Motion capture-based ai educational interaction system

Abstract

The application relates to the technical field of AI education interaction, and discloses an AI education interaction system based on motion capture. The system comprises four modules, namely a motion data acquisition module, a motion quality analysis module, a teaching parameter configuration module and an interaction result generation module. The motion data acquisition module acquires a historical motion monitoring data set of a target user in a preset historical time period through a motion sensor; the motion quality analysis module traverses the data set, analyzes and determines a motion stability factor and a motion accuracy factor; the teaching parameter configuration module indexes the two factors, optimizes search in a teaching configuration space, and determines a target teaching parameter; and the interaction result generation module carries out interaction with the user according to a preset teaching item based on the target teaching parameter, acquires a real-time interaction data set, and obtains a teaching target interaction result after AI recognizer identification. The system can accurately monitor motion, scientifically analyze quality and provide personalized guidance, and improve the teaching experience.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AI education interaction, in particular to an AI education interaction system based on motion capture. BACKGROUND

[0002] In the current education field, with the continuous penetration of intelligent technology, the demand for personalized guidance in practical teaching projects (such as sports action training, artistic performance guidance, professional skill operation, etc.) is increasingly prominent. In the traditional practical teaching mode, teachers often need to face multiple students at the same time, and it is difficult to monitor and analyze the action details of each student in real time and accurately. In most cases, teachers can only observe the action performance of students with their eyes and judge whether the action is standard based on experience. This way is not only easily affected by subjective factors, but also cannot capture subtle deviations in the action process, such as slight fluctuations in action amplitude and subtle changes in action rhythm, resulting in that the evaluation of student action quality is not comprehensive and accurate enough.

[0003] In the traditional teaching mode, teachers usually develop a unified teaching plan according to their own teaching experience and the overall learning situation of students, and it is difficult to conduct targeted teaching according to the individual action characteristics and learning ability of each student. For students with poor action stability, more basic action practice and guidance may be needed; for students with insufficient action accuracy, more detailed action decomposition explanation and correction may be needed. The unified teaching plan cannot meet these individual needs, which greatly reduces the teaching effect, and some students may gradually lose interest in learning because they cannot get suitable teaching guidance, affecting the realization of the overall teaching goal.

[0004] In the process of practical teaching, the real-time learning state and action performance of students lack effective data recording and analysis means. Teachers cannot accurately grasp the action progress of students at different learning stages, and it is also difficult to find the root cause of students' action problems, resulting in that teaching adjustment lacks scientific basis, and teaching strategies cannot be optimized in a timely manner, further affecting teaching quality and efficiency. With the development of education intelligence, a technical solution is needed that can accurately monitor student actions, scientifically analyze action quality, and provide personalized teaching guidance, to solve the above problems existing in traditional practical teaching. SUMMARY

[0005] The purpose of the present application is to provide an AI education interaction system based on motion capture to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides an AI education interaction system based on motion capture, which comprises:

[0007] The action data collection module is configured to perform action sensor data collection and obtain a set of historical action monitoring data of the target user collected by the action sensor within a preset historical time period.

[0008] The action quality analysis module is configured to perform action quality analysis on the set of historical action monitoring data and determine an action stability factor and an action accuracy factor.

[0009] The teaching parameter configuration module is configured to use the action stability factor and the action accuracy factor as indexes to perform an optimized search in a teaching configuration space and determine a target teaching parameter.

[0010] The interaction result generation module is configured to perform teaching interaction with the target user according to a set of preset teaching projects based on the target teaching parameter, obtain a set of real-time interaction data, and use an AI recognizer to identify the set of real-time interaction data and obtain a teaching target interaction result.

[0011] Preferably, the action data collection module is further configured to:

[0012] The set of historical action monitoring data includes a set of historical skeletal point data, a set of historical joint angle data, and a set of historical motion trajectory data, which are collected by a multi-modal sensor array.

[0013] Preferably, the action quality analysis module is further configured to:

[0014] The set of historical skeletal point data is subjected to cluster analysis to obtain an action typical pattern.

[0015] The proportion of data in the set of historical action monitoring data that matches the action typical pattern is calculated to obtain the action stability factor.

[0016] The average angle deviation of the set of historical joint angle data from a standard action model is calculated to obtain the action accuracy factor.

[0017] Preferably, the teaching parameter configuration module is further configured to:

[0018] A plurality of sample action stability factors and a plurality of sample action accuracy factors, as well as a plurality of corresponding sample teaching parameters, are obtained as construction data.

[0019] A two-dimensional space is pre-constructed, where the x-axis of the two-dimensional space is the action stability factor and the y-axis is the action accuracy factor.

[0020] The construction data is input into the two-dimensional space to obtain a plurality of sample space points, and the plurality of sample teaching parameters are used to label the plurality of sample space points to obtain the teaching configuration space.

[0021] Preferably, the teaching parameter configuration module is further configured to:

[0022] a straight line in the teaching configuration space passing through the action stability factor and parallel to the y-axis as a first reference line;

[0023] a straight line in the teaching configuration space passing through the action accuracy factor and parallel to the x-axis as a second reference line;

[0024] the intersection of the first reference line and the second reference line as a reference point;

[0025] construct a search area with the reference point as the center and a preset search radius, wherein the search area includes a plurality of configuration sample space points;

[0026] perform density search on the plurality of configuration sample space points, determine a target configuration sample space point, and take the sample teaching parameter corresponding to the target configuration sample space point as the target teaching parameter.

[0027] Preferably, the teaching parameter configuration module is further configured to:

[0028] calculate the density distribution of the configuration sample space points in the search area;

[0029] identify the center point of the highest density area as an initial target point;

[0030] start from the initial target point, iteratively adjust the search radius, recalculate the density, and take the final center point as the target configuration sample space point until the density change is less than a threshold value.

[0031] Preferably, the action quality analysis module is further configured to:

[0032] obtain a standard action trajectory sequence;

[0033] align the historical motion trajectory data set with the standard action trajectory sequence;

[0034] calculate a set of aligned trajectory deviation angles;

[0035] statistically screen the set of trajectory deviation angles to obtain an average deviation angle as part of the action accuracy factor.

[0036] Preferably, the action quality analysis module is further configured to:

[0037] extract an action feature vector from the historical action monitoring data set;

[0038] use a machine learning model to learn the quality score of the action feature vector to obtain an action quality score;

[0039] determine a motion stability factor and a motion accuracy factor based on the motion quality score.

[0040] Preferably, the interaction result generation module is further configured to:

[0041] collect current motion data in real time during the teaching interaction process;

[0042] calculate a difference value between the current motion data and the expected motion data;

[0043] dynamically adjust the target teaching parameters according to the difference value to generate adaptive teaching instructions.

[0044] Preferably, the system further comprises:

[0045] a feedback update module configured to update the sample teaching parameters in the teaching configuration space according to the teaching target interaction result to optimize subsequent search.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] The motion sensor data collection performed by the motion data collection module can obtain a set of historical motion monitoring data of the target user within a preset historical time period. Compared with the traditional teaching method of relying only on the naked eye observation of the teacher, the method can comprehensively capture various data information in the user motion process, including the amplitude, rhythm, and intensity changes of the motion, and other detailed data, so that the monitoring of the user motion is no longer limited to surface observation, but is deep into the subtle level of the motion, thereby providing a rich and comprehensive data basis for subsequent motion quality analysis, and helping to more objectively understand the motion performance of the user.

[0048] In the motion quality analysis link, the system determines the motion stability factor and the motion accuracy factor by traversing the set of historical motion monitoring data, thereby realizing the quantitative analysis of the motion quality of the user. In the traditional teaching, the evaluation of the motion quality by the teacher is mostly dependent on subjective feelings, and lacks unified and precise evaluation standards. However, the system converts the abstract motion quality into specific analyzable indexes through the two factors, can clearly reflect the specific situation of the user in terms of motion stability and accuracy, such as the motion fluctuation degree of the user when continuously completing a certain motion, the deviation degree of the motion from the standard motion, and the like, so that the evaluation of the motion quality is more scientific and accurate, and the evaluation deviation caused by subjective factors is avoided.

[0049] The teaching parameter configuration module indexes the action stability factor and the action accuracy factor, and performs an optimization search in the teaching configuration space to determine the target teaching parameter, breaking the unified teaching mode in traditional teaching. The traditional teaching scheme often ignores the individual differences of students and cannot develop targeted teaching strategies according to the action characteristics of different students, but the system can accurately match suitable teaching parameters according to the specific action quality indicators of the user, for example, for users with a low action stability factor, more teaching parameters related to basic action exercises can be configured; for users with insufficient action accuracy factor, teaching parameters related to action decomposition explanation and detail correction can be configured, thereby providing personalized teaching guidance for the user and making the teaching scheme more suitable for the actual learning needs of the user.

[0050] The interaction result generation module carries out teaching interaction based on the target teaching parameter, and uses the AI identifier to identify the real-time interaction data set to obtain the teaching target interaction result, so that the teaching process has real-time and feedback. In the teaching interaction process, the system can analyze the user's action data in real time and generate interaction results in a timely manner, so that the user can quickly understand their performance in the current teaching link, and also provides real-time basis for teachers to adjust teaching strategies. Compared with the feedback lag in traditional teaching, the system can timely discover the problems of the user in the teaching interaction process, such as new action deviations of the user when performing a certain teaching project, and help the user adjust the action in a timely manner through real-time feedback to improve the learning effect.

[0051] The whole system combines motion capture technology with AI technology to realize the data-based and intelligent management of the teaching process. Through the analysis of the user's historical action data and real-time interaction data, a complete learning file of the user can be formed, recording the user's action progress and learning state at different stages, which not only facilitates teachers to master the user's learning process comprehensively, but also provides data support for the continuous optimization of subsequent teaching schemes, promotes the development of practical teaching towards more accurate, efficient and personalized direction, and improves the overall teaching quality and the learning experience of the user. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The timing diagram of the AI education interaction system based on motion capture described in the present application;

[0053] Figure 2 The flowchart for constructing the teaching configuration space;

[0054] Figure 3 The flowchart for searching the teaching parameters based on the reference line. DETAILED DESCRIPTION

[0055] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0056] Please refer to Figure 1 The present application provides an AI education interaction system based on motion capture, which comprises a motion data acquisition module, a motion quality analysis module, a teaching parameter configuration module and an interaction result generation module.

[0057] The motion data acquisition module continuously captures the motion information of the target user in a preset historical time period through motion sensors deployed at key parts of the user's body, forming a historical motion monitoring data set, which contains multi-dimensional motion parameters in time series. The motion quality analysis module processes the historical motion monitoring data set, quantifies the motion stability factor by calculating the time sequence consistency and spatial position consistency of the motion, and derives the motion accuracy factor by comparing the spatial feature difference of the standard motion template. The teaching parameter configuration module takes the motion stability factor and the motion accuracy factor as input indexes, and performs optimization search in a pre-constructed teaching configuration space, which maps the corresponding relationship between different motion ability indicators and teaching parameters, to determine the target teaching parameters, including the difficulty level, feedback frequency and prompt mode of the teaching content. The interaction result generation module controls the teaching interaction process according to the target teaching parameters, generates real-time interaction data set by real-time acquisition of the motion data of the user when executing the preset teaching project set, and calls the AI recognizer to perform pattern recognition and intent understanding on the data, finally outputs the teaching target interaction result, such as the motion completion degree score or the learning stage evaluation.

[0058] Embodiment 1: The action data acquisition module uses a multi-modal sensor array for comprehensive data capture, which is composed of multiple types of sensors to form a complete biomechanical information capture network. Inertial measurement units are precisely arranged on the user's limbs, including the upper arm, lower arm, thigh, lower leg, and the thoracic and lumbar spine of the torso. Each unit contains a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer to continuously record the coordinate change data of the skeletal points in three-dimensional space at a sampling frequency of 100 Hz, forming a high-precision historical skeletal point data set. The optical marker point system is captured by eight high-speed infrared cameras arranged around the training site. The cameras collect the spatial position of the reflective marker points at a frequency of 120 frames per second. The rotation Euler angle data of the main joint points are calculated through the principle of triangulation. After time stamp alignment, these data form a historical joint angle data set. Surface electromyography sensors are attached to the skin surface of the user's main muscle groups, including the deltoid, biceps, quadriceps, and gastrocnemius muscles. The sensors record muscle electrical signal activity at a sampling rate of 2000 Hz. Through signal processing algorithms, the electrical signal is converted into muscle activation timing data. Combined with the speed information of the inertial measurement unit, the acceleration characteristics of the limb movement trajectory are derived, and finally a historical movement trajectory data set is formed. The action quality analysis module processes the collected multi-modal data comprehensively. First, the historical skeletal point data set is subjected to cluster analysis. The density-based clustering algorithm is used to identify repeated motion patterns in the data. This algorithm can effectively handle noise points in motion data and find clusters of any shape. The algorithm first calculates the point density in the surrounding area of each data point and groups points with similar densities into the same cluster. The center trajectory of each cluster is defined as the typical action pattern representing the user's common motion characteristics. When calculating the matching proportion of the historical action monitoring data set and the typical action pattern, the dynamic time warping algorithm is used to calculate the similarity between each motion trajectory and the typical pattern. This algorithm can solve the time axis distortion problem caused by different execution speeds. By constructing a cost matrix, the optimal alignment path between the two time series is found, and the minimum cumulative distance is calculated as the similarity measure. When the similarity value exceeds the preset threshold, it is determined that the data matches the typical pattern. Finally, the action stability factor is obtained by dividing the number of matching data by the total data volume. The action quality analysis module also calculates the angle deviation of the historical joint angle data set and the standard action model. The standard action model is provided by a professional action library and contains the time series variation curve of each joint angle in the ideal state. When calculating the average angle deviation, the measured joint angle data and the standard model are first time-normalized to ensure the same time length for both sequences. Then, the absolute difference between the measured joint angle and the standard angle is calculated at each corresponding time point. The arithmetic mean of all time point differences is taken as the action accuracy factor.The process fully considers the space-time characteristics of movement execution, the action stability factor reflects the consistency and repeatability of the user's action pattern, and the action accuracy factor quantifies the closeness of the action execution to the ideal template, and the two factors together constitute a comprehensive evaluation index of the user's action ability.

[0059] Data synchronization and fusion of the multi-modal sensor array is an important technical basis for this embodiment. All sensor data are clock-synchronized through a unified time server, with an error controlled within milliseconds. The data of the inertial measurement unit are noise-eliminated and attitude-solved through Kalman filtering algorithm. The optical marker point data are converted into rotation data in the joint coordinate system through rigid body transformation principle. The surface electromyography data are processed through band-pass filtering, rectification and smoothing to extract muscle activation features. The historical skeletal point data set contains not only spatial coordinate information, but also velocity and acceleration vectors of each point, which are used to analyze the kinetic characteristics of the movement more finely. The extraction process of the action typical pattern considers the feature differences in different movement stages. For periodic movements, the pattern features in a complete cycle are extracted, and for non-periodic movements, key action segments are extracted as typical patterns. The implementation of the dynamic time warping algorithm uses symmetric warping path constraints to limit the slope of the path to avoid excessive distortion of the time relationship. The Euclidean distance is used as the local cost metric when calculating the similarity, and the dynamic programming method is used to search the warping path to ensure that the global optimal solution is found. The angle deviation calculation is performed separately for each main joint, including the flexion, abduction-adduction and rotation angles of the shoulder joint, elbow joint, hip joint and knee joint. The final action accuracy factor is the weighted average of all joint deviation values, and the weight is determined according to the importance of the joint in the specific teaching project. The whole implementation process embodies the comprehensive application of multi-modal data fusion, pattern recognition and motion analysis technology.

[0060] Example 2: see Figure 2, the teaching parameter configuration module needs to obtain a large amount of sample data as a construction basis during implementation, and these sample data are derived from user interaction records accumulated in a historical teaching database. Each sample contains a user's action stability factor value, an action accuracy factor value, and a teaching parameter configuration adopted at the time. The sample action stability factor and the sample action accuracy factor are obtained through the same quality analysis process, ensuring the consistency of the data source. The corresponding sample teaching parameters are effective parameter settings recorded in the actual teaching process, including teaching content difficulty level, feedback time interval, prompt information detail level, and other configuration information in multiple dimensions. When pre-constructing the two-dimensional space, the Cartesian coordinate system is selected as the mathematical basis. The action stability factor is taken as the horizontal axis x-axis, and the larger the value, the more stable and reliable the user's action execution. The action accuracy factor is taken as the vertical axis y-axis, and the higher the value, the better the user's action accuracy. The value range of both coordinate axes is normalized to map to the closed interval of zero to one. This processing enables data of different dimensions to be compared and analyzed in the same space.

[0061] When the constructed data is input into the two-dimensional space, each sample data is mapped to a point on the two-dimensional plane according to its action stability factor value and action accuracy factor value. These points are distributed at different positions in the coordinate system according to the actual values, forming a collection of sample space points. The teaching parameter configuration module uses a database management system to store and manage these space point data. Each point is associated with corresponding sample teaching parameter metadata. These parameters are stored in the point's attribute information in the form of key-value pairs, including teaching speed parameter value, repetition number parameter value, and feedback intensity parameter value, and other specific configurations. Through this mapping method, each point in the two-dimensional space represents a corresponding relationship between a certain action ability state and a teaching strategy. The collection of all points and their associated parameters constitutes the teaching configuration space. This space is essentially an experience knowledge base that stores the matching relationship between various action ability levels and optimal teaching parameter configurations in history. The construction process of the teaching configuration space considers the completeness and representativeness of the data. The sample data is derived from the performance records of multiple users at different teaching stages, covering user groups at various ability levels from beginners to experts, ensuring that the distribution of points in the space has sufficient diversity and coverage. The distribution of sample points in the space reflects the inherent relationship between action ability and teaching parameters. Generally, areas with higher action stability factor and action accuracy factor values correspond to higher-level teaching parameter configurations, while areas with lower values of both factors correspond to more basic teaching parameter settings. The maintenance of the teaching configuration space includes periodically adding new teaching record data points, deleting outdated or invalid data points, and updating the parameter configurations of existing data points, ensuring that the experience knowledge in the space is continuously optimized and improved as the use process deepens.

[0062] The data normalization processing adopts the minimum-maximum scaling method to linearly transform the original action stability factor and action accuracy factor values to the zero interval. The original minimum and maximum values are preserved during the transformation process to allow new data to be transformed according to the same rules. The teaching parameter configuration module establishes a detailed data archive for each sample data point, including user identification, collection time, teaching project type, and other background information. These information helps to understand the context environment of the data points. As new sample data is continuously added, the teaching configuration space gradually becomes richer and expands, forming a more complete action ability-teaching parameter mapping relationship. This space becomes an important basis for subsequent optimization search and provides data support for personalized teaching parameter configuration. The distribution of sample points in the two-dimensional space may exhibit certain clustering characteristics. Some regions have a higher density of points, indicating that these regions correspond to more common action ability states, and the corresponding teaching parameter configuration has been verified multiple times and is more reliable. Some regions have a lower density of points, which may correspond to special or rarely occurring action ability states. The construction quality of the teaching configuration space directly affects the effectiveness of subsequent parameter search, so it is necessary to ensure the quality and quantity of sample data. Typically, a certain number of samples are required to form a statistically significant spatial distribution. During the implementation of the module, a data quality check mechanism is also established to detect outliers and verify consistency for the input sample data, excluding obviously unreasonable data points, and maintaining the data reliability and practicality of the teaching configuration space.

[0063] In a specific application scenario of a yoga teaching system, the teaching parameter configuration module needs to determine the individualized teaching parameters for a student who is practicing the "tree pose". The system has collected motion performance data of 536 other students in the historical database when they were practicing this pose, each sample containing the values of the student's motion stability factor, motion accuracy factor, and the teaching parameter configuration used during the training process. These sample data come from the training records of the past two years, including training data of people at different levels such as amateur students, professional students, and yoga instructors. The module first preprocesses the sample data, normalizing all motion stability factor values and motion accuracy factor values. The motion stability factor is calculated by analyzing the body sway frequency and amplitude of the student when maintaining the tree pose, and the higher the value, the more stable the pose. The motion accuracy factor is calculated by comparing the body angle deviation of the student's actual pose with the standard tree pose, and the higher the value, the more accurate the motion. The normalization process linearly converts the original data to the closed interval of zero to one, so that data collected at different times can be compared on the same scale. The module constructs a two-dimensional coordinate system, with the horizontal axis representing the normalized motion stability factor and the vertical axis representing the normalized motion accuracy factor. Each sample data is mapped to a point in the coordinate system according to the values of its two factors, such as a sample data of a certain amateur student with a stability factor of 0.35 and an accuracy factor of 0.42, which corresponds to a point at the position (0.35, 0.42) in the coordinate system. Each sample point is associated with the teaching parameter configuration used at the time, including the frequency of voice prompts (number of prompts per minute), the speed of visual demonstration (playback rate of standard motion demonstration), and the detail level of correction feedback (simple prompt or detailed explanation). In the coordinate system, the sample points in different regions show obvious distribution rules: the upper left corner area is clustered with students whose stability is poor but accuracy is high, and these points correspond to teaching parameters that require more stability training prompts; the lower right corner area is clustered with samples whose stability is good but accuracy is poor, and the corresponding parameters emphasize the correction of motion details; the middle area is distributed with students whose performance is balanced. All sample points and their associated teaching parameters together form the teaching configuration space of the tree pose, which is actually an experience database that records the most suitable teaching parameter configuration for students of various ability levels in history. When a new student uses the system, the module first collects his training data and calculates the motion stability factor as 0.38 and the motion accuracy factor as 0.45. The system automatically maps the student's data into the teaching configuration space and finds the most suitable teaching parameter configuration within the space, providing data support for subsequent individualized teaching. As the system continuously collects new teaching data, the teaching configuration space will be continuously updated and expanded, making the parameter recommendation more accurate and individualized.

[0064] Example 3: see Figure 3, the teaching parameter configuration module enters the search area construction stage after determining the reference point. The module first reads the action stability factor value and the action accuracy factor value of the current user, and positions these two values as coordinate values in the pre-constructed teaching configuration space. The module draws a straight line passing through the action stability factor value point and parallel to the longitudinal axis as the first reference line, which represents the horizontal position of the current user's action stability; at the same time, a straight line passing through the action accuracy factor value point and parallel to the horizontal axis is drawn as the second reference line, which reflects the vertical position of the user's action accuracy. The intersection of the two reference lines in the two-dimensional space is determined as the reference point, which is determined by the action stability factor and the action accuracy factor, and represents the current comprehensive action ability level of the user. Taking the reference point as the center, the module constructs a circular search area according to the preset initial search radius, and the initial radius value is dynamically determined according to the distribution density of the sample points in the teaching configuration space. When the sample points are sparsely distributed, a larger radius is used, and when the sample points are densely distributed, a smaller radius is used. The search area contains multiple configuration sample space points, each of which is associated with the teaching parameter configuration used by users with similar action ability in history. These parameters include the specific settings of the teaching content promotion speed, the exercise repetition number and the feedback prompt detail level, etc.

[0065] When analyzing the density of the configuration sample space points in the search area, the module uses the kernel density estimation algorithm to calculate the distribution of the points. This algorithm estimates the probability density distribution of the points by placing a kernel function at each sample point position and then superimposing all kernel functions in the region. The density calculation uses the following relationship:

[0066]

[0067] Where: represents the estimated density at position , is the total number of sample points in the search area, is the bandwidth parameter that controls the smoothness of the kernel function, is the selected kernel function, represents the Euclidean distance from position to the th sample point . Through this calculation, the most densely distributed area of the sample points in the search area can be identified.

[0068] After identifying the center point of the highest density area as the initial target point, the module begins to iteratively adjust the search radius, reducing the search radius to a certain proportion of the previous iteration each time, and then recalculating the point density distribution in the new search area. During the iteration process, the density change is continuously monitored, and when the density change of adjacent two iterations is less than the preset threshold, the iteration is stopped, and the center point at this time is determined as the target configuration sample space point. The sample teaching parameters corresponding to this target point are extracted as the final target teaching parameters, including the optimal teaching speed, suitable repetition number and personalized feedback intensity for the current user's action ability level. The whole optimization search process fully considers the distribution characteristics of the historical data in the teaching configuration space, and through the density search method, the teaching parameter configuration most suitable for the current user's action ability level can be found, avoiding the local optimal problem that may be caused by simple nearest neighbor search. The mechanism of iteratively adjusting the search radius ensures that the search accuracy gradually improves with the increase of the number of iterations, and the target configuration sample space point finally found represents the most commonly used and verified effective teaching parameter settings in the history of similar ability users.

[0069] After the teaching parameter configuration module determines the target teaching parameters, it passes these parameters to the interaction result generation module to guide the subsequent personalized teaching interaction process. At the same time, it records the relevant data of this search, including the reference point coordinates, search area size, iteration number and final target point information, which are used for subsequent optimization of the search algorithm and updating of the teaching configuration space. The module also establishes an exception handling mechanism, which uses a backup search strategy to ensure that suitable teaching parameter configurations can always be found when the number of sample points in the search area is too small or the density distribution is too uniform.

[0070] Taking the golf swing teaching system as an example, the teaching parameter configuration module needs to determine the most suitable swing training parameters for an intermediate-level learner. After the preliminary assessment, the learner's action stability factor is 0.62, and the action accuracy factor is 0.58. These two values reflect the learner's current level in maintaining the consistency of the swing action and the accuracy of the shot. The module first locates the reference point in the teaching configuration space. The action stability factor value 0.62 is used to determine the position on the horizontal axis, and a straight line perpendicular to the horizontal axis and passing through this point is drawn as the first reference line. At the same time, the action accuracy factor value 0.58 is used to determine the position on the vertical axis, and a straight line parallel to the horizontal axis and passing through this point is drawn as the second reference line. The two reference lines intersect at the coordinate (0.62, 0.58) of the teaching configuration space, and this intersection point is the reference point of this search, representing the learner's current action ability state. According to the distribution characteristics of the sample points in the teaching configuration space, the module determines the initial search radius as 0.18, and constructs a circular search area with the reference point as the center. This area contains 47 historical sample points, each of which records the teaching parameter configuration used by previous learners with similar action ability levels. These parameters include swing demonstration speed, correction feedback frequency, and training repetition number, etc. The module starts to analyze the density of the sample points in the search area, and uses the kernel density estimation algorithm to calculate the point distribution density in each small area. By scanning the entire search area, it is found that the point distribution is most dense near the coordinate (0.64, 0.61), and this area is gathered with 12 historical sample points. This position is determined as the initial target point. Starting from the initial target point, the module starts the iterative optimization process. First, the search radius is reduced to 0.09, and the point density distribution in the new area is recalculated. It is found that the density center slightly shifts to (0.635, 0.615). Continue the second iteration, the search radius is further reduced to 0.045, and the density center moves to (0.632, 0.617). The third iteration adjusts the radius to 0.0225, and the density center stabilizes at (0.631, 0.618), and the density change from the last iteration is less than the pre-set threshold. The iteration process terminates, and the final target configuration sample point is determined as (0.631, 0.618). The teaching parameters corresponding to this sample point are extracted as the target teaching parameters: the swing demonstration speed is set to 0.8 times the standard speed, allowing the learner to observe the action details more clearly; the correction feedback frequency is set to one voice prompt every three swings, avoiding information overload; the training repetition number is set to 12 times per group, a total of 3 groups, which is suitable for the learner's current ability level. These parameter configurations have been used by multiple learners with similar action abilities and have achieved good training results. The module passes the determined target teaching parameters to the real-time teaching engine, and records the key data of the search process, including the reference point coordinates, the number of iterations, and the final determined sample point information.These data will be used to optimize the search algorithm and update the teaching configuration space in the future. As the system continues to accumulate new teaching data, the search accuracy and parameter matching effect will continue to improve. The entire search process fully considers the distribution characteristics of historical data and finds the most suitable teaching parameter configuration for the current ability level of the student through density clustering and iterative optimization.

[0071] In the implementation process of embodiment 4, the standard action trajectory sequence needs to be obtained as a comparison reference. This sequence is derived from the ideal action record of professional athletes performing a specific teaching project, and contains complete spatial path and timing information from the start to the end of the action. The standard action trajectory is stored in the form of time series data, and each time point corresponds to the three-dimensional spatial coordinates of the key joints of the human body. These coordinate data are collected in a controlled environment through an optical motion capture system, and have high precision and reliability. The module aligns the user's historical motion trajectory data set with the standard action trajectory sequence. The algorithm based on dynamic time warping is used to solve the time length difference problem that may exist between the two sequences. The algorithm finds the optimal path to achieve the maximum similarity matching of the two sequences on the time axis, eliminating the comparison errors caused by different execution speeds.

[0072] When calculating the set of trajectory deviation angles after alignment, the module calculates the spatial angle between the actual motion direction vector and the standard direction vector at each corresponding time point. The direction vector is calculated by the difference of the position coordinates of consecutive time points. For each main joint, such as the wrist, elbow, shoulder, etc., the deviation angle between its motion direction and the standard direction is calculated respectively. The deviation angle values of all joints constitute the complete set of trajectory deviation angles. The module performs statistical screening processing on the set of trajectory deviation angles. First, the mean and standard deviation of all deviation angles are calculated. Then, abnormal values that exceed three times the standard deviation range are excluded. These abnormal values may come from sensor transient errors or accidental mistakes in action execution. Finally, the arithmetic mean of the remaining valid deviation angles is taken as a component of the action accuracy factor.

[0073] The module also extracts a multi-dimensional action feature vector from the historical action monitoring data set. These features include motion amplitude features, acceleration peak value features, joint coordination coefficient features, and other spatio-temporal motion parameters. The motion amplitude feature is obtained by calculating the maximum displacement distance of the joint during the motion process. The acceleration peak value feature extracts the maximum acceleration value in each direction from the acceleration sensor data. The joint coordination coefficient is obtained by calculating the correlation coefficient between different joint motion angles. A machine learning model is used to learn the quality score of the action feature vector. A support vector regression model is used for supervised learning on the pre-labeled training data set. The training data contains a large number of action sample feature vectors that have been scored by experts. The model learns the non-linear mapping relationship between the feature space and the quality score.

[0074] The action quality score output by the machine learning model is used to determine the final action stability factor and action accuracy factor. The variance of the quality scores of multiple action samples over a continuous period of time is used as the action stability factor, reflecting the consistency of the user's action execution. The difference between the quality score and the ideal full score value is used as another component of the action accuracy factor, together with the average value of the trajectory deviation angle, to form a complete evaluation of action accuracy. This process realizes the fusion of multi-source information, considering both the spatial deviation of the movement trajectory from the standard template and the overall evaluation of action quality by the machine learning model, making the final generated action ability factor more comprehensive and reliable. See Table 1.

[0075] Table 1: Trajectory Deviation Angle Calculation Data

[0076]

[0077] A complete data quality control mechanism is established during the implementation of the module. The input historical action monitoring data is subjected to integrity check and noise filtering to remove obviously unreasonable data points. Standard action trajectory sequences are established according to different teaching project types, with each project having a corresponding ideal action template. These templates are regularly reviewed and updated by a team of professional trainers. The training of the machine learning model uses the cross-validation method to optimize the model parameters, and the model is regularly updated with newly collected data to maintain the timeliness of the evaluation capability. The output range of the action quality score is set between zero and one hundred, with a higher score indicating better action quality, and a smaller difference from the full score indicating higher action accuracy.

[0078] In the teaching interaction process, the interaction result generation module starts the real-time data acquisition process, continuously acquiring current action data at a frequency of 50 Hz through motion sensors distributed at key parts of the user's body. These data include kinematic parameters such as joint angle changes, limb displacement speeds, and accelerations. The module calculates the multi-dimensional difference values between the current action data and the expected action data, which are generated according to the teaching project goals and include ideal action execution trajectories, angle change ranges, and timing requirements. The difference value calculation covers quantitative indicators such as spatial position deviation, motion direction deviation, and time synchronization error. According to the real-time calculated difference values, the target teaching parameters are dynamically adjusted. When the difference value exceeds the preset threshold, the parameter adjustment mechanism is triggered. The teaching speed parameter is dynamically adjusted according to the action execution accuracy. When the action execution is not accurate enough, the teaching pace is appropriately reduced, and when the action execution is good, the teaching rhythm is gradually increased. The detail level of the feedback prompts is adjusted according to the user's real-time performance. When significant action deviation is detected, the detail level of the visual prompt information is increased to provide more specific guidance and advice. When the user's action execution is accurate, the prompt frequency is reduced to avoid information overload.

[0079] The generated adaptive teaching instructions are fed back to the user in real time through multiple modalities. Visual feedback is provided through a head-mounted display or a screen display showing standard motion demonstrations and contrast pictures of the current motion. Auditory feedback is provided through voice prompts indicating specific aspects that need to be improved. Tactile feedback is provided through a vibration device to remind the user of deviations in motion execution. The feedback updating module updates the sample teaching parameters in the teaching configuration space based on the teaching target interaction results, which include the motion accuracy trend, the final mastery level, and the learning efficiency indicators recorded during the teaching process. These data, after being standardized, are added to the teaching configuration space as new experience samples. The updating process of the teaching configuration space follows strict data screening standards. Only those teaching parameter configurations that meet certain standards are included in the space as new samples. The motion stability factor and the motion accuracy factor of the new sample points are calculated based on the actual performance data during the teaching process. After the new sample points are added to the teaching configuration space, they participate in the subsequent optimization search process, allowing the teaching parameter configuration to be continuously optimized and improved over time, gradually forming a more accurate and personalized teaching parameter recommendation capability. A data quality monitoring mechanism is established during the real-time data collection process to check the integrity and reasonableness of the sensor data in real time. When data anomalies are detected, the re-collection process is automatically triggered. The difference value calculation uses a weighted comprehensive scoring method, assigning different weight coefficients to spatial deviation, directional deviation, and time deviation based on the characteristics of different teaching projects, ensuring that the difference value can fully reflect the quality of motion execution. The parameter adjustment strategy uses a gradual adjustment method to avoid frequent and large changes in teaching parameters that may affect user experience. The adjustment amplitude is determined based on the size and duration of the difference value.

[0080] The generation of adaptive teaching instructions takes into account the user's current learning state and attention level, assisted by eye tracking and physiological signal monitoring to determine the user's focus level, and adjusts the presentation method and detail of feedback information in a timely manner. The feedback updating module regularly organizes and optimizes the sample data in the teaching configuration space, removing outdated or ineffective sample points to maintain the freshness and effectiveness of the space data. A sample point life cycle management mechanism is established, and older sample points are gradually eliminated as new data is continuously added. The entire implementation process forms a complete closed-loop optimization system, with real-time interaction data guiding teaching parameter adjustment, and teaching effectiveness data optimizing the teaching configuration space. The continuous enrichment of space data improves the accuracy of subsequent parameter search. This self-optimizing mechanism allows the system to adapt to the individual characteristics of different users, providing increasingly accurate teaching parameter configurations over time and forming a continuously improving personalized teaching capability. During the implementation of the modules, attention is paid to the smoothness of the user experience, and all adjustments and changes are implemented gradually to avoid sudden changes that may affect the coherence and comfort of the learning process.

[0081] In the implementation process, the action capture-based AI education interaction system can further combine a multi-agent collaborative architecture to enhance the system's processing capacity for complex teaching tasks and the level of personalized interaction. In the action data acquisition link, the system continuously captures historical action monitoring data, including bone point coordinate sequences, joint angle change curves, and motion trajectory features, through a multi-modal sensor array deployed on key parts of the user's body. After preprocessing, these data are quantitatively evaluated by the action quality analysis module, the user's typical action patterns are identified by clustering algorithms, and the action stability factor and the action accuracy factor are calculated as comprehensive indicators of the user's action ability.

[0082] In the teaching parameter configuration link, the system introduces an agent scheduling mechanism to build a special teaching strategy planning agent and a parameter optimization agent. The teaching strategy planning agent searches collaboratively in the teaching configuration space according to the user's action stability factor and accuracy factor. This space constructs a two-dimensional coordinate system with the action stability factor as the horizontal axis and the action accuracy factor as the vertical axis, where each sample point is associated with the teaching parameters used by similar action ability users in history. The parameter optimization agent identifies the optimal configuration sample point in the search area through a density clustering algorithm and gradually narrows the search range based on an iterative optimization mechanism to ultimately determine the target teaching parameters, such as teaching content advancement rate, feedback frequency, and prompt detail level, etc.

[0083] The interaction result generation module further integrates a multi-agent collaborative processing mechanism to deploy real-time monitoring agents and dynamic adjustment agents during the teaching interaction process. The real-time monitoring agent continuously acquires current action data of the user at a high sampling rate, including real-time information such as joint motion angle and limb displacement speed, and calculates the difference value between the current action data and the expected action data. The dynamic adjustment agent dynamically adjusts the teaching parameters based on the difference value, for example, when the user's action accuracy is detected to decrease, the teaching advancement speed is automatically reduced, and the detail level of visual prompts is increased. At the same time, the system cooperates with the teacher agent and the evaluation agent to perform multi-dimensional identification on the real-time interaction data set to generate teaching target interaction results, such as action completion degree scores or learning stage evaluations.

[0084] In terms of system self-optimization, an autonomous evolution agent is introduced to update the parameter configuration in the teaching configuration space through a continuous learning mechanism. Based on user historical interaction data and real-time feedback, the agent uses automated strategies to fine-tune teaching parameters, including adjusting threshold settings for search radius, optimizing parameter configurations for density clustering algorithms, etc. The feedback update module adds effective teaching parameter configurations as new sample points to the teaching configuration space based on the teaching target interaction results, enabling the system to continuously improve the accuracy and adaptability of parameter recommendations as the frequency of use increases.

[0085] The collaborative operation of the multi-agent architecture in the system is reflected in the hierarchical design of task division and information flow. The perception agent is responsible for the collection and preliminary processing of multi-modal sensor data, the planning agent formulates teaching strategies based on the action quality analysis results, the execution agent controls the feedback output of the interactive interface, and the memory agent maintains the historical data and update records of the teaching configuration space. These agents work collaboratively through a message passing mechanism, for example, when the perception agent detects that the user's action deviates significantly, it immediately notifies the planning agent to re-evaluate the teaching parameters, and the execution agent adjusts the real-time feedback content.

[0086] In specific implementation scenarios, the system can be applied to teaching projects with high requirements for action specification, such as sports training or operation skill guidance. After the system obtains the user's historical action data through the action data acquisition module, the action quality analysis module calculates the action stability factor as 0.65 and the action accuracy factor as 0.70. Based on this, the strategy planning agent in the teaching parameter configuration module locates the reference point in the teaching configuration space, and the parameter optimization agent identifies the target configuration sample point through density search, corresponding to the teaching parameters including the demonstration speed being 0.85 times the standard value and the feedback interval being a prompt after every two action executions. The real-time monitoring agent in the interactive result generation module continuously tracks the user's action data during the teaching process, and when it detects that the joint angle deviation exceeds the threshold, it dynamically adjusts the agent to automatically reduce the demonstration speed and triggers the voice prompt mechanism to provide targeted guidance.

[0087] The multi-agent collaborative architecture of the system not only improves the processing efficiency of complex teaching tasks, but also realizes fine regulation and control of the teaching process through multi-agent division of labor and cooperation. The organic cooperation of perception, planning, execution, and memory agents enables the system to adapt to different users' action characteristics and learning pace, providing a truly personalized teaching interactive experience. At the same time, the self-evolution agent continuously optimizes system performance through a continuous learning mechanism, forming a virtuous development cycle combining data-driven and experience accumulation. In terms of technical implementation, the system uses a distributed agent communication framework, and each agent exchanges data through a lightweight message queue to ensure real-time and reliability. The teaching strategy planning agent and the parameter optimization agent use a strategy generation method based on rule-based reasoning and data-driven combination, considering both the statistical laws of historical data and the dynamic characteristics in real-time interaction. The real-time monitoring agent uses stream data processing technology to perform real-time filtering and feature extraction on sensor data, ensuring the timeliness of action evaluation. The dynamic adjustment agent is based on reinforcement learning mechanism, dynamically optimizing parameter adjustment strategies according to user responses, gradually improving the interactive intelligence level of the system.

[0088] With the introduction of the multi-agent collaborative architecture, the system further enhances the adaptability and optimization efficiency in complex teaching scenarios while maintaining the original motion capture and personalized teaching advantages. The division of labor and information sharing among agents enable the system to more accurately understand user needs, dynamically adjust teaching strategies, and provide a continuously optimized interactive experience. This architectural design not only enhances the technological advancement of the system but also provides a new implementation path for the in-depth application of motion capture technology in the field of educational interaction.

[0089] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements not only includes those elements but also includes other elements not explicitly listed or inherent to such a process, method, article or device.

[0090] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, alternatives, and variations can be made thereto without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-powered educational interactive system based on motion capture, characterized in that, The system includes: The motion data acquisition module is used to perform motion sensor data acquisition and obtain a set of historical motion monitoring data of the motion sensor monitoring the target user within a preset historical time period. The motion quality analysis module is used to traverse the historical motion monitoring data set to perform motion quality analysis and determine the motion stability factor and motion accuracy factor. The teaching parameter configuration module is used to perform an optimization search in the teaching configuration space using the action stability factor and action accuracy factor as indexes to determine the target teaching parameters; Multiple sample action stability factors and multiple sample action accuracy factors, as well as the corresponding multiple sample teaching parameters, are obtained as the construction data; A pre-constructed two-dimensional space is provided, wherein the x-axis of the two-dimensional space is the motion stability factor and the y-axis is the motion accuracy factor. The constructed data is input into the two-dimensional space to obtain multiple sample space points, and the multiple sample space points are identified using the multiple sample teaching parameters to obtain the teaching configuration space; The straight line in the teaching configuration space that passes through the motion stability factor and is parallel to the y-axis is used as the first reference line; The straight line in the teaching configuration space that passes through the action accuracy factor and is parallel to the x-axis is used as the second reference line; The intersection of the first reference line and the second reference line is taken as the reference point; Centered on the reference point, a search area is constructed according to a preset search radius, wherein the search area includes multiple configuration sample space points; A density search is performed on the plurality of configured sample space points to determine the target configured sample space point, and the sample teaching parameters corresponding to the target configured sample space point are used as the target teaching parameters. The interaction result generation module is used to conduct teaching interactions with the target user according to a preset set of teaching projects based on the target teaching parameters, obtain a real-time interaction data set, and use an AI recognizer to identify the real-time interaction data set to obtain the teaching target interaction result.

2. The AI-based educational interactive system based on motion capture as described in claim 1, characterized in that, The motion data acquisition module is also used for: The historical motion monitoring data set is collected using a multimodal sensor array, including historical skeletal point data set, historical joint angle data set, and historical motion trajectory data set.

3. The AI-based educational interactive system based on motion capture as described in claim 2, characterized in that, The motion quality analysis module is also used for: Cluster analysis is performed on the historical skeletal point data set to obtain typical action patterns; The proportion of data in the historical motion monitoring data set that matches the typical motion pattern is statistically analyzed to obtain the motion stability factor. The mean angle deviation between the historical joint angle data set and the standard motion model is calculated to obtain the motion accuracy factor.

4. The AI ​​educational interactive system based on motion capture as described in claim 3, characterized in that, The teaching parameter configuration module is also used for: Calculate the density distribution of spatial points of the configured sample within the search area; Identify the center point of the region with the highest density as the initial target point; Starting from the initial target point, the search radius is iteratively adjusted and the density is recalculated until the density change is less than the threshold. The final center point is then used as the target configuration sample space point.

5. The AI ​​educational interactive system based on motion capture as described in claim 2, characterized in that, The motion quality analysis module is also used for: Obtain the standard motion trajectory sequence; Align the historical motion trajectory data set with the standard motion trajectory sequence; Calculate the set of trajectory deviation angles after alignment; The set of trajectory deviation angles is statistically filtered to obtain the average deviation angle as part of the motion accuracy factor.

6. The AI-based educational interactive system based on motion capture as described in claim 1, characterized in that, The motion quality analysis module is also used for: Extract motion feature vectors from the historical motion monitoring data set; A machine learning model is used to learn a quality score for the action feature vector to obtain an action quality score. Based on the motion quality score, a motion stability factor and a motion accuracy factor are determined.

7. The AI ​​educational interaction system based on motion capture as described in claim 1, characterized in that, The interaction result generation module is also used for: During the interactive teaching process, real-time action data is collected. Calculate the difference between the current action data and the expected action data; The target teaching parameters are dynamically adjusted based on the difference value to generate adaptive teaching instructions.

8. The AI ​​educational interaction system based on motion capture as described in claim 7, characterized in that, The system also includes: The feedback update module is used to update the sample teaching parameters in the teaching configuration space based on the interaction results of the teaching objectives, so as to optimize subsequent searches.

Images