A method for teaching and training and quantitatively evaluating flat acting body language

Abstract

The application relates to the technical field of intelligent teaching training, and discloses a flat teaching body segment intelligent teaching training and quantitative evaluation method, which determines a teaching training task according to a teaching training plan, presents standard demonstration teaching content through a learner terminal, carries out time phase alignment and quantitative teaching evaluation on the actions of learners, obtains local action deviation and action comprehensive scores, generates a teaching error correction guidance text according to the local action deviation, presents action feedback information through the learner terminal, writes the action comprehensive scores and the local action deviation into a student portrait database, determines a persistent weak feature according to historical learning records, matches special training content according to the persistent weak feature, adjusts a subsequent teaching training plan, and the learner carries out targeted training in the next training period. The application forms an intelligent teaching training process combining action sensing processing as a teaching evaluation means, standard demonstration, action evaluation, error correction feedback and special training adjustment.

Description

Technical Field

[0001] This invention relates to the field of intelligent teaching and training technology, specifically to an intelligent teaching and training method and quantitative evaluation method for Pingjiang Opera movements. Background Technology

[0002] As a traditional opera form, Pingjiang Opera has specific physical norms and power requirements for its movements. In conventional teaching and transmission, it mainly relies on the demonstrator's on-site guidance and the learner's imitation and memorization. Due to the lack of objective auxiliary assessment tools, postural deviations that learners exhibit during independent practice cannot be corrected in a timely manner, easily leading to the solidification of incorrect movement habits.

[0003] To improve teaching efficiency, several digital opera teaching support systems have emerged in the industry. For example, existing digital video recording systems can record and review teaching footage, but they only have video storage capabilities and cannot provide quantitative data on movement features. Furthermore, Chinese Patent Publication No. CN121860821A discloses a personalized opera movement teaching support method and system based on multi-agent technology. This method extracts and compares movement features, achieving a certain degree of auxiliary analysis of opera movements. However, for the quantitative evaluation and model construction of specific opera movements such as Pingjiang Opera, the aforementioned existing technologies still have the following significant shortcomings: First, the physical movements instruction in Pingjiang Opera has varying degrees of standardized requirements for different body parts. Some body parts must strictly conform to standard demonstration movements, while others are allowed for reasonable individual differences. Existing movement-assisted teaching methods typically compare learners' movements based on a single standard movement or a uniform evaluation scale. This makes it difficult to determine the permissible range of variation for different body parts based on multiple standard demonstration samples that conform to teaching standards, easily leading to performance evaluation results that do not fully reflect the different standardized requirements in Pingjiang Opera physical movements instruction. Furthermore, the limited number of standard movement samples available for establishing standard movement teaching evaluation benchmarks in specific opera genres like Pingjiang Opera can also affect the stability of the movement characteristics acquired for learners' movement teaching evaluation.

[0004] Secondly, in actual motion comparison, learners' practice sequences and standard demonstration sequences often exhibit temporal phase differences in performance speed and rhythm. Existing evaluation methods largely rely on static frame-to-frame spatial feature calculations, failing to comprehensively consider temporal sequence alignment and the tolerance variance of different torso positions, which easily leads to evaluation errors due to rhythmic inconsistencies. Furthermore, existing technologies primarily output evaluation results as simple geometric distances or similarity values, lacking error correction mechanisms that convert underlying quantitative biases into natural language forms. This makes the evaluation results difficult for learners to directly understand and apply to motion correction.

[0005] Finally, existing motion assessment systems typically end with the completion of a single motion assessment, lacking mechanisms for continuous learning analysis and subsequent content adjustment across multiple teaching cycles. The comprehensive motion scores, localized motion deviations, and typical error records generated by learners' long-term practice are not further utilized to identify persistent learning weaknesses. The system struggles to distinguish between isolated, occasional errors and entrenched errors that persist across multiple practice cycles, and it also fails to automatically determine corresponding specialized training content and adjust the teaching plan for the next cycle based on learners' persistent weaknesses. Therefore, existing systems lack a feedback link between motion assessment results and subsequent teaching content, making it difficult to form a closed-loop teaching and training process that dynamically adjusts teaching content based on learners' long-term learning status. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent teaching and quantitative assessment method for Pingjiang Opera movements. It aims to solve the problems of existing autonomous teaching methods for Pingjiang Opera movements, such as the difficulty in objectively evaluating learners' follow-up movements based on standard demonstrations, the difficulty in directly converting quantitative evaluation results into targeted teaching corrections, the difficulty in using historical learning data from multiple teaching cycles to identify learners' persistent weaknesses, and the inability to dynamically adjust subsequent teaching and training content based on these persistent weaknesses. This invention forms a continuous teaching and training process that combines standard demonstrations, follow-up practice, teaching evaluation, real-time error correction, learning analysis, and adjustments to specific training plans.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] This invention provides an intelligent teaching, training, and quantitative assessment method for Pingjiang Opera body movements, establishing an intelligent teaching closed loop for autonomous training of Pingjiang Opera body movements. This enables learners to complete standard demonstration learning, movement practice, error feedback, and targeted reinforcement training without real-time teacher guidance. The method includes the following steps: The teaching and training task for this lesson is determined according to the teaching and training plan corresponding to the learner, and standard demonstration teaching content related to the teaching and training task is presented to the learner through the learner's terminal. The learner is given a real-time practice video stream generated by following the body movements according to the standard demonstration teaching content. The learner's real-time movements in the real-time practice video stream are time-phase aligned and the movement is quantitatively evaluated to obtain local movement deviations and comprehensive movement scores. Based on the local action deviation, a teaching error correction guide text is generated, and the teaching error correction guide text is presented through the learner's terminal; The comprehensive score of the movement and the local movement deviation are written into the student profile database, and persistent weakness characteristics are determined based on the historical learning records formed by multiple practice cycles in the student profile database. Based on the characteristics of persistent weaknesses, specific training content is determined, and subsequent teaching and training plans are adjusted accordingly.

[0009] By using the above methods, a continuous teaching and training process is formed, which includes standard demonstration, learner practice, quantitative teaching evaluation of movements, immediate error correction, and adjustment of subsequent specialized training content. This allows the evaluation results of the current practice to be used in determining the subsequent teaching content.

[0010] Preferably, the quantitative teaching evaluation of movement uses a standard feature variance distribution matrix as the evaluation basis. The standard feature variance distribution matrix is ​​established as follows: a standard demonstration video is acquired; the absolute coordinates of the skeletal key points in the standard demonstration video are extracted; relative coordinates of the skeletal key points are generated through spatial coordinate system translation; skeleton topology normalization is performed on the relative coordinates of the skeletal key points to generate standard feature skeleton vectors and calculate joint angles; sequence alignment is performed on multiple standard demonstration samples of the same type; spatial distribution variances of different feature dimensions are calculated; and these are combined to form the standard feature variance distribution matrix.

[0011] Preferably, the basic pose extraction network is pre-trained using motion data from similar theatrical genres; the spatial distribution variance is compared with the generalization tolerance threshold to generate binary mask parameters; the generalization tolerance threshold is jointly calibrated based on the pre-training loss convergence value of the basic pose extraction network on a general human pose dataset and the standardization test data of Pingjiangxi action; the frozen and unfrozen states of the feature extraction branches are controlled according to the binary mask parameters, and local fine-tuning is performed on the feature extraction branches in the unfrozen state using standard Pingjiangxi samples.

[0012] Preferably, the time phase alignment includes: extracting real-time skeletal key points from the real-time practice video stream and generating a real-time feature skeleton vector; performing downsampling on the standard demonstration action sequence; calculating the local spatial distance between the real-time feature skeleton vector and the standard feature skeleton vector and constructing a local spatial distance matrix; calculating the time warping path using a dynamic time warping algorithm, and performing time axis resampling according to the time warping path to form a synchronous teaching evaluation sequence pair.

[0013] Preferably, the spatial distribution variance in the standard feature variance distribution matrix is ​​subjected to reciprocal mapping and sum-normalization to obtain dynamic evaluation weights; the normalized cosine distance and normalized angle deviation are calculated based on the synchronous teaching evaluation sequence; the normalized cosine distance and normalized angle deviation are weighted and linearly combined using the dynamic evaluation weights to form local action deviations; the local action deviation sequence is subjected to time-series integral averaging to obtain a comprehensive action score.

[0014] Preferably, typical error nodes are determined based on the local action deviations, and the timestamps, joint names, local action deviations, and comprehensive action scores of the typical error nodes are extracted to form structured teaching evaluation data; the structured teaching evaluation data is formed into a data slice sequence and embedded with a structured prompt word template, and the structured prompt word template is input into a large language model to generate teaching error correction guidance text.

[0015] Preferably, the learner's action screen and the standard demonstration action screen are obtained based on the timestamp of the typical error node, and a skeleton overlay layer is generated; a skeleton highlight comparison image is formed based on the local action deviation; the standard demonstration action screen, the learner's action screen, the skeleton highlight comparison image, the comprehensive action score, and the teaching error correction guidance text are presented in association through the learner's terminal.

[0016] Preferably, the comprehensive action scores within a preset teaching cycle are extracted from the student profile database to form a one-dimensional vector of historical scoring time series; an exponential moving average filter is applied to the one-dimensional vector of historical scoring time series to form a smoothed scoring sequence; a skill qualification score benchmark value is set in conjunction with the standard teaching syllabus for Pingjiang Opera; a bottleneck stage of action is determined based on smoothed scoring nodes that are continuously lower than the skill qualification score benchmark value, and a fixed error point is determined based on the typical error nodes of the bottleneck stage of action, forming the persistent weakness feature.

[0017] Preferably, the specific feature dimensions of the persistent weakness feature are used as input conditions to match in the Pingjiangxi special training mapping table to determine the special training content; the teaching task node associated with the special training content is added to the priority execution position of the teaching and training plan, and the updated teaching and training plan is sent to the learner terminal, so that the learner terminal will prioritize loading the standard demonstration teaching content of the special training content in the next practice cycle.

[0018] Preferably, when the persistent weakness feature does not have a completely consistent feature item in the Pingjiangxi Special Training Mapping Table, the specific action feature of the persistent weakness feature is mapped to the next higher level body function area according to the subordinate relationship of human body structure, and the comprehensive special review content of the next higher level body function area is searched within the expanded search range.

[0019] This invention provides a method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera. It has the following beneficial effects: 1. This invention does not simply output a single action score for the learner's practice movements. Instead, it presents standard demonstration teaching content to the learner based on the current teaching and training plan. After the learner completes the follow-up practice, it generates action teaching evaluation and immediate error correction feedback. Furthermore, it utilizes historical learning status generated from multiple practice cycles to identify persistent weaknesses. Based on persistent weaknesses, it automatically matches corresponding specialized practice content and adjusts subsequent teaching and training plans. This allows the learning evaluation results of the previous teaching cycle to directly change the teaching content presented to the learner in the next teaching cycle, improving the pertinence and continuity of the self-training of body movements in Peking Opera.

[0020] 2. This invention determines the specific body parts and time points of the learner's errors by quantifying time phase alignment and action deviation, and converts the quantified deviation into teaching error correction guidance text. At the same time, it combines standard demonstration screens, actual action screens of learners, and skeleton highlight comparison images for terminal presentation, enabling learners to understand the location of the error and the direction of adjustment, reducing the problem that simply outputting similarity or scoring is not enough to directly guide action correction.

[0021] 3. This invention reduces the impact of differences in body size among learners by transforming local coordinates of the human body and normalizing the topology of the skeleton, and eliminates the time misalignment caused by the difference between the learner's movement speed and the standard demonstration movement speed by using dynamic time warping; further, it determines the teaching evaluation weight of different body parts based on the spatial distribution variance formed by multiple standard demonstration samples, so that strictly standardized movements and movements that allow reasonable individual differences are evaluated on different scales.

[0022] 4. This invention obtains a basic posture extraction network through pre-training with similar opera genres, and selectively controls the freezing and unfreezing of network branches according to the standard feature variance distribution matrix. Only the network parameters related to the specific movements of Pingjiang Opera are locally fine-tuned. While retaining the general human posture recognition ability, it reduces the overfitting caused by small sample full parameter updates and improves the feature extraction stability of Pingjiang Opera specific body movements.

[0023] 5. This invention constructs difficult example samples by screening long-tail abnormal practice data, and improves the accuracy of action evaluation during the long-term operation of the intelligent teaching system when fine-tuning conditions are met, making subsequent teaching feedback more stable. Attached Figure Description

[0024] Figure 1 This is a system architecture diagram according to an embodiment of the present invention.

[0025] Figure 2 This is a flowchart of a closed-loop intelligent teaching and training process according to an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See attached document Figure 1 This invention provides an intelligent teaching, training, and quantitative evaluation system for Peking Opera movements, including a teaching content management module, a learner interaction module, a practice data collection module, an intelligent perception and evaluation module, a teaching feedback and training plan adjustment module, and a teaching data storage module.

[0028] The teaching content management module manages the teaching resources used in teaching Pingjiang Opera movements, including a Pingjiang Opera standard movement library, a repertoire knowledge base, a Pingjiang Opera specialized training mapping table, and teaching and training plans corresponding to different learning stages. The Pingjiang Opera standard movement library stores standard demonstration videos corresponding to multiple Pingjiang Opera standard demonstration movements and establishes the association between standard demonstration videos, repertoire numbers, movement types, and teaching and training tasks.

[0029] The learner interaction module is deployed on the learner's terminal device. It determines the current teaching and training task based on the existing plan and presents the learner with the corresponding standard demonstration teaching content for the Pingjiang Opera. It also receives practice requests from learners and, after evaluating the movements, presents a comprehensive performance score, error correction guidance, and a highlighted skeleton comparison result. In subsequent learning cycles, the learner interaction module also presents specific training content tailored to the learner's persistent weaknesses, based on the adjusted teaching and training plan.

[0030] The practice data acquisition module includes at least a video acquisition device used to capture real-time video streams of learners as they follow the body movements in accordance with the standard demonstration teaching content. The video acquisition device can be a single camera or a multi-angle camera array, acquiring continuous video frames containing the learner's complete body movement area, depending on the deployment environment.

[0031] The intelligent perception and evaluation module includes a basic pose extraction network, a skeleton feature processing component, a motion timing alignment component, and a motion quantification evaluation component. The basic pose extraction network extracts key points of the human skeleton from standard demonstration videos and learner's real-time practice videos; the skeleton feature processing component performs local coordinate transformation, skeleton topology normalization, and joint angle calculation; the motion timing alignment component eliminates the temporal phase difference between the learner's motion rhythm and the standard demonstration motion rhythm; and the motion quantification evaluation component calculates the learner's local motion deviations and overall motion score based on standard motion evaluation benchmarks.

[0032] The teaching feedback and training plan adjustment module is used to identify typical error nodes that need to be corrected based on the local movement deviations in this exercise, and to generate teaching error correction guidance information for learners based on the typical error nodes; it is also used to identify learners' persistent movement weaknesses based on historical evaluation data formed over multiple learning cycles, select corresponding decomposed practice content from the Pingjiang Opera special training content based on movement weaknesses, and adjust subsequent teaching and training plans.

[0033] In one specific embodiment, the teaching feedback and training plan adjustment module can be implemented collaboratively by a planning agent, an execution agent, and a reflection agent. Specifically, the execution agent controls the teaching evaluation and feedback process corresponding to the current teaching and training task; the reflection agent identifies persistent weaknesses based on historical learning data; and the planning agent selects specific training content based on these persistent weaknesses and adjusts subsequent teaching and training plans.

[0034] The teaching data storage module includes a student profile database, which is used to establish a correspondence between learners, teaching tasks and practice time, and to save the comprehensive action scores, local action deviations, typical error nodes and teaching error correction guidance information generated in each practice session, and to support subsequent periodic learning analysis.

[0035] In some implementations, to facilitate collaborative teaching between body movements and singing in traditional opera, the system may also include an acoustic teaching support module. This module acquires standard demonstration audio synchronized with the standard demonstration video and real-time practice audio generated by the learner. It compares the acoustic characteristics of the standard demonstration audio and the real-time practice audio, and uses the acoustic evaluation results as an extended teaching feedback dimension for body movement instruction. The acoustic teaching support module is not a necessary condition for implementing body movement instruction and quantitative evaluation in this embodiment.

[0036] See attached document Figure 2 This invention provides a method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera, which is implemented through the aforementioned system and includes the following steps: S1. Determine the teaching and training task for this session, and present the standard demonstration teaching content of Pingjiang Opera corresponding to the teaching and training task to the learners through the learner's terminal. S2, establish standard action teaching evaluation benchmarks for standard demonstration teaching content; S3, in order to improve the accuracy of learner action recognition and teaching evaluation, adapts the basic posture extraction network pre-trained with action data from similar opera genres to the action features of the opera. S4: Obtain the real-time practice video stream generated by learners performing follow-up exercises based on the standard demonstration teaching content, perform time phase alignment and quantitative teaching evaluation of learners' real-time movements, and obtain local movement deviations and comprehensive movement scores. S5: Based on local movement deviations, identify the typical error nodes that learners need to correct in this teaching and training task, generate real-time teaching and error correction guidance information for the typical error nodes, and present the teaching evaluation results through the learner's terminal. S6. The learner's learning record is formed by the comprehensive score of the action corresponding to this teaching and training task, local action deviations, typical error nodes and teaching error correction guidance information, and written into the learner profile database. S7. Based on the historical learning records formed over multiple practice cycles, identify the learner's persistent weaknesses, determine the corresponding specialized training content based on the persistent weaknesses, and adjust the learner's subsequent teaching and training plan so that the learner can continue to implement targeted practice according to the adjusted teaching and training plan in the next practice cycle.

[0037] In some implementations, the system can also construct difficult example samples based on abnormal practice data collected during long-term operation, and perform supplementary fine-tuning on the basic posture extraction network when the difficult example samples meet the preset fine-tuning conditions, so as to maintain the accuracy of subsequent teaching evaluation results.

[0038] Step S1 specifically includes the following implementation steps.

[0039] S11, obtain the teaching and training plan corresponding to the current learner.

[0040] After a learner logs into the terminal and enters the Pingjiang Opera movement training interface, the learner interaction module retrieves the teaching and training plan corresponding to the current learner. The teaching and training plan includes multiple teaching task nodes arranged in a preset teaching order. Each teaching task node is associated with at least the corresponding song number, movement type, and standard demonstration teaching content.

[0041] For the initial learning stage, a teaching and training plan can be established according to the preset basic teaching syllabus and in order of increasing difficulty of the movements; for learners who have already formed historical learning data, the current teaching task is determined according to the teaching and training plan updated in step S7.

[0042] S12, Determine the teaching and training tasks for this session.

[0043] The learner interaction module reads the teaching task node to be executed in the teaching and training plan, determines the song number and action type corresponding to the teaching task node as the current teaching and training task, and retrieves the standard demonstration video of Pingjiang Opera corresponding to the song number and action type from the teaching content management module.

[0044] S13 presents standard demonstration teaching content.

[0045] The learner's terminal loads a standard demonstration video through a display interface, visually presenting the standard body movements of the Peking Opera to the learner. By observing the standard demonstration video, learners can obtain the standard movement process, rhythm, and body part movement status corresponding to this teaching task, and initiate a follow-up practice request after completing the movement observation.

[0046] In some implementations, the standard demonstration teaching content also includes standard demonstration audio synchronized with the standard demonstration video. The learner's terminal plays the standard demonstration video and standard demonstration audio simultaneously to help the learner understand the correspondence between body movements and performance rhythm.

[0047] Regarding step S2, before learners practice the movements, the system pre-establishes standard movement teaching evaluation benchmarks corresponding to the standard demonstration movements of different Pingjiang Opera.

[0048] S21, Collect and establish standard demonstration action samples for Pingjiang Opera.

[0049] The data acquisition device collects standard demonstration video streams from the instructors for different teaching tasks. The standard demonstration video streams are presented as a two-dimensional matrix sequence of continuous motion images arranged chronologically. In one specific embodiment, the image resolution of the standard demonstration video is set to 1920×1080 pixels, and the frame rate is set to 60fps.

[0050] For the same type of body movements in Peking Opera, multiple standard demonstration sequences can be collected from multiple demonstrators or from the same demonstrator performing the standard movements multiple times. After collection, the key points of the human skeleton in the standard demonstration video are labeled to form a sequence of absolute coordinates of the standard skeletal key points.

[0051] By establishing standard action evaluation benchmarks through multiple standard demonstration samples that conform to teaching norms, subsequent teaching evaluations of learners' actions can not only rely on the absolute human posture of a single demonstrator, but also further consider the reasonable range of motion variations allowed by the standard teaching actions themselves.

[0052] In some implementations, standard demonstration audio of Pingjiang Opera is also collected simultaneously, and corresponding Fu'an dialect text and phoneme annotation data are generated as training samples for subsequent optional acoustic teaching auxiliary models.

[0053] S22, Establish the local skeletal coordinate system for the standard demonstration movement.

[0054] The intelligent perception and evaluation module receives the standard demonstration video stream and analyzes the video sequence frame by frame.

[0055] The basic pose extraction network takes a single-frame RGB image matrix as input and extracts local spatial heatmap features through a two-dimensional convolutional module, then performs keypoint coordinate regression. In one specific embodiment, the basic pose extraction network outputs a three-dimensional absolute coordinate matrix containing 33 human skeletal keypoints.

[0056] After obtaining the absolute coordinates of all skeletal keypoints within the same video frame, the absolute coordinates of the left hip joint keypoint and the right hip joint keypoint are extracted and their average is calculated to generate the absolute coordinates of the pelvic center.

[0057] Using the absolute coordinates of the pelvic center as the origin of the local three-dimensional spatial coordinate system, the absolute coordinates of the other skeletal key points in the video frame are subtracted from the absolute coordinates of the pelvic center, and a spatial coordinate system translation operation is performed to generate the relative coordinates of each skeletal key point in the local three-dimensional spatial coordinate system.

[0058] S23, Perform skeleton topology normalization on the standard demonstration action.

[0059] The intelligent perception and evaluation module extracts the relative coordinates of the neck key points from the relative coordinate set, calculates the Euclidean distance between the neck key points and the origin of the local coordinate system, and defines the obtained distance as the length of the trunk vector.

[0060] To avoid the denominator being zero due to the output of coincident coordinates by the basic pose extraction network under occlusion conditions, a tolerance bias constant is introduced during the calculation process. The reciprocal of the sum of the backbone vector length and the tolerance bias constant is calculated, and this reciprocal is used as a scaling factor. The relative coordinates of each skeleton keypoint within the frame are then multiplied by the scaling factor to obtain a feature skeleton vector without dimensions. Normalization is performed using the following formula: ; In the formula, Represents the length of the main vector; This represents the tolerance bias constant, with a value of 10. -6 ; Representing the The relative coordinates of each skeletal key point in the local three-dimensional space coordinate system; Representing the The feature skeleton vector is generated after normalization calculation of each skeletal key point. By local coordinate system translation and skeleton topology normalization, the differences in height, torso size and absolute position in the camera frame of different standard demonstrators are eliminated, so that the standard features mainly reflect the skeletal topology and force state of the body movements of the performers.

[0061] S24 forms the standard characteristic variance distribution matrix.

[0062] The intelligent perception and evaluation module extracts adjacent body parts with connectivity attributes from the feature skeleton vector based on the connection topology of the human skeleton, and calculates the joint angle between adjacent bone nodes.

[0063] After completing the single-frame feature extraction, multiple standard demonstration samples corresponding to the same teaching action type are obtained, and sequence alignment is performed on different samples according to the action timeline.

[0064] The sequence alignment uses a dynamic time warping algorithm, which finds the optimal time alignment path by constructing a time warping cumulative distance matrix to eliminate frame-level time misalignment caused by different speeds and rhythms when different standard demonstrators perform the same action.

[0065] After time alignment, the skeleton vector spatial coordinate components and joint angles of different standard demonstration samples are extracted at the same action progress node, and the spatial distribution variance of each feature dimension among multiple standard demonstration samples is calculated. The spatial distribution variance is expressed by the following formula: ; In the formula, before calculating the spatial distribution variance, the skeleton coordinate components are kept as dimensionless normalized values, and the joint angles are divided by π to convert them into dimensionless angular features in the range of 0 to 1, so that different types of features are on a unified numerical scale. Representing the Spatial distribution variance of each feature dimension; This represents the total number of samples of the same type of action participating in the statistical calculation, and the value must be greater than or equal to 2. Representing the The nth sample in time-aligned state The feature values ​​of each feature dimension include the three-dimensional spatial coordinate components of the feature skeleton vector and the joint angle values. represent The sample at the th The arithmetic mean of the feature values ​​across each feature dimension; For feature dimension index, This refers to the sample sequence number used in the statistics. The value range is from 1 to N.

[0066] Before calculating the spatial distribution variance, the skeleton coordinate components are kept as dimensionless normalized values, and the joint angles are divided by π to convert them to the range of 0 to 1, so that different types of features are on a uniform numerical scale.

[0067] After calculating the variance of different feature dimensions, the variance values ​​of multiple spatial distributions are combined to form a standard feature variance distribution matrix according to the natural connection topology of human skeletal nodes, and the correlation between the standard feature variance distribution matrix and the corresponding teaching action type is established.

[0068] A smaller spatial distribution variance in the matrix indicates that the corresponding body part or joint angle shows high consistency across multiple standard demonstration samples, reflecting that the body part has relatively strict teaching standards for posture; a larger spatial distribution variance indicates that the corresponding body part is allowed to have a certain range of movement variations.

[0069] Therefore, the standard feature variance distribution matrix serves as the teaching evaluation benchmark for Pingjiangxi body movements, used to determine the evaluation rigor of different body parts during subsequent learner movement evaluation; it can also serve as the parameter basis for the basic posture extraction network to perform Pingjiangxi movement feature adaptation.

[0070] For step S3, in order to obtain the learner's movement characteristics required for quantitative teaching evaluation of movements and to improve the stability of human movement characteristics acquisition during the teaching evaluation of body movements in Pingjiang Opera, this embodiment uses a basic posture extraction network to extract key points of human skeleton from standard demonstration videos and learner real-time practice videos. The basic posture extraction network is pre-trained using movement data from similar opera genres, and then the basic posture extraction network is selectively and locally fine-tuned according to the standard movement teaching evaluation benchmark of Pingjiang Opera.

[0071] S31, load the cross-genre pre-training basic posture extraction network for motion quantification teaching evaluation.

[0072] The intelligent perception and evaluation module retrieves the basic posture extraction network from a pre-set model library. This network, used as a means of acquiring motion features in the quantitative teaching evaluation process, extracts key points of the human skeleton from standard demonstration videos and learners' real-time practice videos. These extracted key points are then used for subsequent skeletal feature processing, temporal phase alignment, local motion deviation calculation, and comprehensive motion scoring.

[0073] The basic pose extraction network is pre-trained using human body movement data from Peking Opera, Min Opera, or other similar opera genres, enabling the network to have basic recognition capabilities for human body contours, skeletal topology, and similar opera movement patterns.

[0074] In one specific embodiment, the basic pose extraction network is configured as a two-dimensional fully convolutional network structure including residual connection modules, and the input channel is a three-channel RGB image matrix.

[0075] The basic pose extraction network consists of a shallow general feature encoding module and a deep opera action feature decoding module. The shallow general feature encoding module includes convolutional layers and pooling layers to extract general visual features such as image edges, textures, and human body contours. The deep opera action feature decoding module includes multi-scale feature fusion and upsampling structures to form high-order pose semantic features corresponding to different human body key points and outputs a heatmap distribution of human body joints.

[0076] By using data from similar opera genres for pre-training, subsequent fine-tuning of the teaching evaluation model does not require building the ability to extract human posture features from scratch, thereby reducing the impact of the limited number of standard samples for a specific opera genre on the training of the teaching evaluation model.

[0077] S32, Generate network parameter update mask based on standard feature variance distribution matrix.

[0078] Read the standard feature variance distribution matrix established in step S2, extract the spatial distribution variance corresponding to different body parts or joint features, compare the spatial distribution variance with the preset generalization tolerance threshold, and generate the corresponding binary mask parameters based on the comparison result.

[0079] The mask parameters correspond to different feature channels in the deep opera action feature decoding module, and the parameter update status of different feature extraction branches is determined by the binarization indicator function.

[0080] The binarization indicator function uses the following formula: ; In the formula, Representative assigned to the first Each feature dimension corresponds to a mask parameter value for the network update channel, and the value state is limited to 0 and 1; The representative quoted from the calculation in step S1 is the first Spatial distribution variance of each feature dimension; Representing the generalization tolerance threshold, it is calibrated to 0.1 based on the test data of opera movement standardization, thus defining the judgment boundary between the rigid standardization and the flexible expansion range of the movement at the data evaluation level.

[0081] In this embodiment, the generalization tolerance threshold can be jointly calibrated based on the convergence value of the pre-training loss of the basic pose extraction network on the general human pose dataset, and combined with the standardization test data of the Pingjiangxi action, for example, by taking 0.1.

[0082] S33 controls the freezing and unfreezing of network branches based on the mask parameters.

[0083] Traverse the relevant feature extraction branches in the basic pose extraction network.

[0084] When the corresponding mask parameter is 0, the corresponding network branch is set to a frozen state, which blocks the gradient update of the convolution kernel weights and bias parameters inside the network branch during subsequent error backpropagation.

[0085] When the corresponding mask parameter is 1, the corresponding network branch is set to the unfrozen state, so that it can receive error gradients and perform parameter updates during the training of standard samples.

[0086] This allows the basic pose extraction network to form a hybrid network state where some general pose feature parameters remain stable, while some parameters related to the specific body movements of the opera are open for learning, in order to reduce overfitting caused by full parameter updates under small sample conditions.

[0087] S34 uses standard samples from Pingjiangxi to perform local fine-tuning on the feature extraction branch in the thawed state.

[0088] The standard demonstration video of Pingjiang Opera is read from the data storage module. The motion images in the standard demonstration video are input into the basic pose extraction network. The network performs forward propagation and outputs the predicted coordinates of the skeletal key points.

[0089] The absolute coordinates of the standard skeletal key points in the Pingjiang Opera, manually annotated by experts, are used as training labels. The mean squared error loss is used to calculate the difference between the predicted key point coordinates and the standard key point coordinates.

[0090] The mean square error is calculated using the following formula: ; In the formula, The value representing the mean squared error loss of attitude estimation; This represents the total number of skeletal keypoints output by the network, and its value is always 33. The first one represents the output of the basic pose extraction network. Predict the absolute coordinates of each key point; The first one, representing manually labeled The absolute coordinates of the key points in the Pingjiang Opera standard corresponding to each key point; This represents the square of the spatial L2 norm of the absolute coordinate difference vector within the interior, and the scattering state of the coordinates is predicted through the quantized feedback constraint of the spatial distance.

[0091] After calculating the loss value, backpropagation is performed. When the error gradient is passed to the deep opera action feature decoding module, the gradient corresponding to the frozen network branch is blocked according to the mask parameters formed in step S32, and it only propagates along the unfrozen network branch.

[0092] The network optimizer updates the weight parameters of the thawed network branches based on the received local error gradient, thereby improving the stability of the basic pose extraction network in acquiring the skeletal key points required for teaching and evaluation of specific body movements in Pingjiang Opera while retaining the ability to acquire general human pose features across different opera genres. This results in a basic pose extraction network for quantitative teaching and evaluation of movements.

[0093] After completing the aforementioned local fine-tuning, the basic posture extraction network is deployed in the intelligent perception and evaluation module. When learners perform follow-up exercises based on the standard demonstration teaching content, the basic posture extraction network receives video frames from the learner's real-time practice video stream and extracts the learner's real-time human skeleton key points. Based on the human skeleton key points, it further forms the learner's motion features, which are used for subsequent temporal phase alignment, local motion deviation calculation, and comprehensive motion scoring to complete the quantitative teaching evaluation of motion.

[0094] Therefore, the aforementioned cross-genre pre-training, binary mask parameter generation, feature extraction branch freezing and unfreezing, and local parameter fine-tuning all serve as the underlying motion feature acquisition process for motion quantification teaching evaluation, used to obtain a basic posture extraction network suitable for teaching evaluation of body movements in Peking Opera. The output of the basic posture extraction network is not the final teaching result of this embodiment, but is used to form learners' local motion deviations and comprehensive motion scores, and further used for real-time teaching error correction, historical learning analysis, and subsequent adjustment of specialized training content.

[0095] For step S4, after learners complete the observation of the standard demonstration teaching content, they perform actual follow-along practice. The basic posture extraction network, which was adapted to the movement features in step S3, serves as the underlying movement feature acquisition method for the quantitative teaching evaluation of movement in this step.

[0096] S41, Collect learner practice videos that are relevant to the current teaching task.

[0097] Learners can initiate follow-up exercises based on the standard demonstration video of the play presented on their learning terminals.

[0098] The learner interaction module maintains the association between the current teaching task number, the standard demonstration action type, and the practice session; the practice data acquisition module simultaneously starts the video acquisition device to acquire the real-time practice video stream generated when the learner performs this teaching task.

[0099] In one specific embodiment, the real-time practice video stream has an image resolution of 1920×1080 pixels and a capture frame rate of 30fps.

[0100] The real-time practice video stream and the current teaching task number are transmitted together to the intelligent perception and evaluation module, thereby determining the standard demonstration movements to be used in subsequent evaluations, and avoiding the incorrect use of standard benchmarks corresponding to different pieces or different body movements for the current learner's practice data.

[0101] The basic posture extraction network obtained in step S3 for motion quantification teaching evaluation receives real-time practice video frames, outputs the distribution of human joint heatmaps, and further decodes to obtain the real-time absolute coordinates of the learner's skeletal key points. The real-time absolute coordinates of the skeletal key points are used to form the learner's motion features required for subsequent temporal phase alignment and motion quantification teaching evaluation.

[0102] The absolute coordinates of the key points of the left and right hip joints are extracted and the mean is calculated to form the pelvic center coordinates of the learner's current video frame. Using the pelvic center coordinates as the origin of the local three-dimensional coordinate system, spatial translation transformation is performed on each real-time skeletal key point to obtain the relative coordinates of the learner's real-time skeletal key points.

[0103] S42 performs scale normalization on learner skeleton features.

[0104] Extract the neck key points from the learner's real-time relative coordinate set, calculate the Euclidean distance between the neck key points and the origin of the local three-dimensional spatial coordinate system, and obtain the length of the real-time backbone vector.

[0105] Following the normalization method in step S23, the scale scaling factor is obtained by combining the length of the real-time backbone vector with the tolerance bias constant. The scale scaling factor is then used to normalize the relative coordinates of all real-time skeleton keypoints to generate a real-time feature skeleton vector without size units.

[0106] By implementing the above methods, the additional distance error caused by differences in height, arm length, and body proportions between learners and standard demonstrators is reduced, allowing subsequent teaching evaluation to focus primarily on the movement and posture itself.

[0107] S43, align the learner's actions with the standard demonstration actions in terms of timing phase.

[0108] When learners practice following a standard demonstration, their speed of movement often differs from the standard demonstration due to factors such as proficiency, reaction time delay, and pauses in movement.

[0109] If comparisons are made directly based on the same video frame number, it may lead to incorrect matching of actions that the learner has not yet completed with actions that the standard demonstration has already completed, thus misjudging rhythm differences as errors in action standardization.

[0110] Therefore, the intelligent perception and evaluation module calls the standard demonstration action sequence corresponding to the current teaching task and assembles the learner's real-time feature skeleton vector into a real-time practice action sequence according to the time axis.

[0111] Considering the difference in initial capture frame rates between the standard demonstration video and the real-time practice video, the standard demonstration motion sequence is first downsampled to match its temporal resolution with that of the real-time practice motion sequence.

[0112] Extract the real-time practice action sequence respectively Real-time feature skeleton vector of frame and standard demonstration action sequence The standard feature skeleton vector of the frame is used to calculate the local spatial distance between the two, and the local spatial distance matrix is ​​formed by traversing the two action sequences.

[0113] The cumulative cost matrix is ​​constructed using the dynamic time warping algorithm, and the optimal time warping path is found under the constraint of a preset time search window.

[0114] The dynamic time warping calculation uses the following formula: ; In the formula, The coordinates in the cumulative cost matrix are The cumulative cost of the location; Represents the real-time practice action sequence number Frame and standard demonstration action sequence The numerical value of local spatial distance between frames; This represents the function operation of finding the minimum value; , and These represent the cumulative cost values ​​corresponding to adjacent preceding coordinate positions in the cumulative cost matrix; The time search window parameter represents the time search window parameter; in one specific embodiment, the time search window parameter is set to 15 frames, corresponding to an action time misalignment tolerance range of approximately 0.5 seconds.

[0115] After calculating the cumulative cost matrix, the process backtracks from the endpoint to the starting point along the direction of decreasing cumulative cost to obtain the time warping path. Based on the time warping path, time axis resampling is performed on the real-time practice action sequence and the standard demonstration action sequence to form synchronous teaching evaluation sequence pairs.

[0116] By aligning the time phase, rhythmic differences caused by learners' movement speed, reaction delay, and local pauses will not be directly judged as errors in body movements, thereby improving the ability of teaching evaluation to reflect the actual standardization of learners' movements.

[0117] S44. Determine the evaluation weights for different movement characteristics based on standard teaching specifications.

[0118] Read the standard feature variance distribution matrix formed in step S24, and extract the spatial distribution variance corresponding to different body parts and joint features in the synchronous teaching evaluation sequence.

[0119] The inverse variance mapping is applied to the spatial distribution variance to obtain the initial evaluation weights. Then, a sum-normalization is performed on all initial evaluation weights to obtain the dynamic evaluation weights. The inverse variance mapping is calculated using the following formula: ; In the formula, Representing the Initial weight coefficients for each feature dimension; The representative quoted from the calculation in step S1 is the first The spatial distribution variance values ​​of each feature dimension; This represents the tolerance bias constant, with a value set to 1×10. -6 This is to avoid numerical overflow when the spatial distribution variance is close to zero.

[0120] Weight normalization is performed using the following formula: ; In the formula, Representing the Dynamic weight coefficients corresponding to each feature dimension; This represents the total number of feature dimensions involved in the evaluation calculation; This represents the sum of the initial weight coefficients for all feature dimensions. The sum-normalization calculation limits the numerical range of all dynamic weight coefficients to between 0 and 1, while maintaining the mathematical constraint that the sum of the dynamic weight coefficients for all dimensions is 1.

[0121] For key body parts that exhibit high consistency across multiple standard demonstrators, with relatively small spatial distribution variance, corresponding to stricter teaching standards for Pingjiang Opera, a higher evaluation weight is assigned in the learner's movement evaluation. For body parts where reasonable differences in movement are permissible between different standard demonstrators, the evaluation weight is appropriately reduced.

[0122] This means that teaching evaluation no longer uses the same fixed weights to evaluate all body parts, but rather dynamically determines the rigor of the evaluation based on the standardized distribution of the standard movements in the Pingjiang Opera.

[0123] S45, calculate local movement deviation and overall movement score.

[0124] For each pair of aligned video frames in the synchronous teaching evaluation sequence, the learner's real-time feature skeleton vector and the standard feature skeleton vector are extracted respectively.

[0125] Based on the torso connection direction features, the cosine similarity between the real-time action direction vector and the standard action direction vector is calculated, and the cosine similarity is converted into normalized cosine distance.

[0126] Based on the joint flexion and extension characteristics, the learner's joint angle and the joint angle corresponding to the standard demonstration movement are extracted, the absolute difference between the two angles is calculated, and the difference is divided by π to obtain the normalized angle deviation.

[0127] Using the dynamic evaluation weights formed in step S44, a weighted linear combination is performed on the normalized cosine distance and the normalized angle deviation to form the local motion deviation corresponding to the current video frame.

[0128] The following formula is used to calculate local motion deviation: ; In the formula, This represents the local deviation value corresponding to the currently aligned video frame; The total number of feature dimensions involved in the evaluation; The representative quotes the first generation calculated from step S34. Dynamic weight coefficients corresponding to each feature dimension; Represents the first in the real-time feature skeleton vector The direction vector of the line connecting the torso; The first matching element in the standard feature skeleton vector The direction vector of the line connecting the torso; This represents the operation of finding the Euclidean norm of a vector space; This represents the tolerance bias constant, with a value set to 1×10. -6 ; Represents the first in the real-time feature skeleton vector The joint angle values ​​corresponding to each feature dimension are expressed in radians, with the unit symbol being rad. The value of the joint angle corresponding to the matching in the standard feature skeleton vector is represented by rad. Represents the constant value of pi; and These represent the linear balance coefficients of the normalized cosine distance and the normalized angle deviation, respectively. Skeletal spatial orientation and joint flexion-extension state have equal evaluation weight, and both are set to 0.5.

[0129] By traversing all synchronized video frames, a local motion deviation sequence on the timeline is obtained. This local motion deviation sequence is then subjected to temporal integral averaging and a percentage-based inverse mapping to obtain the comprehensive motion score for this teaching task. The comprehensive motion score uses the following formula: ; In the formula, The overall score representing the actions generated by the mapping; This represents the total number of aligned video frames contained in the synchronous evaluation sequence pair; Representing the The local deviation value corresponding to each aligned video frame. This represents performing a temporal-dimensional summation operation on the local deviation values ​​of all aligned video frames; This represents the operation of a function that takes non-negative extreme values. It provides learners with a quantitative evaluation of their overall performance in body movements through a comprehensive action score, and also provides a basis for generating targeted error correction guidance by analyzing local action deviations at various time points.

[0130] For step S5, based on the local movement deviation and comprehensive movement score obtained in step S4, instant teaching feedback that can be directly used for learner movement correction is generated.

[0131] S51, identify typical teaching error points that need to be corrected for learners.

[0132] The teaching feedback and training plan adjustment module traverses the local motion deviation and specific motion feature dimension deviation corresponding to each video frame in the synchronous teaching evaluation sequence.

[0133] The deviation values ​​of different motion characteristics are compared with a preset warning tolerance threshold. In one specific embodiment, the warning tolerance threshold can be set in the range of 0.15 to 0.20, for example, 0.15.

[0134] When the deviation of at least one motion feature dimension in a video frame exceeds the warning tolerance threshold, the corresponding video frame is identified as an abnormal motion frame.

[0135] To avoid generating numerous repeated prompts for the same error during continuous actions, abnormal action frames are clustered according to a preset time window. In one specific embodiment, the time window is set to 15 consecutive frames, corresponding to approximately 0.5 seconds.

[0136] Abnormal action frames within the same time window are grouped according to different body parts or feature dimensions, and the abnormal action frames with the largest local action deviations are retained as typical error nodes in each group.

[0137] For each typical error node, extract the corresponding timestamp, the name of the body part or joint where the deviation occurred, the direction of the deviation, and the degree of the deviation.

[0138] Typical error nodes indicate the specific body parts and time stages of the action that learners need to correct in the current teaching task.

[0139] S52 generates teaching error correction guidance information based on typical error nodes.

[0140] The timestamps, joint names, movement deviations, and overall scores of the typical error nodes are encapsulated according to the preset key-value correspondence specifications to form structured teaching evaluation data.

[0141] In one specific embodiment, a JSON data structure can be used to form a data slice sequence, and the data slice sequence can be embedded in a structured prompt word template.

[0142] The large language model deployed on the server is invoked, and the structured teaching evaluation data is input into the large language model.

[0143] The large language model can employ a decoder network structure based on a multi-layer masked self-attention mechanism. During the model training phase, error correction guidance texts from Pingjiangxi teaching experts are pre-collected, and a training sample relationship is established between objective motion capture deviation data and corresponding standard error correction phrases. Through fine-tuning of instructions, the large language model learns the mapping relationship between motion deviation data and Pingjiangxi teaching terminology.

[0144] When performing real-time teaching feedback, the data slice sequence generated in step S51 is used as a mandatory context constraint to construct a structured prompt word template that includes system roles, objective action deviation data, and generation rules.

[0145] The generation rules limit the large language model to generate error correction content only for actual action time nodes and body parts in the data slice sequence, and make the degree of action adjustment described in the text match the actual deviation value.

[0146] In one specific embodiment, the output sampling temperature of the large language model can be set to 0.2, and the kernel sampling parameter can be set to 0.8 to reduce random output.

[0147] The large language model outputs corresponding teaching correction guidance text, which transforms the skeletal posture deviation, which was originally only represented by numerical values, into natural language teaching information containing specific movement parts, corresponding time nodes, and movement adjustment directions, making it easier for learners to directly implement movement correction according to the teaching correction guidance information.

[0148] S53 creates a visual teaching feedback interface on the learner's terminal.

[0149] Based on the timestamps corresponding to typical error nodes, the learner's action footage and the standard demonstration action footage at the corresponding moments are extracted from the real-time practice video and the standard demonstration video, respectively.

[0150] The learner's real-time skeletal keypoint connection relationship and the standard demonstration skeletal connection relationship are overlaid in the two action frames respectively to generate a skeleton overlay layer.

[0151] Different skeletal parts are displayed differently based on local motion deviations.

[0152] In one specific embodiment, skeletal connections that do not exceed the warning tolerance threshold are displayed in a conventional manner; skeletal connections and joint positions that exceed the warning tolerance threshold are displayed in a warning manner, and the deviation angle or deviation value is presented at the corresponding position, thereby generating a skeleton highlight comparison map.

[0153] The learner terminal displays the standard demonstration action screen, the learner's practice action screen, the skeleton highlight comparison image, the comprehensive action score, and the teaching error correction guidance text in a preset teaching feedback format.

[0154] As a result, learners can simultaneously observe the standard movements, their own actual movements, and the specific body parts that need correction, and then re-perform the corresponding body movements based on the teaching and error correction guidance information.

[0155] In some implementations, if the acoustic teaching aid module is also enabled, the acoustic feature comparison results corresponding to the standard demonstration audio and the learner's real-time practice audio can be displayed synchronously in the teaching feedback interface.

[0156] For step S6, after providing immediate teaching feedback to the learner, the system will generate a structured learning record of the evaluation results of this exercise.

[0157] Extract the teaching task number, song number, practice time, overall movement score, local movement deviations at each time point, typical error points, and corresponding teaching correction guidance information, and establish a relationship with the current learner.

[0158] The evaluation data corresponding to each video frame is reorganized into a nested document structure according to time sequence. In one specific embodiment, a learner's practice profile can be constructed using a JSON structure, and a time-series evaluation array can be created in the secondary data nodes, with the timestamp of the corresponding video frame used as the array element index.

[0159] Under different timestamp indexes, normalized angle deviation, normalized cosine distance, typical error node identifiers, and teaching error correction guidance texts can be attached.

[0160] The learner practice profiles that have completed the structured encapsulation are written into the learner profile database.

[0161] To improve data write stability, asynchronous non-blocking input / output interfaces and a local circular queue can be used as write buffers. Newly generated practice files first enter the local buffer queue, and the data is released after receiving a write success response from the student profile database; if no success response is received within a preset waiting time, the write operation is re-executed.

[0162] In one specific embodiment, the response wait time can be set to 500 milliseconds.

[0163] By continuously recording learning data corresponding to different teaching tasks and practice cycles, the student profile database forms a continuous learning status record that reflects the long-term changes in learners' body movements during opera performances, providing a data foundation for subsequent continuous weakness analysis and adjustments to specific teaching plans.

[0164] After accumulating learner learning data over multiple practice cycles, the system executes step S7, which further transforms the evaluation results of the preceding actions into a basis for selecting subsequent teaching content.

[0165] This process includes acquiring learning data in stages, identifying persistent weaknesses, and adjusting specific training plans.

[0166] S71, Obtain historical learning data within the preset teaching cycle. The system's global scheduler initiates phased learning analysis according to a preset learning cycle.

[0167] In one specific embodiment, the preset learning period can be set to 7 calendar days. The time-series task engine wakes up the reflective agent used for learning analysis according to the set period.

[0168] The reflective agent extracts all learner practice files within the corresponding time range from the learner profile database based on the target learner and the preset learning cycle derived from the current time.

[0169] By traversing each practice file, we can extract the overall score of the corresponding teaching task and the local deviations in the movement.

[0170] Considering that learners may accidentally exit or experience network interruptions during actual practice, resulting in incomplete practice records, the extracted historical records are effectively cleaned.

[0171] In one specific embodiment, based on a 30fps video capture benchmark, single practice records with a total number of video frames less than 300 are marked as incomplete records and excluded from subsequent statistical processes.

[0172] For the learning archives retained after cleaning, the comprehensive scores of the actions corresponding to different exercises are spliced ​​together according to the absolute time sequence of the practice sessions to form a one-dimensional vector of historical scores that reflects the learner's stage-by-stage skill changes.

[0173] S72, Identify persistent weaknesses based on historical learning data. Learners' scores may fluctuate during a single practice session due to changes in attention, physical condition, or occasional mistakes. Therefore, a single low score should not be used to determine long-term motor weaknesses.

[0174] The reflective agent performs an exponential moving average filter on the one-dimensional vector of historical score time series.

[0175] The exponential moving average is calculated using the following formula: ; In the formula, The sequence number for the practice session; Set as the current number The smoothed score value corresponding to each practice session; Represents the current number extracted from the preceding sequence. Overall score of actions corresponding to this practice session ; Indicates the first The smoothed score value corresponding to each practice session; This is the set smoothing attenuation coefficient. The smoothing attenuation coefficient is then adjusted based on empirical data. The value is set in the range of 0.2 to 0.3. In this example, 0.25 is used so that the practice evaluation results that are more recent in time will receive a relatively higher calculation weight.

[0176] After filtering, a smooth scoring sequence is formed.

[0177] In conjunction with the standard teaching syllabus for Pingjiang Opera, a benchmark value for skill qualification is set. In one specific embodiment, the benchmark value for skill qualification is set at 75 points.

[0178] If the number of valid practice records in the current teaching cycle is less than the preset statistical requirements, the determination of persistent weaknesses will not be carried out temporarily.

[0179] In one specific implementation, when there are fewer than 5 valid scoring nodes, the system waits for subsequent practice data to continue accumulating.

[0180] When five consecutive smooth scoring points are detected to be below the skill qualification score benchmark, the corresponding time range is determined as the learner's action bottleneck stage.

[0181] Based on the time range corresponding to the bottleneck stage of the movement, the typical error nodes and local movement deviation data within that time interval are retrieved in reverse, and the frequency of error occurrence is statistically analyzed according to the body part or movement feature dimension.

[0182] Specific body parts or joint movements that occur frequently enough to meet preset conditions are identified as fixed error points, and the movement features corresponding to these fixed error points are structured and encapsulated as persistent weakness features.

[0183] Persistent weaknesses are used to characterize specific body movements or body parts that learners have not yet reached the preset teaching requirements after multiple practice cycles, rather than directly identifying a single, occasional error as a long-term learning weakness.

[0184] S73, the corresponding standard demonstration teaching content is re-called according to the adjusted teaching and training plan.

[0185] Once learners' persistent weaknesses are identified, the teaching feedback and training plan adjustment module no longer only outputs evaluation results, but also uses persistent weaknesses as the basis for controlling the selection of teaching content for the next teaching cycle.

[0186] The reflective agent sends persistent weaknesses to the planning agent.

[0187] The teaching content management module pre-establishes a mapping table for Pingjiang Opera special training.

[0188] The Pingjiang Opera Specialized Training Mapping Table uses the movement characteristics corresponding to the human skeleton lines, joint positions, or functional areas of the body as search conditions, and associates different movement weaknesses with decomposed practice pieces that target the corresponding local force exertion techniques or movement specifications.

[0189] The planning agent uses the specific feature dimensions corresponding to persistent weaknesses as input conditions and matches them in the Pingjiangxi special training mapping table.

[0190] When a directly matching motion feature is found, the corresponding decomposed practice piece number is extracted.

[0191] When there is no completely consistent feature in the specific training mapping table for the current persistent weakness feature, the specific action feature is mapped to the next higher level of body function area according to the subordinate relationship of human body structure, and the corresponding comprehensive special review content is searched within the expanded search scope.

[0192] For example, a specific weakness in the movement of the right shoulder and elbow joints can be further mapped to the upper limb movement area, and a comprehensive exercise program targeting upper limb strength can be found.

[0193] After determining the target-specific training content, the planning agent reads the teaching and training plan queue corresponding to the current learner.

[0194] The teaching and training plan queue is arranged according to the difficulty of the movements in the basic teaching syllabus or the preset teaching sequence under normal circumstances.

[0195] For the identified persistent weaknesses, the next regular teaching task in the current queue is temporarily suspended, and the teaching task node corresponding to the special training content is inserted into the priority execution position in the current teaching and training plan.

[0196] After completing the adjustment of the teaching and training plan, the updated teaching and training plan will be sent to the corresponding learner's terminal.

[0197] When a learner initiates the next practice session, the learner's terminal first reads the adjusted teaching and training plan, and prioritizes loading the specific breakdown practice pieces corresponding to persistent weaknesses, as well as the corresponding standard demonstration teaching content.

[0198] Learners practice again based on the newly presented specialized training content, and the system re-executes the processes of motion video capture, posture extraction, temporal phase alignment, motion teaching evaluation, and real-time error correction feedback.

[0199] In some implementations, to ensure the accuracy of motion teaching evaluation during long-term use of the intelligent teaching and training system, auxiliary iterative optimization can be performed on the basic posture extraction network based on long-tail anomaly data generated by learners' actual practice.

[0200] This process serves as a maintenance mechanism for the underlying teaching evaluation model and is independent of the teaching process in step S73, which involves adjusting the teaching and training plan based on learners' persistent weaknesses.

[0201] When analyzing historical learning data, the reflective agent can also simultaneously detect abnormal local motion deviations and abnormal spatial displacement of video frames.

[0202] A difficulty deviation threshold is set based on the normal range of motion of human skeletal nodes. In one specific embodiment, the difficulty deviation threshold can be set in the range of 0.35 to 0.45.

[0203] When the motion feature deviation in a video frame exceeds the hard case deviation threshold, or the relative displacement of the human body's two-dimensional coordinates in adjacent timestamp video frames exceeds the preset pixel displacement tolerance, the corresponding video frame will be identified as a candidate hard case that exceeds the conventional parsing capability of the current pose extraction network.

[0204] The pixel shift tolerance can be set in the range of 30 to 50 pixels, for example, 40 pixels.

[0205] Abnormal video frames that meet the criteria are extracted, and the original abnormal image, absolute timestamp, and corresponding skeletal key point coordinates are encapsulated into a difficult example sample data package and stored in the long-tail data cluster.

[0206] The system's global scheduler monitors the cumulative status of difficult sample data.

[0207] In one specific embodiment, the fine-tuning trigger baseline is set to 10,000 sample frames. When the total number of difficult examples reaches the fine-tuning trigger baseline, the model fine-tuning service is initiated.

[0208] The model fine-tuning service uses a visual basic network model based on the Transformer architecture to predict the coordinates of key points in difficult examples with high confidence. The obtained key point coordinates are used as ground truth labels to form a fine-tuning supplementary dataset.

[0209] During the fine-tuning phase, the shallow general feature encoding module in the basic pose extraction network is frozen, and only the high-order semantic feature fusion parameters in the deep opera action feature decoding module are opened.

[0210] The weighted mean square error function is used as the model fine-tuning loss function: ; In the formula, This represents the overall loss value output during the network fine-tuning phase. This indicates the total number of fine-tuning samples contained in the current training batch. The system forcibly sets the lower limit of the batch size to 16, avoiding the risk of computational overflow caused by the denominator approaching 0. This indicates the sequence number of the fine-tuned sample within the current training batch; This indicates that the basic pose extraction network is for the first... The predicted real-time skeletal keypoint coordinates of the current output of a difficult example screen. Indicates that the calibration channel is the first The true values ​​of precise keypoint coordinates assigned to a difficult sample; Set as the weighting coefficient for sample difficulty. The calculation formula is set as follows ,in This represents the normalized angle deviation corresponding to the current data frame recorded in the preceding mining stage. It uses a natural constant... A positive correlation mapping is established for the exponential function of the base, so that edge samples with larger initial calculation deviations receive higher penalty weights in the gradient descent iteration calculation.

[0211] In one specific embodiment, the lower limit of the fine-tuning batch size is set to 16, and the learning rate is set to 1×10⁻⁶. -5 .

[0212] Once the overall training loss curve converges, the updated base pose is output, network weights are extracted, and deployed to the front-end teaching evaluation service.

[0213] Through this auxiliary network optimization mechanism, the stability of acquiring human movement features required for quantitative teaching evaluation of movement is maintained under special clothing obstruction, unconventional lighting, or complex movement conditions, thereby improving the reliability of learners' movement teaching evaluation results and subsequent teaching error correction feedback; the auxiliary network optimization mechanism is only used as the underlying maintenance method of movement teaching evaluation and does not change the intelligent teaching and training process of determining persistent weaknesses based on learners' historical learning progress and adjusting subsequent teaching and training plans.

[0214] In some implementations, to help learners grasp the coordination between body movements, singing style, and rhythm in Peking Opera, the system can also add acoustic teaching assistance functions on the basis of the above-mentioned body movement teaching and training process.

[0215] The data acquisition module synchronously acquires standard demonstration audio of Pingjiang Opera and real-time practice audio of learners.

[0216] The general acoustic model is pre-trained using audio data from similar opera genres and employs a sequence processing structure that includes one-dimensional convolutional layers and long short-term memory network layers.

[0217] First, the standard audio time signal is divided into fixed-length audio frames, for example, 25ms in length, with a frame shift of 10ms. Then, a short-time Fourier transform is performed on each fixed-length audio frame, and the Mel frequency cepstral coefficient characteristic matrix is ​​calculated.

[0218] The general acoustic model receives the Mel frequency cepstral coefficient feature matrix, performs temporal convolution and temporal dependency calculation, and outputs the predicted phoneme probability distribution corresponding to different audio moments.

[0219] Artificial phoneme annotations corresponding to the standard text of the Fu'an dialect were used as acoustic training labels, and cross-entropy loss was used to calculate the difference between the predicted phoneme probability distribution and the standard phoneme labels.

[0220] The cross-entropy loss is calculated using the following formula: ; In the formula, The cross-entropy loss value represents the acoustic calculation. This represents the total number of phoneme classification categories preset in the acoustic model; The first one in the real label data The one-hot encoded state value of each phoneme category, with the actual phoneme category corresponding to the annotation having a value of 1 and the other categories having a value of 0; The first, representing the output of the general acoustic model The predicted probability distribution values ​​for each phoneme category. To avoid numerical overflow caused by the independent variable approaching 0 during logarithmic operations, a positive minimum tolerance constant is pre-stacked onto the predicted probability distribution values ​​during the calculation.

[0221] By using full-parameter backpropagation, the general acoustic model is gradually adapted to the specific pronunciation characteristics of the Fu'an dialect, resulting in an acoustic teaching aid model for Pingjiang Opera.

[0222] When learners practice body movement, the acoustic features corresponding to the learners' real-time practice audio can be extracted simultaneously, compared with the acoustic features corresponding to the standard demonstration audio along a unified time axis, and a dual-track acoustic feature comparison chart can be generated.

[0223] The learner terminal can present the acoustic feature dual-track comparison map together with the comprehensive action score, the skeleton highlight comparison map, and the teaching error correction guidance text.

[0224] Acoustic teaching aids only expand the dimensions of teaching feedback and do not change the body-based intelligent teaching process in this embodiment, which is centered on standard movement demonstration, learner practice, movement teaching evaluation, real-time error correction, continuous weakness identification, and adjustment of specific training plans.

[0225] When a learner exhibits a persistent weakness of insufficient left arm elevation across multiple cycles, the teaching feedback and training plan adjustment module identifies the corresponding decomposed action from the specialized training mapping table, inserts the specialized action into the priority task of the next teaching cycle, and re-presents the corresponding standard demonstration teaching content to the learner's terminal.

[0226] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera, characterized in that, Includes the following steps: The teaching and training tasks are determined according to the teaching and training plan, and the standard demonstration teaching content is presented on the learner's terminal. The system acquires a real-time practice video stream generated by learners following the standard demonstration teaching content, performs time phase alignment and quantitative teaching evaluation on the learners' real-time movements in the real-time practice video stream, and obtains local movement deviations and comprehensive movement scores. Based on the local motion deviations, instructional error correction guidance text is generated and presented by the learner's terminal; The comprehensive score of the movement and the local movement deviation are written into the student profile database. The persistent weakness characteristics are determined based on the historical learning records in the student profile database. The specific training content is determined based on the persistent weakness characteristics, and the subsequent teaching and training plan is adjusted for the determination of the standard demonstration teaching content for the next teaching cycle. The establishment of the standard characteristic variance distribution matrix includes: Obtain standard demonstration videos; Extract the absolute coordinates of the skeletal key points from the standard demonstration video, and generate the relative coordinates of the skeletal key points through a spatial coordinate system translation operation; Perform skeleton topology normalization on the relative coordinates of the skeletal key points to generate standard feature skeleton vectors, and calculate joint angles; Sequence alignment is performed on multiple standard demonstration samples of the same type, spatial distribution variance is calculated, and the samples are combined to form the standard feature variance distribution matrix.

2. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 1, characterized in that, To obtain the learner movement features required for the quantitative teaching evaluation of the movements, a basic posture extraction network is used to extract key points of the human skeleton. The basic posture extraction network is pre-trained with movement data from similar theatrical genres. The spatial distribution variance is compared with the generalization tolerance threshold to generate binary mask parameters; The generalization tolerance threshold is jointly calibrated by the pre-training loss convergence value of the basic pose extraction network on a general human pose dataset and the standardization test data of the storytelling action. When the binary mask parameter is 0, the feature extraction branch corresponding to the binary mask parameter is frozen; when it is 1, the feature extraction branch is unfrozen. By using standard samples from Pingjiang Opera to perform local fine-tuning on the feature extraction branch, a basic posture extraction network for the quantitative teaching evaluation of the movement is obtained.

3. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 2, characterized in that, The time phase alignment includes: Extract the relative coordinates of real-time skeletal key points from the real-time practice video stream to generate a real-time feature skeleton vector; The standard demonstration action sequence is downsampled, and the local spatial distance between the real-time feature skeleton vector and the standard feature skeleton vector is calculated to construct a local spatial distance matrix. The time warping path is calculated using a dynamic time warping algorithm, and time axis resampling is performed based on the time warping path to form synchronous teaching evaluation sequence pairs.

4. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 3, characterized in that, Perform inverse mapping and summation normalization on the spatial distribution variance in the standard feature variance distribution matrix to obtain the dynamic evaluation weights; Based on the aforementioned synchronous teaching evaluation sequence, the normalized cosine distance and normalized angle deviation are calculated; The local action deviation is formed by performing a weighted linear combination of the normalized cosine distance and the normalized angle deviation using the dynamic evaluation weights. The local action deviation sequence is subjected to time-series integral averaging to obtain the comprehensive action score.

5. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 4, characterized in that, The generation of the teaching error correction guidance text includes: Typical error nodes are determined based on the local action deviations; Extract the timestamps, joint names, local action deviations, and comprehensive action scores of the typical error nodes to form structured teaching evaluation data; The structured teaching evaluation data is divided into data slice sequences and embedded with structured prompt word templates. The structured prompt word template is input into the large language model to generate the teaching error correction guidance text.

6. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 5, characterized in that, Based on the timestamps of the typical error nodes, learner action images and standard demonstration action images are obtained, and a skeleton overlay layer is generated. A skeleton highlight comparison image is generated based on the local motion deviation; The learner terminal displays the standard demonstration action screen, the learner's action screen, the skeleton highlight comparison image, the action comprehensive score, and the teaching error correction guidance text.

7. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 1, characterized in that, Extract comprehensive action scores within a preset teaching cycle from the student profile database to form a one-dimensional vector of historical score time sequence. An exponential moving average filter is applied to the historical time-series one-dimensional vector of scores to form a smoothed score sequence; Set skill qualification benchmark values ​​based on the standard teaching syllabus for Pingjiang Opera; A bottleneck stage in the action is determined when five consecutive smooth scoring nodes are below the skill qualification scoring benchmark value. Based on the typical error nodes of the action bottleneck stage, solidified error points are identified and the persistent weakness features are formed.

8. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 7, characterized in that, Using the specific feature dimensions of the persistent weaknesses as input conditions, the content of the specialized training is determined by matching in the Pingjiangxi specialized training mapping table. Insert the teaching task nodes corresponding to the specific training content into the priority execution position of the teaching and training plan; The updated teaching and training plan will be sent to the learner's terminal. The learner's terminal will prioritize loading the standard demonstration teaching content of the specific training content in the next practice session.

9. The method for intelligent teaching, training, and quantitative evaluation of body movements in Peking Opera according to claim 8, characterized in that, When the persistent weakness feature does not have a completely consistent feature item in the Pingjiangxi Special Training Mapping Table, the specific action feature of the persistent weakness feature is mapped to the next higher level of body function area according to the subordinate relationship of human body structure. Search within the expanded search scope for comprehensive and specialized review content for the aforementioned higher-level bodily functional areas.

Images