AI virtual coach training system based on standard action matching and deviation feedback

Abstract

The invention discloses an AI virtual coach training system based on standard action matching and deviation feedback, which relates to the technical field of AI virtual coach training systems and comprises a user modeling module, an action acquisition module, a template matching module, a deviation calculation module, a feedback generation module, an interactive presentation module and a learning optimization module. The user modeling module is used for modeling a registered user by adopting a body parameter acquisition and health data analysis method to obtain a user personalized feature vector; the action acquisition module is used for capturing actions executed by a user in real time by adopting a multi-source sensor fusion method to obtain a time sequence containing key point coordinates; and the template matching module is used for comparing the time sequence of the key point coordinates with corresponding actions in a preset standard action template library by adopting an improved dynamic time warping (DTW) algorithm to obtain an optimal matching path and a corresponding minimum matching cost.

Description

Technical Field

[0001] The present invention relates to the technical field of AI virtual coach training systems, and in particular to an AI virtual coach training system based on standard action matching and deviation feedback. Background Art

[0002] The AI virtual coach training system is an intelligent system that uses artificial intelligence technology to simulate the functions of a human coach. Therefore, how to use advanced technologies to improve the intelligence level and safety of the AI virtual coach training system has become one of the most pressing issues to be addressed.

[0003] In the field of AI virtual coach training systems, existing virtual coach systems often use universal motion templates and feedback mechanisms, failing to fully consider individual differences among users. This results in inaccurate and ineffective training suggestions. Traditional methods are also easily affected by environmental factors during the motion capture process, leading to inaccurate recognition of key point coordinates, which in turn affects subsequent deviation analysis and feedback generation. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an AI virtual coach training system based on standard action matching and deviation feedback to solve the problem that existing virtual coach systems often adopt universal action templates and feedback mechanisms, fail to fully consider the individual differences of users, resulting in the provided training suggestions not being accurate and effective. In addition, traditional methods are easily affected by environmental factors during the motion capture process, resulting in inaccurate recognition of key point coordinates, which in turn affects subsequent deviation analysis and feedback generation.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides an AI virtual coach training system based on standard action matching and deviation feedback, which includes: User modeling module, action acquisition module, template matching module, deviation calculation module, feedback generation module, interactive presentation module and learning optimization module; The user modeling module is used to model the registered user using a body parameter collection and health data analysis method to obtain a user personalized feature vector; The motion acquisition module is used to capture the user's motions in real time using a multi-source sensor fusion method to obtain a time series containing key point coordinates; The template matching module is used to compare the time series of key point coordinates with the corresponding actions in the preset standard action template library using the improved dynamic time warping (DTW) algorithm to obtain the optimal matching path and the corresponding minimum matching cost; The deviation calculation module is used to perform frame-by-frame deviation analysis on the user action based on the matching path and the standard action template to obtain the three-dimensional space deviation vector of each key point and its comprehensive deviation score; The feedback generation module is used to generate personalized correction suggestions based on the comprehensive deviation score and the user's personalized feature vector using a rule-model hybrid reasoning method; The interactive presentation module is used to present personalized correction suggestions in a visual manner on the user terminal interface, guide the user to adjust the action in real time, and receive interactive operations such as user confirmation or correction feedback; The learning optimization module is used to update the template matching strategy, deviation calculation weight and feedback generation logic based on the user's historical training records and feedback responses using an online incremental learning method to continuously optimize system performance.

[0007] As a preferred solution of the AI virtual coach training system based on standard action matching and deviation feedback of the present invention, the method of using body parameter collection and health data analysis to model the registered user and obtain the user's personalized feature vector is specifically carried out as follows: Collecting the physical information actively filled in by the registered user to obtain a set of original physiological indicator data, wherein the original physiological indicator data set includes height, weight, and body fat percentage; The user's smart bracelet or watch acquires real-time health data to obtain a set of basic health parameters, including resting heart rate and maximum oxygen uptake; A questionnaire evaluation mechanism is used to analyze the health status questionnaire submitted by the user to obtain a set of health risk identifiers, wherein the set of health risk identifiers includes chronic disease flags and recent sports injury records; Using a behavior preference selection interface to guide the user to complete the training habit setting, and obtain a behavior preference parameter set, wherein the behavior preference parameter set includes a training frequency preference and a feedback method preference; Identifying and scoring the user's movements during a standardized flexibility test and a static balance test to obtain a set of functional ability parameters, the set of functional ability parameters including a flexibility score and a balance ability score; The height and weight were normalized using a linear normalization function to obtain normalized height and normalized weight values respectively; The body fat percentage is normalized using a scaling function to obtain a normalized body fat value; The relative oxygen uptake function is used to normalize the maximum oxygen uptake to obtain the normalized oxygen uptake value A three-choice classification function was used to quantify the training frequency preference and feedback mode preference, and normalized frequency factors and mode factors were obtained respectively. The normalized body parameters are integrated using a multi-dimensional weighted fusion formula to obtain body dimension sub-vectors; The fusion formula with health inhibition factor was used to integrate the normalized oxygen uptake value, flexibility score and balance ability score to obtain the functional dimension sub-vector; The preference quantification formula is used to fuse the training frequency preference and feedback mode preference to obtain the behavior dimension sub-vector; Use vector concatenation to concatenate body dimension sub-vectors , functional dimension subvector and the behavioral dimension subvector Merge them sequentially to get the complete user personalized feature vector, which is expressed as: ; ; in, Represents a vector concatenation operation.

[0008] As a preferred solution of the AI virtual coach training system based on standard action matching and deviation feedback of the present invention, wherein: the multi-source sensor fusion method is used to capture the user's actions in real time to obtain a time series containing key point coordinates, and the specific steps are as follows: The deep learning model HRNet is used to estimate human posture in the video streams collected by the camera and depth camera, and the coordinate set of key points of the human body in each frame is extracted; The accelerometer and gyroscope integrated algorithm is used to analyze the inertial measurement unit (IMU) data collected by the wearable device, calculate the angle change rate and angular acceleration of each joint, and convert it into a posture adjustment vector in three-dimensional space; A spatiotemporal alignment algorithm is used to align the key point coordinates and attitude adjustment vectors obtained from different sensors in time and space, eliminating data deviations caused by delays or position differences between sensors and generating a unified timestamp. The Kalman filter is used to smooth the key point coordinates and posture adjustment vectors after spatiotemporal alignment to reduce noise interference, improve data accuracy, and output a smoothed human posture dataset. The quaternion interpolation algorithm SLERP is used to complete the missing key points between adjacent frames in the smoothed human posture dataset to ensure that each frame of data is complete and continuous, forming a complete motion trajectory; The time series construction function is used to convert the completed motion trajectory into a time series of key point coordinates indexed by time; The normalization processing function is used to normalize the key point coordinates in the time series to adapt it to the action template library of users with different body shapes, and the final standardized key point coordinate time series is generated.

[0009] As a preferred solution of the AI virtual coach training system based on standard action matching and deviation feedback described in the present invention, the improved dynamic time warping (DTW) algorithm is used to compare the time series of key point coordinates with the corresponding actions in the preset standard action template library to obtain the optimal matching path and the corresponding minimum matching cost. The specific steps are as follows: Using videos of standard movements performed by professional athletes or coaches as input, a deep learning model is used to extract the key frames of each movement and the key coordinates of the human body posture to build a standard movement template library. The weighted distance function is used to locally optimize the adjacent key points in each standard action template. The expression is: ; in, and Represent the three-dimensional coordinates of two key points respectively, and Represent the direction angles of the two key points, and is the weight coefficient; The standardized key point coordinate time series is used as the data input for the user's action, which is then smoothed to remove noise interference and generate a smoothed user action time series. The improved dynamic time warping (DTW) algorithm is used to calculate the distance matrix between the user's action time series and each item in the standard action template library. ; The path cumulative cost function is used to correct the cumulative cost of each step, and the expression is: ; in, From the starting point to The cumulative cost of the point, is the angle penalty factor, For the The weight of the key points, and Represents user actions and standard action templates in The position at the key point; Use the backtracking algorithm to reversely search the minimum cumulative cost path in the cumulative cost matrix from the end point to obtain the optimal matching path and record all points on the path ; The minimum cumulative cost formula is used to calculate the average cost of the optimal matching path as the final minimum matching cost. , the expression is: ; in, represents the number of points on the optimal matching path, Represents the cumulative cost of a point on the path.

[0010] As a preferred solution of the AI virtual coach training system based on standard action matching and deviation feedback of the present invention, the specific steps are as follows: based on the matching path and the standard action template, the user action is subjected to frame-by-frame deviation analysis to obtain the three-dimensional space deviation vector of each key point and its comprehensive deviation score. The three-dimensional space deviation vector formula is used to calculate the key point deviation of each pair of frames on the matching path to obtain the three-dimensional space deviation vector of each key point. The expression is: ; in, Represents the user action time series Frame The three-dimensional coordinates of the key points, Indicates the standard action template Frame The three-dimensional coordinates of the key points, Indicates the Keypoints in frame The three-dimensional space deviation vector on ; For each pair of frames on the matching path A comprehensive deviation score is calculated to obtain a comprehensive deviation score for the frame, and the comprehensive deviation score is normalized.

[0011] As a preferred solution of the AI virtual coach training system based on standard action matching and deviation feedback described in the present invention, wherein: based on the comprehensive deviation score and the user's personalized feature vector, a rule-model hybrid reasoning method is used to generate personalized correction suggestions, the specific steps are as follows: A lightweight neural network model is used to build a personalized feedback prediction model; The threshold judgment function is used for evaluation, and the expression is: ; in, The action qualification threshold preset by the system, if > , indicating that there is an obvious deviation in the action and correction is required; Use the rule matching algorithm to match the condition items in the rule base one by one, and extract all the rule sets that meet the current user status; The neural network forward propagation function is used to predict the user status model and generate a preliminary correction suggestion vector, which is expressed as: ; in, represents the output layer activation function, represents the hidden layer activation function, 、 are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. 、 are the bias terms, Indicates that the user feature vector and the deviation score are concatenated into the input vector; A weighted fusion function is used to fuse the rule inference results and model prediction results to generate the final personalized correction suggestion vector.

[0012] As a preferred solution of the AI virtual coach training system based on standard movement matching and deviation feedback described in the present invention, the personalized correction suggestions are presented in a visual manner on the user terminal interface, guiding the user to adjust the movement in real time, and receiving the user's confirmation or correction feedback interactive operation. The specific steps are as follows: A text conversion function is used to convert the final personalized correction suggestion vector into a text description suitable for different presentation forms; A graphics rendering engine is used to draw the user's current and ideal postures in three-dimensional space based on the key point coordinates and deviation vectors of the user's movements. An arrow is added between the two to indicate the direction that needs to be adjusted. Use speech synthesis technology to convert it into voice prompts and adjust the voice tone according to the language style set by the user; Use split-screen or multi-view mode to display the user's current posture, ideal posture, and correction suggestions on the same interface so that the user can quickly understand the adjustments needed; The use of gesture recognition or voice command interfaces allows users to confirm the correction suggestions given by the system through simple gestures or voice commands.

[0013] As a preferred solution of the AI virtual coach training system based on standard action matching and deviation feedback described in the present invention, wherein: based on the user's historical training records and feedback responses, an online incremental learning method is used to update the template matching strategy, deviation calculation weight and feedback generation logic to continuously optimize system performance. The specific steps are as follows: Use data cleaning algorithms for preprocessing to remove outliers and duplicate records to obtain a clean data set; A feature extraction algorithm is used to extract key feature vectors, which include the user's height ratio, weight ratio, body fat level, cardiopulmonary endurance, flexibility, and balance stability personalized features; Use label generation algorithm to generate corresponding label vectors based on user feedback responses; The template matching strategy is updated using the gradient descent method, and each training record is iteratively optimized by minimizing the loss function; Use the weighted error square sum formula to update the deviation calculation weight vector; The Q-learning algorithm in reinforcement learning is used to update the feedback generation logic. Each state-action is evaluated by maximizing the cumulative reward function, which is expressed as: ; in, For instant rewards, is the discount factor, For the next state; The updated template matching strategy, deviation calculation weight vector, and feedback generation logic are encapsulated into a structured data format and passed to each module of the system for real-time updating of the system configuration.

[0014] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the AI virtual coach training system based on standard action matching and deviation feedback as described in the first aspect of the present invention is implemented.

[0015] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the AI virtual coach training system based on standard action matching and deviation feedback as described in the first aspect of the present invention.

[0016] The beneficial effects of the present invention are as follows: human motion data is synchronously collected through multiple devices such as cameras, depth cameras, and IMU sensors, and after spatiotemporal alignment, Kalman filtering, and smooth interpolation processing, a standardized key point coordinate time series is formed, thereby achieving high-precision and continuous capture of human motion in three-dimensional space, providing accurate and stable raw data input for subsequent motion comparison and deviation analysis. Based on the standard motion template library, an improved dynamic time warping DTW algorithm is adopted, combined with a weighted distance function and an angle penalty factor, to compare the user motion time series, output the optimal matching path and the minimum matching cost, achieve nonlinear time alignment between user motion and standard motion, solve the matching problem caused by inconsistent motion rhythm, and provide a reliable path basis for deviation analysis. Based on the matching path and the standard motion template, the three-dimensional spatial deviation of each frame of user motion is calculated, and the weight coefficient and the angle deviation term are combined to generate a comprehensive deviation score, achieving a quantitative evaluation from local key point deviation to overall motion quality, and providing data support for personalized correction suggestions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Schematic diagram of the AI virtual coach training system based on standard action matching and deviation feedback in Example 1. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0022] Example 1, with reference to Figure 1, which is the first embodiment of the present invention, provides an AI virtual coach training system based on standard action matching and deviation feedback, including: User modeling module, action acquisition module, template matching module, deviation calculation module, feedback generation module, interactive presentation module and learning optimization module; The user modeling module is used to model registered users using body parameter collection and health data analysis methods to obtain user personalized feature vectors; Furthermore, the physical information actively filled in by registered users is collected to obtain a set of original physiological indicator data, including height, weight, and body fat percentage; The user's smart bracelet or watch acquires real-time health data and obtains a set of basic health parameters, including resting heart rate and maximum oxygen uptake; A questionnaire evaluation mechanism is used to analyze the health status questionnaire submitted by users to obtain a set of health risk identifiers, which includes chronic disease markers and recent sports injury records; A behavior preference selection interface is used to guide the user to complete the training habit setting and obtain a behavior preference parameter set, which includes a training frequency preference and a feedback method preference; Identifying and scoring the user's movements during standardized flexibility tests and static balance tests to obtain a set of functional ability parameters, including a flexibility score and a balance ability score; The height and weight were normalized using a linear normalization function to obtain normalized height and normalized weight values respectively; The body fat percentage is normalized using a scaling function to obtain a normalized body fat value; The relative oxygen uptake function is used to normalize the maximum oxygen uptake to obtain the normalized oxygen uptake value A three-choice classification function was used to quantify the training frequency preference and feedback mode preference, and normalized frequency factors and mode factors were obtained respectively. The normalized body parameters are integrated using a multi-dimensional weighted fusion formula to obtain body dimension sub-vectors; The fusion formula with health inhibition factor was used to integrate the normalized oxygen uptake value, flexibility score and balance ability score to obtain the functional dimension sub-vector; The preference quantification formula is used to fuse the training frequency preference and feedback mode preference to obtain the behavior dimension sub-vector; Use vector concatenation to concatenate body dimension sub-vectors , functional dimension sub-vector and the behavioral dimension subvector Merge them sequentially to get the complete user personalized feature vector, which is expressed as: ; ; in, Represents vector concatenation operation; It should be noted that through multi-dimensional data fusion and normalization processing, a scalable and computable personalized feature vector is constructed to provide basic support for subsequent action recognition and feedback generation. This method not only improves the system's personalized service capabilities, but also enhances its adaptability to users with different physiques, especially in terms of sports risk control.

[0023] The motion acquisition module is used to capture the user's actions in real time using a multi-source sensor fusion method to obtain a time series containing key point coordinates; Furthermore, the deep learning model HRNet is used to estimate human posture in the video streams collected by the camera and depth camera, and extract the coordinate set of key points of the human body in each frame of the image; The accelerometer and gyroscope integrated algorithm is used to analyze the inertial measurement unit (IMU) data collected by the wearable device, calculate the angle change rate and angular acceleration of each joint, and convert it into a posture adjustment vector in three-dimensional space; A spatiotemporal alignment algorithm is used to align the key point coordinates and attitude adjustment vectors obtained from different sensors in time and space, eliminating data deviations caused by delays or position differences between sensors and generating a unified timestamp. The Kalman filter is used to smooth the key point coordinates and posture adjustment vectors after spatiotemporal alignment to reduce noise interference, improve data accuracy, and output a smoothed human posture dataset. The quaternion interpolation algorithm SLERP is used to complete the missing key points between adjacent frames in the smoothed human posture dataset to ensure that each frame of data is complete and continuous, forming a complete motion trajectory; The time series construction function is used to convert the completed motion trajectory into a time series of key point coordinates indexed by time; A normalization function is used to normalize the key point coordinates in the time series to adapt it to the motion template library of users of different body shapes, and the final normalized key point coordinate time series is generated; It should be noted that the adoption of a multi-source sensor fusion strategy, combining the advantages of vision and inertial sensing, effectively solves the limitations of a single device in complex environments such as occlusion and lighting changes. At the same time, through spatiotemporal alignment and interpolation completion mechanisms, the continuity and stability of motion capture are ensured, providing reliable data input for high-precision motion analysis.

[0024] The template matching module uses the improved dynamic time warping (DTW) algorithm to compare the time series of key point coordinates with the corresponding actions in the preset standard action template library to obtain the optimal matching path and the corresponding minimum matching cost; Furthermore, by using videos of standard movements performed by professional athletes or coaches as input, the deep learning model is used to extract the key frames of each movement and the key coordinates of the human body posture to build a standard movement template library; The weighted distance function is used to locally optimize the adjacent key points in each standard action template. The expression is: ; in, and Represent the three-dimensional coordinates of two key points respectively, and Represent the direction angles of the two key points, and is the weight coefficient; The standardized key point coordinate time series is used as the data input for the user's action, which is then smoothed to remove noise interference and generate a smoothed user action time series. The improved dynamic time warping (DTW) algorithm is used to calculate the distance matrix between the user's action time series and each item in the standard action template library. ; The path cumulative cost function is used to correct the cumulative cost of each step, and the expression is: ; in, From the starting point to The cumulative cost of the point, is the angle penalty factor, For the The weight of the key points, and Represents user actions and standard action templates in The position at the key point; Use the backtracking algorithm to reversely search the minimum cumulative cost path in the cumulative cost matrix from the end point to obtain the optimal matching path and record all points on the path ; The minimum cumulative cost formula is used to calculate the average cost of the optimal matching path as the final minimum matching cost. , the expression is: ; in, represents the number of points on the optimal matching path, Represents the cumulative cost of a point on the path; It should be noted that the improved DTW algorithm introduces an angle penalty factor and a key point weight adjustment mechanism, which enables the system to maintain a high matching accuracy when facing actions with inconsistent rhythms or large differences in posture. This method is particularly suitable for training scenarios for non-professional users and significantly improves the robustness and generalization ability of action recognition.

[0025] Deviation calculation module, used to perform frame-by-frame deviation analysis of user actions based on matching paths and standard action templates, and obtain the three-dimensional spatial deviation vector and comprehensive deviation score of each key point; Furthermore, the three-dimensional space deviation vector formula is used to calculate the key point deviation of each pair of frames on the matching path to obtain the three-dimensional space deviation vector of each key point, which is expressed as: ; in, Represents the user action time series Frame The three-dimensional coordinates of the key points, Indicates the standard action template Frame The three-dimensional coordinates of the key points, Indicates the Keypoints in frame The three-dimensional space deviation vector on ; For each pair of frames on the matching path Performing a comprehensive deviation score calculation to obtain a comprehensive deviation score for the frame, and performing normalization processing on the comprehensive deviation score; It should be noted that the calculation method of the frame-by-frame three-dimensional spatial deviation vector can accurately locate the specific problem points in the action execution process, while the comprehensive deviation score reflects the quality of the action as a whole. The combination of fine-grained and global evaluation provides a scientific basis for subsequent feedback generation and enhances the system's closed-loop feedback capability.

[0026] The feedback generation module is used to generate personalized correction suggestions based on the comprehensive deviation score and the user's personalized feature vector using a rule-model hybrid reasoning method; Furthermore, a lightweight neural network model is used to build a personalized feedback prediction model; The threshold judgment function is used for evaluation, and the expression is: ; in, The action qualification threshold preset by the system, if > , indicating that there is an obvious deviation in the action and correction is required; Use the rule matching algorithm to match the condition items in the rule base one by one, and extract all the rule sets that meet the current user status; The neural network forward propagation function is used to predict the user status model and generate a preliminary correction suggestion vector, which is expressed as: ; in, represents the output layer activation function, represents the hidden layer activation function, 、 are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. 、 are the bias terms, Indicates that the user feature vector and the deviation score are concatenated into the input vector; A weighted fusion function is used to fuse the rule inference results and model prediction results to generate the final personalized correction suggestion vector; It should be noted that the rule-model hybrid reasoning mechanism takes into account both expert experience and the flexibility of machine learning, making the correction suggestions both authoritative and personalized. The design of the weighted fusion function further enhances the system's adaptability to different user states and achieves more efficient and targeted training guidance.

[0027] The interactive presentation module is used to present personalized correction suggestions in a visual manner on the user terminal interface, guide the user to adjust the action in real time, and receive interactive operations such as user confirmation or correction feedback; Furthermore, a text conversion function is used to convert the final personalized correction suggestion vector into a text description suitable for different presentation forms; A graphics rendering engine is used to draw the user's current and ideal postures in three-dimensional space based on the key point coordinates and deviation vectors of the user's movements. An arrow is added between the two to indicate the direction that needs to be adjusted. Use speech synthesis technology to convert it into voice prompts and adjust the voice tone according to the language style set by the user; Use split-screen or multi-view mode to display the user's current posture, ideal posture, and correction suggestions on the same interface so that the user can quickly understand the adjustments needed; Using gesture recognition or voice command interfaces allows users to confirm correction suggestions given by the system through simple gestures or voice commands; It should be noted that the visual guidance mechanism, combined with various forms such as graphic comparison, text description and voice prompts, significantly improves the user's efficiency in understanding movement deviations and their willingness to correct them. In addition, the introduction of gesture or voice confirmation functions enhances the convenience of human-computer interaction and provides a strong guarantee for improving user experience and training participation.

[0028] The learning optimization module is used to continuously optimize system performance by updating template matching strategies, deviation calculation weights, and feedback generation logic using online incremental learning methods based on users' historical training records and feedback responses; Furthermore, data cleaning algorithms are used for preprocessing to remove outliers and duplicate records to obtain a clean data set; A feature extraction algorithm is used to extract key feature vectors, which include the user's height ratio, weight ratio, body fat level, cardiopulmonary endurance, flexibility, and balance stability. Use label generation algorithm to generate corresponding label vectors based on user feedback responses; The template matching strategy is updated using the gradient descent method, and each training record is iteratively optimized by minimizing the loss function; Use the weighted error square sum formula to update the deviation calculation weight vector; The Q-learning algorithm in reinforcement learning is used to update the feedback generation logic. Each state-action is evaluated by maximizing the cumulative reward function, which is expressed as: ; in, For instant rewards, is the discount factor, For the next state; The updated template matching strategy, deviation calculation weight vector, and feedback generation logic are encapsulated into a structured data format and passed to each system module for real-time system configuration updates. It should be noted that the online incremental learning mechanism enables the system to continuously absorb user training records and feedback responses, dynamically adjust template matching strategies, deviation calculation weights and feedback generation logic, thereby continuously improving the system's intelligence level and personalized service capabilities. This mechanism effectively extends the product life cycle, reduces the frequency of manual intervention, and improves the system's sustainable development capabilities.

[0029] This embodiment also provides a computer device, which is suitable for the case of an AI virtual coach training system based on standard action matching and deviation feedback, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the AI virtual coach training system based on standard action matching and deviation feedback proposed in the above embodiment.

[0030] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0031] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI virtual coach training system based on standard action matching and deviation feedback as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0032] In summary, the present invention synchronously collects human motion data through multiple devices such as cameras, depth cameras, and IMU sensors. After spatiotemporal alignment, Kalman filtering, and smooth interpolation processing, a standardized key point coordinate time series is formed, which realizes high-precision and continuous capture of human motion in three-dimensional space, and provides accurate and stable raw data input for subsequent motion comparison and deviation analysis. Based on the standard motion template library, an improved dynamic time warping DTW algorithm is adopted, combined with a weighted distance function and an angle penalty factor, to compare the user motion time series, output the optimal matching path and the minimum matching cost, realize nonlinear time alignment between user motion and standard motion, solve the matching problem caused by inconsistent motion rhythm, and provide a reliable path basis for deviation analysis. Based on the matching path and the standard motion template, the three-dimensional spatial deviation of each frame of user motion is calculated, and the weight coefficient and the angle deviation term are combined to generate a comprehensive deviation score, which realizes quantitative evaluation from local key point deviation to overall motion quality, and provides data support for personalized correction suggestions.

[0033] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An AI virtual coach training system based on standard action matching and deviation feedback, characterized by: include: User modeling module, action acquisition module, template matching module, deviation calculation module, feedback generation module, interactive presentation module and learning optimization module; The user modeling module is used to model the registered user using a body parameter collection and health data analysis method to obtain a user personalized feature vector; The motion acquisition module is used to capture the user's actions in real time using a multi-source sensor fusion method to obtain a time series containing key point coordinates; The template matching module is used to compare the time series of key point coordinates with the corresponding actions in the preset standard action template library using the improved dynamic time warping (DTW) algorithm to obtain the optimal matching path and the corresponding minimum matching cost; The deviation calculation module is used to perform frame-by-frame deviation analysis on the user action based on the matching path and the standard action template to obtain the three-dimensional space deviation vector of each key point and its comprehensive deviation score; The feedback generation module is used to generate personalized correction suggestions based on the comprehensive deviation score and the user's personalized feature vector using a rule-model hybrid reasoning method; The interactive presentation module is used to present personalized correction suggestions in a visual manner on the user terminal interface, guide the user to adjust the action in real time, and receive interactive operations such as user confirmation or correction feedback; The learning optimization module is used to update the template matching strategy, deviation calculation weight and feedback generation logic based on the user's historical training records and feedback responses using an online incremental learning method to continuously optimize system performance.

2. The AI virtual coach training system based on standard movement matching and deviation feedback according to claim 1, characterized in that: The method of using body parameter collection and health data analysis to model registered users and obtain user personalized feature vectors is specifically performed as follows: Collecting the physical information actively filled in by the registered user to obtain a set of original physiological indicator data, wherein the original physiological indicator data set includes height, weight, and body fat percentage; The user's smart bracelet or watch acquires real-time health data to obtain a set of basic health parameters, including resting heart rate and maximum oxygen uptake; A questionnaire evaluation mechanism is used to analyze the health status questionnaire submitted by the user to obtain a set of health risk identifiers, wherein the set of health risk identifiers includes chronic disease flags and recent sports injury records; Using a behavior preference selection interface to guide the user to complete the training habit setting, and obtain a behavior preference parameter set, wherein the behavior preference parameter set includes a training frequency preference and a feedback method preference; Identifying and scoring the user's movements during a standardized flexibility test and a static balance test to obtain a set of functional ability parameters, the set of functional ability parameters including a flexibility score and a balance ability score; The height and weight were normalized using a linear normalization function to obtain normalized height and normalized weight values respectively; The body fat percentage is normalized using a scaling function to obtain a normalized body fat value; The relative oxygen uptake function is used to normalize the maximum oxygen uptake to obtain the normalized oxygen uptake value A three-choice classification function was used to quantify the training frequency preference and feedback mode preference, and normalized frequency factors and mode factors were obtained respectively. The normalized body parameters are integrated using a multi-dimensional weighted fusion formula to obtain body dimension sub-vectors; The fusion formula with health inhibition factor was used to integrate the normalized oxygen uptake value, flexibility score and balance ability score to obtain the functional dimension sub-vector; The preference quantification formula is used to fuse the training frequency preference and feedback mode preference to obtain the behavior dimension sub-vector; Use vector concatenation to concatenate body dimension sub-vectors , functional dimension subvector and the behavioral dimension subvector Merge them sequentially to get the complete user personalized feature vector, which is expressed as: ； ； in, Represents a vector concatenation operation.

3. The AI virtual coach training system based on standard movement matching and deviation feedback according to claim 2, characterized in that: The multi-source sensor fusion method is used to capture the user's actions in real time to obtain a time series containing key point coordinates. The specific steps are as follows: The deep learning model HRNet is used to estimate human posture in the video streams collected by the camera and depth camera, and the coordinate set of key points of the human body in each frame is extracted; The accelerometer and gyroscope integrated algorithm is used to analyze the inertial measurement unit (IMU) data collected by the wearable device, calculate the angle change rate and angular acceleration of each joint, and convert it into a posture adjustment vector in three-dimensional space; A spatiotemporal alignment algorithm is used to align the key point coordinates and attitude adjustment vectors obtained from different sensors in time and space, eliminating data deviations caused by delays or position differences between sensors and generating a unified timestamp. The Kalman filter is used to smooth the key point coordinates and posture adjustment vectors after spatiotemporal alignment to reduce noise interference, improve data accuracy, and output a smoothed human posture dataset. The quaternion interpolation algorithm SLERP is used to complete the missing key points between adjacent frames in the smoothed human posture dataset to ensure that each frame of data is complete and continuous, forming a complete motion trajectory; The time series construction function is used to convert the completed motion trajectory into a time series of key point coordinates indexed by time; The normalization processing function is used to normalize the key point coordinates in the time series to adapt it to the action template library of users with different body shapes, and the final standardized key point coordinate time series is generated.

4. The AI virtual coach training system based on standard movement matching and deviation feedback according to claim 3, characterized in that: The improved dynamic time warping (DTW) algorithm is used to compare the time series of key point coordinates with the corresponding actions in the preset standard action template library to obtain the optimal matching path and the corresponding minimum matching cost. The specific steps are as follows: Using videos of standard movements performed by professional athletes or coaches as input, a deep learning model is used to extract the key frames of each movement and the key coordinates of the human body posture to build a standard movement template library. The weighted distance function is used to locally optimize the adjacent key points in each standard action template. The expression is: ; in, and Represent the three-dimensional coordinates of two key points respectively, and Represent the direction angles of the two key points, and is the weight coefficient; The standardized key point coordinate time series is used as the data input for the user's action, which is then smoothed to remove noise interference and generate a smoothed user action time series. The improved dynamic time warping (DTW) algorithm is used to calculate the distance matrix between the user's action time series and each item in the standard action template library. ; The path cumulative cost function is used to correct the cumulative cost of each step, and the expression is: ; in, From the starting point to The cumulative cost of the point, is the angle penalty factor, For the The weight of the key points, and Represents user actions and standard action templates in The position at the key point; Use the backtracking algorithm to reversely search the minimum cumulative cost path in the cumulative cost matrix from the end point to obtain the optimal matching path and record all points on the path ; The minimum cumulative cost formula is used to calculate the average cost of the optimal matching path as the final minimum matching cost. , the expression is: ; in, represents the number of points on the optimal matching path, Represents the cumulative cost of a point on the path.

5. The AI virtual coach training system based on standard movement matching and deviation feedback according to claim 4, characterized in that: Based on the matching path and the standard action template, the user action is analyzed frame by frame to obtain the three-dimensional space deviation vector of each key point and its comprehensive deviation score. The specific steps are as follows: The three-dimensional space deviation vector formula is used to calculate the key point deviation of each pair of frames on the matching path to obtain the three-dimensional space deviation vector of each key point. The expression is: ； in, Represents the user action time series Frame The three-dimensional coordinates of the key points, Indicates the standard action template Frame The three-dimensional coordinates of the key points, Indicates the Keypoints in frame The three-dimensional space deviation vector on ; For each pair of frames on the matching path A comprehensive deviation score is calculated to obtain a comprehensive deviation score for the frame, and the comprehensive deviation score is normalized.

6. The AI virtual coach training system based on standard movement matching and deviation feedback according to claim 5, characterized in that: The method of generating personalized correction suggestions based on the comprehensive deviation score and the user's personalized feature vector by using a rule-model hybrid reasoning method is as follows: A lightweight neural network model is used to build a personalized feedback prediction model; The threshold judgment function is used for evaluation, and the expression is: ; in, The action qualification threshold preset by the system, if > , indicating that there is an obvious deviation in the action and correction is required; Use the rule matching algorithm to match the condition items in the rule base one by one and extract all the rule sets that meet the current user status; The neural network forward propagation function is used to predict the user status model and generate a preliminary correction suggestion vector, which is expressed as: ; in, represents the output layer activation function, represents the hidden layer activation function, 、 are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. 、 are the bias terms, Indicates that the user feature vector and the deviation score are concatenated into the input vector; A weighted fusion function is used to fuse the rule inference results and model prediction results to generate the final personalized correction suggestion vector.

7. The AI virtual coach training system based on standard movement matching and deviation feedback according to claim 6, characterized in that: The personalized correction suggestions are presented in a visual manner on the user terminal interface, guiding the user to adjust the action in real time, and receiving the user's confirmation or correction feedback interactive operation. The specific steps are as follows: A text conversion function is used to convert the final personalized correction suggestion vector into a text description suitable for different presentation forms; A graphics rendering engine is used to draw the user's current and ideal postures in three-dimensional space based on the key point coordinates and deviation vectors of the user's movements. An arrow is added between the two to indicate the direction that needs to be adjusted. Use speech synthesis technology to convert it into voice prompts and adjust the voice tone according to the language style set by the user; Use split-screen or multi-view mode to display the user's current posture, ideal posture, and correction suggestions on the same interface so that the user can quickly understand the adjustments needed; The use of gesture recognition or voice command interfaces allows users to confirm the correction suggestions given by the system through simple gestures or voice commands.

8. The AI virtual coach training system based on standard movement matching and deviation feedback according to claim 7, characterized in that: Based on the user's historical training records and feedback responses, the online incremental learning method is used to update the template matching strategy, deviation calculation weight and feedback generation logic to continuously optimize the system performance. The specific steps are as follows: Use data cleaning algorithms for preprocessing to remove outliers and duplicate records to obtain a clean data set; A feature extraction algorithm is used to extract key feature vectors, which include the user's height ratio, weight ratio, body fat level, cardiopulmonary endurance, flexibility, and balance stability personalized features; Use label generation algorithm to generate corresponding label vectors based on user feedback responses; The template matching strategy is updated using the gradient descent method, and each training record is iteratively optimized by minimizing the loss function; Use the weighted error square sum formula to update the deviation calculation weight vector; The Q-learning algorithm in reinforcement learning is used to update the feedback generation logic. Each state-action is evaluated by maximizing the cumulative reward function, which is expressed as: ； in, For instant rewards, is the discount factor, For the next state; The updated template matching strategy, deviation calculation weight vector, and feedback generation logic are encapsulated into a structured data format and passed to each module of the system for real-time updating of the system configuration.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the AI virtual coach training system based on standard action matching and deviation feedback are implemented as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the AI virtual coach training system based on standard action matching and deviation feedback are implemented as described in any one of claims 1 to 8.

Images