the first set of training actions to be corrected and the second set of training actions to be corrected, and extract the training actions that have been corrected among them; Mark the training actions consecutive to the first correction action as relevant influencing training actions; Analyze and process based on relevant influencing training actions, and calculate the ratio of similar point positions; Mark the relevant influencing actions in the combination to be analyzed with a ratio of similar point positions greater than the threshold as relevant corrective actions; Obtain all relevant corrective actions, combine them with the first corrective training action, and mark them as a relevant corrective action group; After the first corrective training, in the set of second corrective training actions, select the first-ranked training action to be corrected, mark it as the second corrective training action, perform the second corrective training, and repeat the process of identifying relevant influencing training actions, and obtain the relevant corrective action group; Repeat the above process of obtaining the relevant corrective action group until all training actions to be corrected meet the standards; Construct a relevant corrective action group query table for all relevant corrective action groups, and obtain relevant corrective action group query tables of several trainers to construct a relevant corrective action group query library.

[0009] As a further solution of the present invention: the correction process is as follows: Obtain the deviation value sequence of each error point position within the training cycle of the first-ranked training action to be corrected; Perform mean processing on the deviation values corresponding to each error point position to obtain the deviation mean; Based on the deviation mean, correct the error point positions of the first-ranked action to be corrected until the trainer reaches the standard training action for this action and stops the correction training.

[0010] As a further solution of the present invention: the process of obtaining the ratio of similar point positions is as follows: Obtain the deviation value sequences of the first corrective training action and the relevant influencing training actions respectively; Based on the combination analysis of the first corrective training action and any one of the relevant influencing training actions, mark it as a combination to be analyzed, and calculate the similarity representation value of the same key point positions in each combination to be analyzed; The similarity representation value is calculated by taking the absolute value of the cosine similarity; Mark the key point positions with a similarity greater than the similarity threshold as similar point positions; Count the number of the same key point positions and the number of similar point positions in each combination to be analyzed, and perform calculation processing on the number of similar point positions and the number of the same key point positions to obtain the ratio of similar point positions.

[0011] As a further solution of the present invention: the process of obtaining the correction suggestion combination is as follows: Obtain the coincidence ratio value and the similarity degree value; Perform weighted summation calculation on the coincidence ratio value and the similarity degree value to obtain the comparison judgment value; Set a comparison judgment threshold, and mark the group of actions to be selected corresponding to the first training action greater than the comparison judgment value threshold as the correction suggestion combination.

[0012] As a further solution of the present invention: The process of obtaining the coincidence ratio and the similarity degree value is as follows: If a trainer action deviation alarm occurs during the real-time training of the trainer, mark the corresponding training action as the training action to be compared. Based on the training action to be compared, compare it with the relevant correction action groups in the relevant correction action group query library, obtain the relevant correction action group with the training action to be compared as the first training action, and mark it as the action group to be selected. Obtain the error points of the first training action of the action group to be selected and the deviation value sequence corresponding to the error points. Compare the error points of the action to be compared with the error points of all the first training actions respectively to obtain the number of coincident error points, and perform a ratio process on the number of coincident error points and the number of error points of the action to be compared to obtain the coincidence ratio. Calculate the similarity representation values of the deviation value sequences of the error points of the action to be compared and the deviation value sequences of the error points of all the first training actions respectively.

[0013] As a further solution of the present invention: It includes: Point comparison module: Collect the image sequence of the trainer's actions through the image acquisition device, extract the key points of the training actions and compare them with the standard action points to obtain the error points. Action evaluation module: Evaluate the severity of the training actions by analyzing and calculating the error data of the error points of the trainer during the training cycle, and construct a set of training actions to be corrected. Relevant action group construction module: Carry out corrective training on the training actions to be corrected, obtain relevant correction groups, and construct a relevant correction action group query library. Correction suggestion module: Based on the constructed relevant correction action group query library, provide correction action selection suggestions for the trainer when the action deviation occurs again, and obtain the correction suggestion combination.

[0014] The beneficial effects of the present invention are as follows: 1. The present invention uses the image acquisition device to collect the image sequence of the training actions, extracts the key point coordinates and compares them with the standard actions, accurately calculates the deviation value using the Mahalanobis distance, can clearly mark the error points, and comprehensively and quantitatively calculates the action deviation evaluation coefficient, so as to screen out the training actions to be corrected and construct a set of training actions to be corrected, thus comprehensively and intuitively reflecting the quality and problem degree of the training actions, providing a clear and accurate object for the subsequent corrective training, effectively avoiding blind correction, and improving the training efficiency and resource utilization efficiency; 2. The present invention corrects and trains the first training action to be corrected, combines the comparative analysis of the subsequent second set of training actions to be corrected, and constructs a query library of relevant correction action groups based on the data of multiple trainers, so as to provide an individualized combination of correction suggestions for a single trainer. Also, considering the situations of different trainers, when a trainer has an action deviation, targeted correction suggestions can be obtained from the query library, reducing ineffective or incorrect training methods, enabling the trainer to master the correct actions faster and shortening the time to reach the standard actions. 3. The relevant correction action groups in the query library of the present invention are compared to screen out a combination of correction suggestions. If there are multiple combinations of correction suggestions, real-time feedback and correction of the training actions can be carried out, enabling the trainer to train more targeted, improving the accuracy and standardization of training, helping the trainer achieve the training goals faster, and enhancing the training effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below with reference to the accompanying drawings.

[0016] Figure 1 is a flowchart of the steps of a training action recognition and correction method based on image point position comparison according to the present invention; Figure 2 is an architecture diagram of a training action recognition and correction system based on image point position comparison according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to make the technical means, creative features, achieved purposes and functions of the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments. Embodiment 1

[0018] As Figure 1 shown, a training action recognition and correction method based on image point position comparison described in an embodiment of the present invention includes the following steps: Step 1: Collect an image sequence of a trainer performing an action through an image acquisition device, extract the key points of the training action and compare them with the key points of the standard action to obtain error points; In some embodiments, an image acquisition device is selected and set at different angles of the trainer to obtain all-round action information of the trainer. Among them, the image acquisition device can be a camera, such as a depth camera (Microsoft Kinect), which can not only obtain color images but also depth information, helping to more accurately capture the three-dimensional spatial position of the human body; Set the acquisition frame rate. For example, set the acquisition frame rate to 60 frames per second, and set the image resolution, such as 1920×1080, to ensure that the image details are sufficient for subsequent key point extraction; Obtain the number of training actions of the sports event that the trainer is training, and obtain the images of each training action. Extract the key point coordinates of the training actions based on each action, and compare them with the key point coordinates of the corresponding standard actions in the standard action image library; Use the Mahalanobis distance to calculate the deviation value between the key point coordinates of the training actions and the key point coordinates of the standard actions; Among them, the specific calculation process of the deviation value is as follows: Based on the extracted key point coordinates of the training actions, obtain the vector composed of the key point coordinates of the training actions , and the vector dimension corresponds to the number of key points; In the standard action image library, for each training action, extract the key point coordinates of the standard actions, and organize the key point coordinates of the obtained standard action samples into a matrix , this matrix has n rows and m columns, where n represents the standard action samples and m represents the number of key points, that is , and R is the set of real numbers; Calculate the mean vector of the key point data of the standard actions , and the calculation formula is: , where is the i-th row vector of the matrix , that is, the key point coordinate vector of the i-th standard action sample; Calculate the covariance matrix S, and the calculation formula is: , where is the deviation vector between the key point coordinate vector of the i-th and standard action samples and the mean vector, and the covariance matrix ; Calculate the Mahalanobis distance D between the key points of the training actions and the key points of the standard actions, and the calculation formula is: , where is the inverse matrix of the covariance matrix S; Then the Mahalanobis distance D is the deviation value between the key point coordinates of the training actions and the key point coordinates of the standard actions; Set the deviation threshold. Among them, the deviation threshold is set by implementers in this field based on the characteristics of sports events. For example, for gymnastics movements with high posture requirements, the threshold needs to be set smaller, while for relatively flexible movements, such as some shooting movements in basketball, the threshold can be set larger; Compare the deviation value with the deviation threshold. If the deviation value is less than or equal to the deviation threshold, mark this key point as a normal point. If the deviation value is greater than the deviation threshold, mark this key point as an error point; Step 2: Based on the error points, evaluate the severity of the training actions by analyzing and calculating the error data of the trainer during the training period, and construct a set of training actions to be corrected; In some embodiments, the total number of trainings for a sports event within a training cycle is set, where the number of trainings is summarized and set according to the characteristics of the training event and the training objectives; Based on any single training, obtain the error points of each training action, and count the number of error points and the number of key points of each training action; Calculate the ratio of the number of error points of the training action to the number of key points to obtain the error point occupancy ratio; Calculate the difference between the deviation value of each error point and the deviation threshold respectively, calculate the ratio of the obtained difference to the deviation threshold to obtain the error degree value, sum and average the error degree values of all error points to obtain the average error degree; Perform a weighted summation calculation on the error point occupancy ratio and the average error degree to obtain the error characterization value; Based on any single training action, count the number of times its error points appear, and calculate the ratio to the total number of trainings to obtain the error frequency occupancy ratio; Multiply the error characterization value by the error frequency occupancy ratio to obtain the action deviation evaluation coefficient; It should be noted that the function of calculating the action deviation evaluation coefficient is as follows: comprehensively considering multiple factors such as the error point occupancy ratio, the average error degree, and the error frequency occupancy ratio to calculate the action deviation evaluation coefficient, more comprehensively evaluating the quality and problem degree of the training action, being able to accurately screen out the training actions that need to be corrected, making the subsequent corrective training more targeted, improving the training efficiency and resource utilization efficiency, and using a specific numerical value, i.e., the action deviation evaluation coefficient, to measure the training action, with clear and intuitive results, facilitating the trainer and coach to understand and grasp the training situation; Set the action deviation evaluation coefficient threshold, mark the training actions less than or equal to the action deviation evaluation coefficient threshold as qualified actions, and mark the training actions greater than the action deviation evaluation coefficient as training actions to be corrected; Among them, the action deviation evaluation coefficient threshold is set by those skilled in the art according to experience; Obtain all the training actions to be corrected and construct a set of training actions to be corrected; The technical solution of this embodiment is as follows: First, image acquisition devices are set at different angles of the trainer to collect a sequence of training action images. The coordinates of the key points of the training action are extracted and compared with the corresponding coordinates in the standard action image library. The Mahalanobis distance is used to calculate the deviation value, which is then compared with the deviation threshold set according to the characteristics of the sports event to mark the error points. Then, based on these error points and combined with the total number of training times within the training cycle, a series of operations such as counting the number of error points, the number of key points, calculating the ratio of the number of error points, and the average value of the error degree are performed to obtain an error characterization value, which is then multiplied by the ratio of the number of error times to obtain an action deviation evaluation coefficient, so as to distinguish qualified actions from training actions to be corrected and construct a set of training actions to be corrected; Thus, the difference between the training action and the standard action can be obtained. Through quantitative calculation, the severity of the training action can be comprehensively evaluated, providing a clear and definite object for subsequent targeted corrective training, thereby effectively improving the accuracy and standardization of the training action, helping the trainer better master the action essentials, and enhancing the training effect. Embodiment 2

[0019] As Figure 1 shown, based on Embodiment 1, a method for identifying and correcting training actions based on image point comparison according to an embodiment of the present invention includes the following steps: Step Three: Based on the training actions to be corrected, perform corrective training, and obtain relevant corrective groups after the corrective training, and construct a query library for relevant corrective action groups; In some embodiments, the set of training actions to be corrected is marked as the first set of training actions to be corrected, and the training actions to be corrected in the first set of training actions to be corrected are arranged in order; Select the training action to be corrected ranked first, mark it as the first corrective training action, and perform the first corrective training. The corrective process is as follows: Obtain the deviation value sequence of each error point within the training cycle of the training action to be corrected ranked first, where the deviation value sequence is the deviation value corresponding to the error point marked for this key point within the training cycle; Perform mean processing on the deviation values corresponding to each error point to obtain the deviation mean; Based on the deviation mean, correct the error points of the training action to be corrected ranked first until the trainer reaches the standard training action for this training action to be corrected and stops the corrective training; After the first corrective training is completed, the trainer performs another training cycle of action training and constructs a second set of training actions to be corrected; It should be noted that the acquisition method of the training actions to be corrected is the same as the processes of Steps One and Two in the above Embodiment 1, and will not be elaborated here; Compare the training actions to be corrected in the first set of training actions to be corrected and the second set of training actions to be corrected, and extract the training actions that have been corrected among them; It should be noted that the training actions that have been corrected refer to the training actions that are in the first set of training actions to be corrected but not in the second set of training actions to be corrected; Exemplarily, if the first set of training actions to be corrected is A = {training action 2, training action 3, training action 4, training action 6, training action 7}, and the second set of training actions to be corrected is B = {training action 6, training action 7}, then the training actions that have been corrected are training action 2, training action 3, and training action 4; Mark the training actions that are consecutive to the first corrected action as related influencing training actions; Based on the first corrected training action and the related influencing training actions, obtain the deviation value sequences of the first corrected training action and the related influencing training actions respectively; Based on the combination analysis of the first corrected training action and any one of the related influencing training actions, mark it as the combination to be analyzed, and calculate the similarity representation values of the same key points in each combination to be analyzed; If the deviation value vector of the key point in the first corrected training action is and the deviation value vector of the key point in the related influencing training action is ; Its similarity value is calculated using cosine similarity, and the calculation formula is: The absolute value of the calculated cosine similarity value is the similarity representation value; Set a similarity threshold, mark the key points greater than the similarity threshold as similar points, and mark the key points less than or equal to the similarity threshold as non-similar points; Based on any one combination to be analyzed, count the number of the same key points, and count the number of similar points, and calculate the ratio of the number of similar points to the number of the same key points to obtain the ratio of similar points; Set a ratio threshold of similar points, mark the related influencing actions in the combination to be analyzed greater than the ratio threshold of similar points as related corrective actions, and mark the related influencing actions in the combination to be analyzed less than or equal to the ratio threshold of similar points as non-related corrective actions; Among them, the ratio threshold of similar points is set by those skilled in the art according to experience; Obtain all the related corrective actions, and based on all the related corrective actions, combine them with the first corrected training action and mark it as a related corrective action group; After the first corrective training, among the set of secondary corrective training actions, select the training action ranked first to be corrected, mark it as the secondary corrective training action, conduct secondary corrective training, repeat the process of identifying relevant influencing training actions, and obtain relevant corrective action groups; Repeat the above process of obtaining relevant corrective action groups until all training actions to be corrected meet the standards; Obtain all the above relevant corrective action groups and construct a relevant corrective action group query table; Obtain the relevant corrective action group query tables of several trainers and construct a relevant corrective action group query library; The relevant corrective action group query library is constructed based on the data of multiple trainers. It can provide personalized corrective advice combinations for individual trainers and also comprehensively consider the situations of different trainers, having certain generality and reference value; Step 4: Based on the constructed relevant corrective action group query library, provide corrective action selection suggestions for the trainer when there is an action deviation again, and obtain a corrective advice combination; In some embodiments, during the real-time training of the trainer, if a trainer action deviation alarm occurs, obtain the training action and mark it as the training action to be compared; Based on the training action to be compared, compare it with the relevant corrective action groups in the relevant corrective action group query library, obtain the relevant corrective action group with the training action to be compared as the first training action, and mark it as the action group to be selected; Based on the action group to be selected, obtain the error points of its first training action and the deviation value sequence corresponding to the error points; Compare the error points of the action to be compared with the error points of all the first training actions respectively to obtain the number of overlapping error points. Process the ratio of the number of overlapping error points to the number of error points of the action to be compared to obtain the overlapping ratio; Calculate the similarity representation values of the deviation value sequences of the error points of the action to be compared and the deviation value sequences of the error points of all the first training actions respectively; It should be noted that the calculation process of the similarity representation value is the same as the calculation process in Step 3 and will not be elaborated here; Perform a mean process on the similarity representation values of all overlapping error points to obtain a similarity degree value; Perform a weighted sum calculation on the overlapping ratio and the similarity degree value to obtain a comparison judgment value; Set a comparison judgment threshold, mark the action group to be selected corresponding to the first training action with a comparison judgment value greater than the threshold as the corrective advice combination, and mark the action group to be selected corresponding to the first training action with a comparison judgment value less than or equal to the threshold as a non-corrective advice combination; Among them, the comparison judgment threshold is set by those skilled in the art according to experience. Obtain a combination of correction suggestions. The trainer can select the combination of correction suggestions for combined correction to reduce ineffective or incorrect training methods, master the correct actions faster, shorten the time to reach the standard training actions, improve training efficiency, help the trainer achieve the training goals faster, enhance the training effect, correct more pertinently, and improve the accuracy and standardization of training. It should be noted that if there are multiple combinations of correction suggestions, select the combination of correction suggestions that contains the most training actions. The technical solution of this embodiment is as follows: First, mark the set of training actions to be corrected as the first set of training actions to be corrected and sort them. Select the first training action to be corrected for correction training until the action reaches the standard. Then, perform another training cycle of actions to construct a second set of training actions to be corrected. Compare the two sets to find the corrected actions, obtain the relevant influencing training actions, calculate the similarity representation values and the ratio of similar points at the same key points as the first corrected training action, screen out the relevant correction actions and combine them with the first corrected training action to form a relevant correction action group. Repeat this process until all the training actions to be corrected reach the standard, construct a query table of relevant correction action groups, and then collect the query tables of several trainers to construct a query library. When the trainer has an action deviation during real-time training, obtain the training action to be compared, compare it with the relevant correction action groups in the query library, calculate the coincidence ratio and the similarity degree value, and obtain the comparison judgment value by weighted summation. Select the combination of correction suggestions according to the comparison judgment threshold. Thus, it is possible to reduce ineffective or incorrect training methods, enable the trainer to master the correct actions faster, help the trainer achieve the training goals faster, and correct more pertinently through precise analysis and comparison. Embodiment 3

[0020] As Figure 2 shown, based on Embodiment 1 and Embodiment 2, a training action recognition and correction system based on image point position comparison according to an embodiment of the present invention specifically includes: Point position comparison module: Collect an image sequence of a trainer performing an action through an image acquisition device, extract the key points of the training action and compare them with the standard action points to obtain the error points. Action evaluation module: Based on the error points, evaluate the severity of the training action by analyzing and calculating the error data of the trainer during the training cycle, and construct a set of training actions to be corrected. Relevant action group construction module: Based on the training actions to be corrected, perform correction training, and obtain relevant correction groups after the correction training, and construct a query library of relevant correction action groups. Correction Suggestion Module: Based on building a query library of relevant correction action groups, it provides correction action selection suggestions for trainers when action deviations occur again and obtains a combination of correction suggestions.

[0021] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A training action recognition correction method based on image point comparison, characterized in that: The following steps are involved: The image sequence of the trainee performing the action is collected by the image acquisition device, and the key points of the training action are extracted and compared with the standard action points to obtain the error points; By analyzing and calculating the error data of the trainee's error points during the training cycle, the severity of the training action is evaluated and a set of training actions to be corrected is constructed; Performing corrective training on the corrective training action to be corrected, obtaining relevant corrective groups, and building a query library of relevant corrective action groups; Based on the construction of a query library of relevant corrective action groups, corrective action selection suggestions are provided to trainees when movement deviations occur again, and a correction suggestion combination is obtained.

2. The training action recognition and correction method based on image point comparison according to claim 1 is characterized in that: The process of obtaining the error point is as follows: Obtain the number of training movements of the sports item being trained by the trainee and the image of each training movement; Based on the image of each action, the key point coordinates of the training action are extracted and compared with the key point coordinates of the corresponding standard action in the standard action image library; The Mahalanobis distance is used to calculate the deviation between the key point coordinates of the training action and the key point coordinates of the standard action; The deviation value is compared with the deviation threshold. If the deviation value is greater than the deviation threshold, the key point is marked as an error point.

3. The training action recognition and correction method based on image point comparison according to claim 1 is characterized in that: The process of obtaining the set of training actions to be corrected is as follows: Analyze the error points of each training action to obtain the error representation value and the error frequency ratio; The error characterization value is multiplied by the error frequency ratio to obtain the action deviation evaluation coefficient; Mark the training movements that are greater than the movement deviation assessment coefficient as training movements to be corrected; Get all the training actions to be corrected and build a set of training actions to be corrected.

4. The training action recognition and correction method based on image point comparison according to claim 3 is characterized in that: The process of obtaining the error characterization value and the error frequency ratio value is as follows: Based on any training, obtain the error points of each training action, and count the number of error points and key points of each training action; Calculate the ratio of the number of error points and the number of key points in the training action to obtain the error point ratio; The deviation value of each error point is calculated separately from the deviation threshold, and the obtained difference is calculated as a ratio of the deviation threshold to obtain the error degree value. The error degree values ​​of all error points are summed and averaged to obtain the error degree mean; The error point ratio and the error degree mean are weighted and calculated to obtain the error representation value; The number of error points in each training action is counted, and the ratio is calculated with the total number of training times to obtain the percentage of error times.

5. The training action recognition and correction method based on image point comparison according to claim 1 is characterized in that: The process of constructing the query library of related correction action groups is as follows: Mark the set of training actions to be corrected as the first set of training actions to be corrected, and arrange them in order; The first corrective training action to be corrected is used as the first corrective training action, and the first corrective training is performed; After the first correction training, the trainee will conduct another training cycle of movements and build a second set of training movements to be corrected; Compare the first set of training actions to be corrected with the actions to be corrected in the second set of training actions to be corrected, and extract the corrected training actions; Mark the training movements that are consecutive to the first corrective movement as relevant impact training movements; Analyze and process the relevant influencing training actions and calculate the percentage of similar points; Mark the relevant influencing actions in the combination to be analyzed that are greater than the similar point ratio threshold as relevant corrective actions; Obtain all relevant corrective actions, combine them with the first corrective training action, and mark them as a relevant corrective action group; After the first correction training, in the set of secondary correction training actions, the first-ranked training action to be corrected is selected and marked as the secondary correction training action, and the secondary correction training is performed, and the process of identifying the relevant influencing training actions is repeated to obtain the relevant correction action group; Repeat the above process of obtaining relevant corrective action groups until all training actions to be corrected meet the standards; A related corrective action group query table is constructed for all related corrective action groups, and related corrective action group query tables of several trainees are obtained to construct a related corrective action group query library.

6. The training action recognition and correction method based on image point comparison according to claim 5 is characterized in that: The correction process is: Obtaining a sequence of deviation values ​​of each error point in a training cycle of the first-ranked training action to be corrected; Perform mean processing on the deviation value corresponding to each error point to obtain the deviation mean; Based on the deviation mean, the error point of the first-ranked action to be corrected is corrected until the trainee stops the corrective training when the action to be corrected reaches the standard training action.

7. The training action recognition and correction method based on image point comparison according to claim 5 is characterized in that: The process of obtaining the similarity point ratio value is as follows: Obtaining deviation value sequences of the first corrective training action and the related influencing training actions respectively; Perform a combination analysis based on the first correction training action and any related influencing training action, mark them as combinations to be analyzed, and calculate similar representation values ​​of the same key points in each combination to be analyzed; The similarity representation value is calculated by taking the absolute value of cosine similarity; Mark key points with a similarity greater than the similarity threshold as similar points; The number of identical key points and the number of similar points of each combination to be analyzed are counted, and the number of similar points and the number of identical key points are calculated and processed to obtain the ratio of similar points.

8. The training action recognition and correction method based on image point comparison according to claim 1 is characterized in that: The process of obtaining the correction suggestion combination is as follows: Obtain the overlap ratio and similarity value; The overlap ratio value and the similarity value are weighted and summed to obtain the comparison judgment value; A comparison judgment threshold is set, and the action group to be selected corresponding to the first training action greater than the comparison judgment value threshold is marked as a correction suggestion combination.

9. The training action recognition and correction method based on image point comparison according to claim 8 is characterized in that: The process of obtaining the overlap ratio value and the similarity value is as follows: If a trainee's action deviation alarm appears during real-time training, the corresponding training action will be marked as a training action to be compared; Based on the training action to be compared, compare with the relevant corrective action groups in the relevant corrective action group query library, obtain the relevant corrective action group with the training action to be compared as the first training action, and mark it as the action group to be selected; Obtain the error point of the first training action of the action group to be selected and the deviation value sequence corresponding to the error point; The error points of the action to be compared are overlapped with the error points of all the first training actions respectively, the number of overlapped error points is obtained, and the number of overlapped error points is ratioed with the number of error points of the action to be compared to obtain the overlap ratio value; The similarity representation values ​​of the error point deviation value sequence of the action to be compared and the error point deviation value sequence of all the first training actions are calculated respectively.

10. A training action recognition and correction system based on image point comparison, characterized in that: The system is used to execute the method described in any one of claims 1 to 9, and the system comprises: Point comparison module: collects the image sequence of the trainee's movements through the image acquisition device, extracts the key points of the training movements and compares them with the standard movement points to obtain the error points; Action evaluation module: By analyzing and calculating the error data of the trainee's error points during the training cycle, the severity of the training action is evaluated and a set of training actions to be corrected is constructed; Related action group construction module: perform correction training on the corrective training action, obtain the related correction group, and build a related correction action group query library; Correction suggestion module: Based on building a query library of relevant corrective action groups, it provides corrective action selection suggestions for trainees when movement deviations occur again, and obtains a combination of correction suggestions.

Images