Traditional sports skill action acquisition and evaluation system and method based on visual recognition

Abstract

The present application belongs to the field of computer vision and human motion recognition, and specifically relates to a traditional sports skill action collection and evaluation system and method based on visual recognition. The system includes: video collection and shooting guidance, posture estimation, key point preprocessing, action segmentation and rhythm analysis, traditional skill action knowledge base and index system, rule evaluation and risk judgment, aging adaptation stratification and personalized threshold, feedback and report, and data and compliance unit. The system processes human key points and normalizes them, identifies action stages and rhythms, conducts multi-dimensional evaluation in combination with expert-level features, and dynamically adjusts thresholds for different physical conditions. The present application can accurately capture the essentials of relaxation, opening and closing, and center of gravity conversion in skills, provide error correction suggestions based on anatomical logic, improve teaching professionalism, reduce the risk of sports injuries, and enhance system robustness and privacy compliance.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and human motion recognition, specifically relating to a system and method for collecting and evaluating traditional sports movements based on visual recognition. Background Technology

[0002] With the continuous evolution of computer vision and artificial intelligence technologies, motion analysis technology based on visual recognition has been widely applied in the fields of smart sports and rehabilitation training. Traditional sports exercises, as an important component of the national fitness system, require scientific and digital data collection and evaluation to improve practice efficiency and the quality of their transmission. By constructing a key point model of the human skeleton and combining it with real-time posture estimation algorithms, a digital evaluation system can capture the movement trajectory and morphological characteristics of practitioners, providing technical support for the standardized practice of traditional sports exercises.

[0003] An automated evaluation system for traditional physical exercises such as Baduanjin (Eight Pieces of Brocade) needs to transform complex limb movements into quantifiable feature parameters to achieve dynamic monitoring of the practice process. The core of this technology lies in extracting key point sequences of the human body using deep learning models and comparing them with pre-defined exercise movement logic to comprehensively evaluate the practitioner's movement completion, coordination, and stability in both temporal and spatial dimensions.

[0004] Existing assessment methods often focus on the geometric alignment of general fitness movements, lacking an indicator system specific to the unique movement principles of traditional exercises, such as relaxation, opening and closing, upright posture, and weight transfer. This leads to a disconnect between feedback information and actual teaching needs. Current systems do not adequately consider the physiological differences of core audiences, such as the middle-aged and elderly. Their rigid error correction logic and inflexible threshold settings are difficult to adapt to individual joint mobility, easily resulting in misjudgments or increased risk of sports injuries. Furthermore, traditional systems exhibit poor robustness in complex lighting environments and often neglect the coordinated assessment of breathing rhythm and mental focus within the exercises. In addition, the lack of class-based statistics and compliant data storage mechanisms for teaching management scenarios makes it difficult to meet the needs of professional teaching and large-scale applications.

[0005] Therefore, there is a need for a system and method for collecting and evaluating traditional sports movements based on visual recognition. Summary of the Invention

[0006] The purpose of this invention is to provide a traditional sports exercise movement acquisition and evaluation system based on visual recognition, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A visual recognition-based system for collecting and evaluating traditional sports exercises includes a video acquisition and shooting guidance unit, a posture estimation unit, a key point preprocessing unit, a movement segmentation and rhythm analysis unit, a traditional sports exercise movement knowledge base and indicator system unit, a rule evaluation and risk assessment unit, an age-appropriate stratification and personalized threshold unit, a feedback and reporting unit, and a data and compliance unit. The video acquisition and shooting guidance unit is used to acquire real-time video streams or record short clips, and is configured to detect the shooting environment in real time. When it detects that the light intensity is lower than the preset brightness threshold, the proportion of the human body in the picture is lower than the preset space proportion threshold, or the key parts of the human body are outside the frame, it automatically generates corresponding environmental adjustment guidance information and camera position adjustment instructions to ensure the quality of subsequent image processing. The pose estimation unit is configured to parse the image frames acquired by the video acquisition and shooting guidance unit, extract the human skeleton key point sequence, and generate a confidence parameter representing the detection reliability for each key point, thereby converting the dynamic video image into structured spatial coordinate features. The key point preprocessing unit is used to receive the key point sequence output by the pose estimation unit and perform dynamic denoising and smoothing, missing key point completion and scale normalization operations. By referring to the length of the human torso, the key point coordinates at different shooting distances are mapped to a unified logical space coordinate system. The movement segmentation and rhythm analysis unit is used to identify specific movement stages in traditional sports exercises based on the coordinate sequence provided by the key point preprocessing unit, including but not limited to the starting stage, unfolding stage, holding stage and retraction stage, and to calculate the movement speed, holding time and limb shaking frequency of each stage to reflect the continuity and stability of the practitioner's movements. The traditional exercise movement knowledge base and indicator system unit configuration is used to store expert-level standard movement characteristics and core error correction indicators for different traditional sports exercises, and to solidify teaching experience into quantifiable angle thresholds, distance parameters, trajectory characteristics and rhythm patterns, providing a benchmark for subsequent evaluation. The rule evaluation and risk assessment unit is used to combine the traditional exercise movement knowledge base and indicator system unit to perform multi-dimensional calculations on the features output by the movement segmentation and rhythm analysis unit, evaluate the degree of limb alignment, the size of movement amplitude, left and right symmetry and the trajectory of center of gravity transfer, and generate risk warning labels for postures that do not conform to physiological and mechanical characteristics. The age-appropriate stratification and personalized threshold unit configuration is used to dynamically adjust the logical judgment threshold in the rule evaluation and risk judgment unit according to the user's age range, discomfort area information and training stage, so as to realize differentiated evaluation logic adaptation for people with different physical conditions. The feedback and reporting unit is used to transform the evaluation conclusions generated by the rule evaluation and risk assessment unit into interactive prompts, providing a limited number of core error correction suggestions in real-time mode, and generating a comprehensive evaluation report including sub-item scores, main error analysis and historical progress curves in after-class mode. The data and compliance unit is used to manage various types of data generated by the system. By default, it is configured to store only anonymized key features, scoring data, and error labels, and to automatically destroy the original video data by not retaining it or retaining it for a short period of time according to the preset security policy.

[0008] Preferably, the video acquisition and shooting guidance unit includes a real-time ambient light monitoring subunit and a human presence detection subunit. The real-time ambient light monitoring subunit analyzes the image histogram distribution and issues supplementary lighting suggestions when the average brightness distribution deviates from a preset range. The human presence detection subunit calculates the logical distance from the top of the head and the bottom of the feet to the edge of the viewfinder and issues a distance adjustment prompt when the logical distance is less than a preset safety boundary value.

[0009] Furthermore, the pose estimation unit uses a deep convolutional network structure to perform multi-level feature extraction. By modeling the spatiotemporal correlation between adjacent image frames, it calculates the motion vectors of key points on the continuous time axis. When the confidence parameter of a key point is lower than a preset reliability threshold, the pose estimation unit will automatically mark the frame as a low-quality frame and notify subsequent units to adopt a conservative evaluation strategy.

[0010] Furthermore, when performing dynamic denoising and smoothing processing, the key point preprocessing unit uses a smoothing filter based on prediction logic to perform weighted correction on the key point coordinates of the current frame according to the motion speed and acceleration information of the previous frame, so as to eliminate key point jumps caused by occlusion or sudden changes in light and shadow; when performing scale normalization, the length of the line connecting the center point of both shoulders and the center point of the hip is used as the reference scale unit, and all limb lengths are converted into ratios to the reference scale unit.

[0011] Furthermore, the motion segmentation and rhythm analysis unit monitors the derivative of the motion trajectory of key joints, identifies the moment when the motion speed approaches zero as the stage switching point of the motion, and records the duration of each motion in a preset holding posture; the unit also generates a jitter coefficient characterizing the stability of the motion by calculating the sum of displacement deviations of key points at the limb ends within a small time window.

[0012] Furthermore, the traditional exercise movement knowledge base and indicator system unit pre-sets multiple dedicated evaluation dimensions for traditional sports exercises such as Baduanjin, including upright posture, shoulder and elbow angle, and smoothness of center of gravity transfer. For each movement, the knowledge base contains no less than a pre-set number of interpretable evaluation rules, and each rule is associated with the corresponding anatomical logic and teaching points.

[0013] Furthermore, when calculating the degree of limb alignment, the rule evaluation and risk judgment unit determines whether the action is in place by calculating the deviation of the angle between the distal joint, proximal joint, and associated support point from the preset standard angle; when calculating the center of gravity transfer trajectory, it estimates the real-time distribution of the center of gravity by using the positional relationship of the key points of the two feet and the vertical projection point of the torso's central axis, and triggers a balance risk warning when the center of gravity exceeds the preset support surface range.

[0014] Furthermore, when the age-appropriate stratification and personalized threshold unit detects that the user belongs to a preset elderly group or has a specific joint injury record, it will automatically widen the tolerance range of angle deviation and change the error correction logic from pursuing absolute standard to pursuing safe range of motion, avoiding giving strong error correction instructions that may induce joint strain, and replacing rigid error prompts with soft suggestions.

[0015] Furthermore, during real-time interaction, the feedback and reporting unit is configured to output only the highest priority error correction information within a preset time interval, with no more than a preset number of messages, in order to reduce the cognitive burden on users. The progress curve in the post-class report provides users with a visual prediction of practice trends by comparing the current practice data with the user's historical best data and the average data of the same group.

[0016] Furthermore, the data and compliance unit protects the coordinates of key points stored using asymmetric encryption technology and provides users with the right to withdraw their data at any time, allowing them to delete evaluation records stored in the cloud or locally. For classroom teaching scenarios, the data and compliance unit is also equipped with a data aggregation function, which supports the statistical analysis of common errors made by all students in a class and the generation of a heat map of common problems for teaching administrators to refer to.

[0017] Furthermore, the system may optionally include a template matching unit configured to load the teacher's standard demonstration sequence of movements, non-linearly align the learner's movement sequence with the standard sequence using a dynamic time warping algorithm, accurately locate the time segment with the largest deviation in the learner's movements, and output a diagram illustrating the difference in key point positions under that segment.

[0018] Furthermore, the system is configured with anomaly triage logic. When the confidence parameters of multiple consecutive frames are all below a preset threshold, or when a sudden and severe shift in the trainee's center of gravity is detected, accompanied by complete instability of the body posture, the system will immediately stop motion evaluation and trigger a safety abort mechanism, pushing rest suggestions or anomaly confirmation prompts to the terminal interface.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. The traditional sports exercise movement acquisition and evaluation system based on visual recognition provided by this invention can accurately capture the key points of relaxation, opening and closing, center of gravity shift and rhythm pause in traditional exercises by constructing a movement knowledge base and indicator system specifically for traditional exercises. This solves the technical defects of mismatched indicators and incorrect feedback in general fitness systems, and improves the relevance and professionalism of digital teaching of traditional sports exercises.

[0020] 2. By introducing age-appropriate stratification and personalized threshold mechanisms, this invention can dynamically adjust the assessment standards according to practitioners of different ages and physical conditions, avoiding the risk of sports injuries to middle-aged and elderly people that may be caused by pursuing excessive standards, and enhancing the system's audience adaptability and safety of use.

[0021] 3. The interpretable rule evaluation engine adopted in this invention, combined with template matching technology, changes the black box mode of traditional artificial intelligence evaluation. It can provide users with clear improvement suggestions based on anatomical logic, making the evaluation results more convincing and actionable. At the same time, through confidence detection and anomaly diversion mechanism, it greatly improves the robustness of the system in complex home or classroom environments and reduces the error correction caused by environmental interference such as occlusion and lighting.

[0022] 4. The closed-loop function designed for teaching management scenarios in this invention achieves full-process coverage from individual training to class management, and from real-time assistance to post-class assessment. Through common problem statistics and progress curve analysis, it improves the level of digital management in physical education teaching. In terms of data privacy protection, through key point desensitization storage and minimum retention strategies, it reduces privacy compliance pressure and storage costs while meeting assessment needs, laying a solid technical foundation for the large-scale promotion of the system. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the age-appropriate stratification and personalized threshold dynamic adjustment in this invention; Figure 3 This is a flowchart illustrating the logical flow of action segmentation recognition and motion rhythm analysis in this invention. Figure 4This is a logical diagram illustrating the rule evaluation and exercise risk determination based on the exercise knowledge base in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of video acquisition environment monitoring and multi-mode evaluation feedback in this invention. Detailed Implementation

[0024] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0025] A traditional sports exercise movement acquisition and evaluation system based on visual recognition includes a video acquisition and shooting guidance module, a posture estimation module, a key point preprocessing module, a movement segmentation and rhythm analysis module, a traditional exercise movement knowledge base and indicator system module, a rule evaluation and risk assessment module, an age-appropriate stratification and personalized threshold module, a feedback and reporting module, and a data and compliance module.

[0026] The video capture and shooting guidance module, serving as the system's data input, is deployed on a terminal device with image capture capabilities. This terminal device includes, but is not limited to, smartphones, tablets, smart TVs, or dedicated interactive teaching terminals. The video capture and shooting guidance module is configured to capture real-time video streams or record short clips of the trainee using a high-resolution complementary metal-oxide-semiconductor sensor. The frame rate is set to no less than 30 frames per second to ensure continuous motion capture. The video capture and shooting guidance module integrates an environmental perception submodule for real-time detection of the shooting environment before formal capture. Specifically, this submodule analyzes the average brightness and contrast of the current image to determine if the light intensity is below a preset brightness threshold. Simultaneously, it uses human bounding box detection logic to calculate the proportion of the human body in the frame. If this proportion is below a preset spatial proportion threshold, it means the trainee is too far from the camera, and the system automatically generates a "Please move closer to the camera" guidance message. The video acquisition and shooting guidance module also has an edge overflow detection function, which monitors in real time whether the top of the human head, the soles of the feet, and the ends of both sides of the limbs are within the safe range of the viewfinder. If the key parts are detected to be out of range, a camera position adjustment command is generated to guide the user to adjust the shooting angle, providing standardized raw materials for subsequent high-precision image processing.

[0027] The pose estimation module is connected to the video acquisition and shooting guidance module via an internal data bus, and is responsible for converting unstructured pixel-level video frames into structured skeleton features. The pose estimation module uses a deep learning model as its core engine. This deep learning model is preferably a visual transformer network based on a self-attention mechanism or a multi-layer residual convolutional neural network, which is pre-trained on a human motion dataset containing millions of samples. It can robustly identify key points of the human skeleton, including but not limited to the tip of the nose, both eyes, both ears, both shoulders, both elbows, both wrists, both hips, both knees, and both ankles. The pose estimation module is configured to perform forward inference on each frame of the image, outputting the two-dimensional or three-dimensional spatial coordinates of each key point, and simultaneously generating a confidence parameter representing the reliability of the detection for each key point. This confidence parameter is a value between 0 and 1, reflecting the model's confidence in determining the location of that point. When practitioners encounter limb obstruction or large body rotations while practicing exercises such as Baduanjin, the posture estimation module can combine the spatiotemporal correlation of adjacent frames and use a time series prediction model to estimate the coordinates of the obstructed points, thus maintaining the integrity of the skeletal structure.

[0028] The keypoint preprocessing module is connected to the output of the pose estimation module, and its main responsibility is to refine the original output keypoint sequence. The keypoint preprocessing module first performs dynamic denoising and smoothing, employing a smoothing filter based on a one-Euclidean filter or a Kalman filter algorithm. Different cutoff frequencies are set for the motion frequencies of different joints to eliminate coordinate jumps caused by image sensor noise or model jitter while preserving as much realistic motion detail as possible. For missing keypoints with a confidence level below a preset threshold, the keypoint preprocessing module performs missing keypoint completion, using linear interpolation or cubic spline interpolation techniques to complete the coordinates by referencing the point's motion trajectory in previous and subsequent frames. Furthermore, to eliminate the influence of different user body proportions, shooting distances, and terminal focal lengths on the evaluation results, the keypoint preprocessing module performs scale normalization. The specific logic of this scale normalization operation is as follows: First, identify the core skeletal length of the human torso, that is, the distance between the center points of the shoulders and the center points of the hips, and use it as the reference scale unit. Then, map the original pixel coordinates of all key points to a unified logical space coordinate system with the center of gravity of the human body as the origin and the reference scale unit as the unit scale.

[0029] The motion segmentation and rhythm analysis module receives a preprocessed coordinate sequence, aiming to segment semantically informational stages of the exercise from a continuous flow of movements. For traditional martial arts like Baduanjin, the movement logic typically follows a cycle of starting, unfolding, holding, and retracting. The module is internally equipped with a motion feature extraction engine that calculates the motion vector magnitude of key joints (such as the wrist and knee) to obtain the velocity change curve of the movement. By monitoring the minimum points of the velocity curve and the points of acceleration reversal, the module identifies moments when the velocity approaches zero and acceleration abruptly changes, defining these as stage transition points. For example, in the "Drawing the Bow to Shoot the Eagle" movement, the module marks the moment when the hands are fully extended and pause briefly as a transition from the unfolding to the holding phase. Simultaneously, the module is also configured to calculate the duration of the limbs in the holding phase and analyze the minute displacement deviations of key point coordinates within that duration, calculating the limb shaking frequency, which characterizes the stability of the movement. These temporal characteristics provide a quantitative basis for assessing whether practitioners have achieved the "combination of movement and stillness, and smooth and continuous flow" required by traditional exercises.

[0030] The traditional exercise movement knowledge base and indicator system module is the core of the entire system, storing expert-level standard movement models for different exercise categories. This traditional exercise movement knowledge base is not a simple static image library, but a multi-dimensional database composed of a series of interpretable rules based on anatomical logic. For each exercise, the knowledge base defines 5 to 8 core error-correction indicators, covering aspects such as posture alignment, shoulder and elbow angles, weight transfer trajectory, limb symmetry, and range of motion. For example, in the "Hands Supporting the Sky to Regulate the Three Burners" movement of the Eight Pieces of Brocade, the indicator system specifies in detail the range of elbow extension angles when the arms are raised, the tilt angle of the palms facing upwards, and the tolerance for deviation from the vertical axis of the torso. Furthermore, the traditional exercise movement knowledge base and indicator system module can optionally load standard demonstration templates recorded by professional instructors. These templates contain complete spatiotemporal evolution characteristics of key points and can serve as the highest reference standard for system evaluation.

[0031] The rule evaluation and risk assessment module, acting as the logical reasoning layer, is responsible for comparing the practitioner's real-time characteristics with indicators in the knowledge base across multiple dimensions. This module uses a vector angle algorithm to calculate the geometric relationship between adjacent skeletal segments, determining the degree of limb alignment. Specifically, the system calculates the angle formed by distal joints (e.g., wrist), proximal joints (e.g., elbow), and associated support points (e.g., shoulder), and performs a difference calculation with the standard angle in the indicator system. When calculating the center of gravity transfer trajectory, this module constructs a support polygon using the real-time positional relationship of key points on both feet and estimates the projection point of the body's center of gravity on the horizontal plane based on the mass distribution model of various parts of the torso. If the projection point trajectory shows that the center of gravity movement is not smooth or deviates too far from the center of the support surface, the system determines it as a risk of instability. Furthermore, the rule evaluation and risk assessment module also has biomechanical conflict detection logic, capable of identifying abnormal postures that violate the normal range of human movement and generating risk warning labels to prevent potential sports injuries.

[0032] The age-appropriate stratification and personalized threshold module is a specially optimized design for the core target audience of this invention. This module stores physiological norm data for practitioners of different age groups and is configured to dynamically adjust the logical judgment thresholds in the rule assessment and risk determination module based on the user's age range, self-reported discomfort areas, and current training stage (e.g., beginner, intermediate, or rehabilitation maintenance). For practitioners over 65 years old, the module automatically implements a "soft evaluation strategy," widening the tolerance range of movement range, for example, increasing the standard knee angle for squatting from the preset 90 degrees to 120 degrees. If the user selects "shoulder pain" during the initialization phase, the module instructs the assessment engine to lower the requirements for upper limb lifting height, shifting the error correction logic from pursuing absolute "standardization" to pursuing "effectiveness" within a safe range of activity. This differentiated assessment logic adaptation effectively alleviates the frustration experienced by middle-aged and elderly users who cannot meet the standards of younger users, and also technically reduces the possibility of joint strain caused by forcibly correcting movements.

[0033] The feedback and reporting module is responsible for transforming the underlying calculation results into understandable interactive information. This module supports two main feedback modes: real-time interaction mode and post-lesson analysis mode. In real-time interaction mode, to avoid excessive text or voice commands interfering with the learner's breathing and focus, the feedback and reporting module is configured to output no more than two highest-priority error correction suggestions, such as prioritizing prompts like "shift your weight to the left" rather than correcting minor fingertip angles. In post-lesson analysis mode, the module calls the report generation engine to summarize the movement scores throughout the practice process and analyze the achievement of various indicators. The report includes a radar chart of sub-scores, analysis of the most typical errors during practice, and a progress curve based on historical data. The progress curve compares current practice data with the user's best data over the past month, showing a trend prediction in movement stability or rhythm control, enhancing the learner's sense of engagement and accomplishment.

[0034] The data and compliance module is responsible for data security management throughout the system's entire lifecycle. Considering the sensitivity of human skeleton data and video data, this module is configured by default to store only anonymized keypoint feature vectors, scoring data, and misclassification labels. The keypoint feature vectors undergo asymmetric encryption, making it impossible to reconstruct them from actual video images, thus protecting user privacy. For the original video data, this module implements a minimal retention strategy, providing "evaluate and destroy" or "short-term retention" options. Users can independently set the retention period for videos on the server (e.g., 48 hours) and retain the right to withdraw and completely delete cloud records at any time. Furthermore, in scenarios such as school teaching or elderly care institutions, this module also provides a data aggregation interface, supporting the statistical analysis of anonymized evaluation results for all members within a class or group, generating a common error heatmap to assist teaching administrators in accurately identifying weaknesses in teaching.

[0035] Example 2: Based on the traditional sports exercise movement acquisition and evaluation system based on visual recognition described in Example 1, this example provides an alternative implementation method using an edge-cloud collaborative distributed architecture, which aims to further improve the response speed and computing efficiency in large-scale concurrent scenarios.

[0036] In this architecture, the video acquisition and shooting guidance module, the posture estimation module, and the key point preprocessing module are deployed on the practitioner's local terminal (edge ​​side), while the movement segmentation and rhythm analysis module, the traditional exercise movement knowledge base and indicator system module, the rule evaluation and risk judgment module, and the data and compliance module are deployed on the cloud server (center side).

[0037] Specifically, the pose estimation module of the local terminal utilizes a lightweight, compressed deep convolutional network for real-time inference. To adapt to the limited computing resources at the edge, this pose estimation module employs knowledge distillation technology to extract knowledge from a high-precision, large-scale teacher model, training a lightweight student model with only 1 / 10 the number of parameters. This lightweight model can run on the mobile terminal's graphics processing unit with low power consumption, outputting the coordinate sequence of human key points in real time.

[0038] The keypoint preprocessing module performs preliminary denoising and normalization of coordinates locally, then compresses the skeleton features into a data packet of a specific format and sends it to the cloud via an encrypted transmission protocol. The cloud server cluster utilizes its powerful computing capabilities to simultaneously process skeleton streams sent from tens of thousands of terminals. The motion segmentation and rhythm analysis module deployed in the cloud employs a Long Short-Term Memory network or a spatiotemporal graph convolutional network to perform deep temporal modeling of the received coordinate sequences. By analyzing the derivatives of the motion trajectories of key points in three-dimensional space, this keypoint preprocessing module can identify minute momentum changes in traditional exercises and accurately capture the transient characteristics of "accumulating energy, releasing power, and concluding the exercise" in the Eight Pieces of Brocade.

[0039] In the cloud, the traditional exercise movement knowledge base and indicator system module is constructed as a high-performance graph database. The standard movement of each exercise is represented as a graph structure containing spatial topological relationships and temporal evolution patterns. Node attributes include the ideal angles and speeds of the joints, while edge attributes represent the constraints between limbs. During evaluation, the rule-based assessment and risk determination module uses a graph matching algorithm to calculate the similarity between the practitioner's current movement diagram and the standard movement diagram, deriving a comprehensive evaluation score. This graph-based evaluation method is more robust than simple angle comparison and can better handle perspective distortion problems caused by differences in shooting angles.

[0040] The age-appropriate stratification and personalized threshold module maintains a user profile database in the cloud. Whenever a user begins practice, this module retrieves their historical health records and exercise habits based on their unique identifier. If it detects that a practitioner has repeatedly triggered knee joint pressure warnings during recent practice, the module updates their personalized threshold parameters in real time and sends the updated "protective assessment logic" to the rules engine. This centralized profile management approach allows assessment standards to dynamically evolve with the practitioner's physical condition, achieving truly personalized and precise instruction.

[0041] The feedback and reporting module, within this distributed architecture, utilizes streaming push technology to deliver evaluation conclusions to local terminals in real time. Based on the delivered logical tags, the local terminals invoke pre-built local voice libraries or graphic materials for interactive output, reducing the impact of network fluctuations on user experience. Furthermore, a commonality analysis submodule is additionally configured in the cloud to summarize the overall movement achievement rate of students in a class in real time during large-scale teaching scenarios (such as school physical education classes). This commonality analysis submodule can automatically identify common movement deviations present in more than 50% of students in the class (such as a generally high center of gravity) and immediately push teaching strategy suggestions to the teacher's end.

[0042] The data and compliance module, within a distributed architecture, undertakes more stringent gateway verification tasks. It employs differential privacy technology to perturb the uploaded skeleton data, ensuring that even if the cloud database is attacked, attackers cannot reconstruct the user's true physical details from the skeleton features. All evaluation records are timestamped and have obfuscated geolocation tags, meeting the needs of traditional sports culture dissemination statistics while strictly adhering to relevant data security regulations.

[0043] Example 3: Based on the above examples, this example describes an enhanced system implementation method, the core of which is the introduction of a template matching module to achieve deeper action alignment and operability feedback.

[0044] The template matching module is configured as an advanced assessment plugin, working in parallel with the traditional exercise movement knowledge base and indicator system module, as well as the rule assessment and risk determination module. The template matching module integrates a dynamic time warping algorithm, which can handle non-linear scaling of two movement sequences over time. After a practitioner completes a set of movements, the template matching module automatically loads the professional teacher's standard demonstration sequence stored in the knowledge base.

[0045] The alignment logic of the template matching module is as follows: First, it centers the trainee's keypoint sequence and the standard demonstration sequence in the spatial dimension; then, it uses a dynamic time warping algorithm to find the optimal matching path on the time axis. Since the speed of different trainees' movements often differs from the standard demonstration, this non-linear alignment technique can find the mapping relationship between each instant in the trainee's movement and the corresponding moment in the standard movement.

[0046] After alignment, the template matching module performs difference feature extraction. It accurately pinpoints which time segment (e.g., from the 15th to the 18th second) within the entire exercise routine deviates most from the standard movement. This deviation includes not only Euclidean distance in spatial coordinates but also deviations in movement speed and limb angles. The feedback and reporting module utilizes this comparative data to generate a "maximum deviation analysis chart" in the post-lesson report. This chart overlays the skeletal outlines of the practitioner and instructor on the screen, highlighting key points of maximum positional deviation with striking colors, providing practitioners with intuitive improvement targets.

[0047] Furthermore, the system in this embodiment is also configured with anomaly routing logic. This anomaly routing logic, as a special branch of the rule evaluation and risk determination module, is configured as a safety circuit breaker mechanism. When the keypoint confidence parameter output by the attitude estimation module is lower than a preset threshold for more than a preset number of consecutive frames (e.g., 15 consecutive frames), that is, when the system determines that the current lighting is too dim, the occlusion is too severe, or the background interference is too strong, making it impossible to provide a reliable evaluation, the anomaly routing logic will immediately trigger an "evaluation pause" command. At this time, the system will not output any error correction information, but will instead prompt the user through the feedback and reporting module that "the detection environment is poor, the evaluation has been paused, please adjust the environment."

[0048] Furthermore, this anomaly triage logic is also configured to identify severe instability. If the system detects a significant and illogical shift in the practitioner's center of gravity projection point within a very short time (e.g., within 0.2 seconds), or a severe deflection of the body's main axis exceeding 45 degrees accompanied by a sharp drop in key point confidence, the system will determine that the practitioner may have fallen or experienced severe shaking. At this point, the system will immediately trigger a safety halt mechanism, stopping all assessment logic and displaying a large message on the terminal interface asking, "Are you feeling unwell?", along with rest suggestions or an anomaly confirmation prompt based on preset logic. This feature offers significant practical value and humanistic care for middle-aged and elderly users practicing alone at home.

[0049] In this embodiment, the traditional exercise movement knowledge base and indicator system module further expands its multimodal assessment capabilities. In addition to basic visual recognition, this module is configured to access third-party biosensor data, such as a smart heart rate monitor or pulse oximeter connected via Bluetooth. While assessing movements, the system can simultaneously monitor the practitioner's heart rate variability or real-time heart rate. The rule assessment and risk judgment module correlates this physiological data with movement intensity. If it detects that the practitioner's heart rate is too high while performing a certain exercise, the system automatically shifts the assessment weight from the standardization of the movement to the gentleness of the movement, and suggests that the practitioner reduce the range of motion, thus achieving a comprehensive digital monitoring attempt of the "form, spirit, intention, and qi" in traditional sports exercises.

[0050] Regarding data compliance, this embodiment introduces blockchain evidence storage technology. The digital fingerprint (hash value) of each generated after-class assessment report is recorded on a school-level or regional-level private blockchain. This provides tamper-proof evidence for evaluating students' physical education homework, rehabilitation training records for middle-aged and elderly people, and scoring for various exercise competitions. Through this method, the data and compliance module establishes a trustworthy digital teaching quality evaluation system while protecting user privacy.

[0051] To further enhance the system's professionalism, the rule evaluation and risk assessment module also introduces "skill assessment indicators" for specific exercises in Baduanjin. These indicators are based on a comprehensive calculation of the stability of the limbs at their peak positions and the smoothness of their speed during the recovery process. For example, in the "shaking head and wagging tail to clear heart fire" movement, the system focuses on analyzing the smoothness of the movement trajectory of the key point at the Dazhui acupoint during torso rotation. By calculating the rate of change of curvature of the trajectory curve, it assesses whether the practitioner has truly achieved the "smooth and fluid" movement. These in-depth biomechanical indicators are converted into intuitive percentage scores and displayed through the feedback and reporting module, making the originally abstract key points of the exercises quantifiable and traceable.

[0052] Finally, this system supports highly modular expansion. Developers can import new indicator plugins into the traditional exercise movement knowledge base and indicator system module through standardized interfaces, based on different exercise categories (such as Wuqinxi, Tai Chi, etc.), without modifying the system's core visual engine and preprocessing logic. This "plug-and-play" architecture design gives the system strong versatility and enables it to quickly adapt to the ever-evolving application scenarios of national fitness.

[0053] In summary, this invention constructs a closed-loop system through the collaborative operation of multiple modules, encompassing environmental perception, high-precision identification, intelligent preprocessing, professional assessment, and safety feedback. It solves the superficial alignment problem of traditional digital assessment of exercise methods, which often relies on "reading pictures and recognizing characters." Furthermore, through age-appropriate layering, an interpretable rule engine, and an anomaly triage mechanism, it provides a professional, safe, refined, and user-friendly digital sports training solution.

[0054] Example 4: Based on the system architecture described above, this example focuses on illustrating a system deployment model for classroom teaching management scenarios. In this model, the system is integrated into a centralized teaching management platform, supporting multi-terminal access and centralized data analysis.

[0055] In a classroom environment, multiple students simultaneously practice exercises within the camera's coverage area. The video capture and shooting guidance module is configured to support multi-target tracking logic, capable of using human detection algorithms to locate multiple independent practitioners in a single frame and assigning a unique tracking identifier to each individual. The pose estimation module employs a multi-task parallel processing strategy, simultaneously extracting skeletal keypoint sequences from multiple targets.

[0056] To address the computational load balancing issue in multi-target scenarios, the keypoint preprocessing module introduces a dynamic importance scoring mechanism. This mechanism dynamically allocates computational resources based on the learner's distance from the camera, the degree of occlusion, and their role in the current teaching task (e.g., a student being randomly selected for evaluation). For students located at the core of the frame or highlighted by the teacher, the system uses a more precise deep learning model for pose estimation; while for targets at the edges, a faster tracking algorithm with lower computational overhead is employed.

[0057] The traditional exercise movement knowledge base and indicator system module provides a set of "common assessment benchmarks" in the teaching mode. These benchmarks are set in real-time by the instructor based on the teaching focus of the lesson. For example, if the lesson emphasizes "sinking the shoulders and dropping the elbows," the instructor can increase the weight of shoulder and elbow-related indicators with a single click through the management interface. The rule assessment and risk judgment module will adjust the contribution of each indicator to the total score in real-time according to the instructor's settings, ensuring that the assessment feedback is highly consistent with the classroom teaching objectives.

[0058] The feedback and reporting module features a "layered display" function in teaching mode. For students, the system uses augmented reality (AR) technology to directly overlay error correction information onto their real-time mirror image, such as using red lines to mark tilted torsos or arrows to indicate the direction of weight shift. For teachers, the module provides a global monitoring dashboard that displays each student's achievement status in real time. Through common problem statistics, teachers can clearly see which movements are the weak points for the whole class (e.g., 80% of students have insufficient torso rotation angle when performing the "Five Labors and Seven Injuries Looking Back" exercise), allowing for timely adjustments to the teaching content.

[0059] In this scenario, the data and compliance module serves as the archiver for teaching evaluations. It supports summarizing the practice data of all students in the class into a teaching quality heatmap and generating phased progress curves. This data not only serves as a component of students' regular grades but also provides quantitative evidence for evaluating the effectiveness of physical education reforms in schools. To meet the privacy protection requirements of a campus environment, all data stored on the management platform is anonymized. Unless specifically authorized, the system cannot associate skeleton data with students' real names and faces.

[0060] Furthermore, the system in this embodiment also incorporates a social incentive mechanism. The feedback and reporting module generates a "Daily Master Ranking" or "Fastest Improvement Ranking" based on evaluation scores and displays it within a controlled class circle. This incentive mechanism can enhance students' enthusiasm for practicing traditional sports exercises. Simultaneously, the age-appropriate stratification and personalized threshold module adjusts the threshold to an encouraging "challenge mode" for adolescent students, guiding them to pursue higher movement quality by appropriately raising the standards.

[0061] Through this in-depth optimization for teaching management scenarios, the present invention is not only a personal training tool, but has evolved into a highly efficient teaching support system. It fills the gap in traditional physical education classes by using visual recognition technology, addressing the inability to monitor the details of each student's movements in real time. This improves the teaching efficiency of physical education teachers and enables the large-scale, standardized teaching of traditional physical exercises.

[0062] In summary, the system described in this embodiment demonstrates the broad application prospects of this invention in the field of professional education through multi-target recognition, teaching benchmark setting, AR real-time feedback, and class-based data analysis.

[0063] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. The functional modules, sub-units, and hardware support in the system can be flexibly combined and adjusted according to the computing resources, sensor specifications, and network conditions of the actual application environment. Without departing from the core logic of the present invention, parameter refinement, logic branch addition, or sensor multimodal expansion for specific traditional exercises are all extensions of the technical solution of the present invention. The present invention is not only applicable to Baduanjin, but can also be extended to various traditional sports with structured movement patterns, such as Yijinjing, Wuqinxi, and Taijiquan, as well as modern rehabilitation training scenarios.

Claims

1. A system for collecting and evaluating traditional sports movements based on visual recognition, characterized in that, include: The video acquisition and shooting guidance unit is used to acquire real-time video streams or record short clips, and is configured to detect the shooting environment in real time. When the environmental parameters are detected to be lower than the preset threshold, the unit automatically generates corresponding environmental adjustment guidance information and camera position adjustment instructions. The pose estimation unit is configured to parse the image frames acquired by the video acquisition and shooting guidance unit, extract the sequence of key points of the human skeleton, and generate a confidence parameter representing the detection reliability for each key point. The key point preprocessing unit is used to receive the key point sequence output by the pose estimation unit and perform dynamic denoising and smoothing, missing key point completion and scale normalization operations. The movement segmentation and rhythm analysis unit is used to identify the movement stages in traditional sports exercises based on the coordinate sequence provided by the key point preprocessing unit, and to calculate the movement characteristic data of each stage to reflect the continuity and stability of the practitioner's movements. The traditional exercise movement knowledge base and indicator system unit is configured to store expert-level standard movement characteristics and core error correction indicators for different traditional sports exercises. The rule evaluation and risk assessment unit is used to combine the traditional exercise movement knowledge base and indicator system unit to perform multi-dimensional calculations on the features output by the movement segmentation and rhythm analysis unit, evaluate the standardization of the movement, and generate risk warning labels for postures that do not conform to physiological and mechanical characteristics. An age-appropriate stratification and personalized threshold unit is configured to dynamically adjust the logical judgment threshold in the rule evaluation and risk judgment unit based on the user's physiological background information and training stage. The feedback and reporting unit is used to transform evaluation conclusions into interactive prompts and generate evaluation reports that include sub-scores, error analysis, and historical progress curves in different modes. The data and compliance unit is used to manage various types of data generated by the system and to perform encrypted storage or automatic destruction of the data according to preset security policies.

2. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, The video acquisition and shooting guidance unit includes a real-time ambient light monitoring subunit and a human body entering the camera subunit; The real-time ambient light monitoring subunit analyzes the image histogram distribution and issues supplementary lighting suggestions when the average brightness distribution deviates from a preset range. The human body detection subunit identifies the spatial proportion of the trainee in the frame using a human body detection algorithm. When key parts of the human body exceed the frame or the spatial proportion is lower than a preset spatial proportion threshold, a corresponding camera position adjustment instruction is generated. The human body detection subunit calculates the logical distance from the top of the head and the bottom of the feet to the edge of the viewfinder, and issues a distance adjustment prompt when the logical distance is less than a preset safety boundary value.

3. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, The pose estimation unit uses a deep convolutional network structure to perform multi-level feature extraction. By modeling the spatiotemporal correlation between adjacent image frames, it calculates the motion vectors of key points on continuous time axes. When the confidence parameter of a certain key point is lower than the preset reliability threshold, the pose estimation unit marks the adjacent image frame as a low-quality frame and notifies the subsequent units to adopt a conservative evaluation strategy. When performing dynamic denoising and smoothing processing, the key point preprocessing unit uses a smoothing filter based on prediction logic to perform weighted correction on the key point coordinates of the current frame according to the motion speed and acceleration information of the previous frame, so as to eliminate key point jumps caused by occlusion or sudden changes in light and shadow. When performing scale normalization, the key point preprocessing unit identifies the core skeleton length of the human torso and uses it as the reference scale unit to map the original pixel coordinates of all key points to a unified logical space coordinate system with the center of gravity of the human body as the origin and the reference scale unit as the unit scale.

4. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, The motion segmentation and rhythm analysis unit monitors the motion trajectory derivatives of key joints and identifies the moment when the motion speed approaches zero and the acceleration changes direction as the phase switching point of the motion, dividing the continuous motion into the starting phase, the unfolding phase, the holding phase, and the retraction phase. The motion segmentation and rhythm analysis unit is also configured to record the duration of each motion in a preset holding posture, and generate a limb shaking coefficient that characterizes the stability of the motion by calculating the total displacement deviation of key points at the end of the limb within a small time window. The motion characteristic data includes the motion speed, dwell time, and limb tremor coefficient at each stage.

5. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, The indicator system in the traditional exercise movement knowledge base and indicator system unit covers the upright posture, the angle of shoulder sinking and elbow dropping, the trajectory of center of gravity transfer, limb symmetry, and the range of motion of the movements; For each movement, the knowledge base defines multiple interpretability rules based on anatomical logic, and each rule is associated with corresponding biomechanical indicators and teaching points. The traditional exercise movement knowledge base and indicator system unit is also configured to load standard demonstration templates recorded by professional teachers. These standard demonstration templates contain complete spatiotemporal evolution characteristics of key points and serve as the highest reference criteria for system evaluation.

6. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, When calculating the degree of limb alignment, the rule evaluation and risk judgment unit uses a vector angle algorithm to calculate the angle between the distal joint, proximal joint and associated support point, and obtains the deviation value between it and the preset standard angle to determine whether the action is in place. When calculating the center of gravity transfer trajectory, the rule evaluation and risk judgment unit uses the positional relationship of the key points of the two feet to construct a support polygon, and combines the vertical projection points of the torso central axis to estimate the real-time distribution of the human body's center of gravity. When the vertical projection point exceeds the preset ratio range of the supporting polygon constructed by the positional relationship, the system triggers a balance risk warning; In addition, the rule assessment and risk determination unit also has a biomechanical conflict detection logic, which is used to identify abnormal postures that violate the physiological limits of normal human movement and to associate them with risk warnings.

7. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, The age-appropriate stratification and personalized threshold unit stores physiological norm data for trainees of different age groups, and dynamically executes differentiated assessment logic to adapt based on the user's age range, self-reported discomfort area information, and current training stage. For users in the preset advanced age group or with a history of joint injury, the age-appropriate stratification and personalized threshold unit implements a soft evaluation strategy by widening the tolerance range of angle deviation. This changes the error correction logic from pursuing preset standards to pursuing the effectiveness of actions within a safe range of activity, so as to avoid joint strain caused by forcibly correcting actions. It also replaces rigid error prompts with preset suggestions.

8. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, In real-time interactive mode, the feedback and reporting unit is configured to output only the highest priority error correction information within a preset time interval, with no more than a preset number of messages. In the post-lesson analysis mode, the feedback and reporting unit summarizes the action scores of the entire practice process and generates a comprehensive evaluation report that includes a radar chart of sub-item scores, analysis of typical error points, and prediction of practice trends. The system is also configured with an abnormal flow diversion logic. When the confidence parameter of the key point output by the posture estimation unit is lower than the preset threshold for more than a preset number of consecutive frames, or when the system detects that the trainee's center of gravity projection point has a violent shift that does not conform to physical logic within a preset time interval and is accompanied by complete instability of body posture, the system immediately stops the motion evaluation and triggers the safety abort mechanism, pushing rest suggestions or abnormal situation confirmation prompts to the terminal interface.

9. The traditional sports exercise movement acquisition and evaluation system based on visual recognition according to claim 1, characterized in that, The system adopts a distributed architecture with edge-cloud collaboration. The video acquisition and shooting guidance unit, posture estimation unit, and key point preprocessing unit are deployed on the local terminal, while the action segmentation and rhythm analysis unit, traditional exercise action knowledge base and indicator system unit, rule evaluation and risk judgment unit, and data and compliance unit are deployed on the cloud server. The pose estimation unit of the local terminal uses a lightweight compressed deep learning model to perform real-time inference and sends the processed skeleton feature data packet to the cloud server through an encrypted transmission protocol. The cloud server uses a graph database to construct a standard movement model, where node attributes include the ideal angle and speed of the joints, and edge attributes represent the constraint relationships between limbs. The similarity between the exerciser's current movement diagram and the standard movement diagram is calculated using a graph matching algorithm.

10. A method for collecting and evaluating traditional sports movements based on visual recognition, characterized in that: The traditional sports exercise movement acquisition and evaluation system based on visual recognition, as described in any one of claims 1-9, is used to acquire and evaluate traditional sports exercise movements.

Images