A Student Physical Education Movement Correction System Based on Image Recognition

Abstract

This invention belongs to the field of computer image recognition technology, specifically disclosing a student physical education movement correction system based on image recognition. The system includes an individual physiological parameter modeling module, a standard movement dynamic adaptation module, a multimodal data acquisition and processing module, a group relative assessment and incentive module, and a personalized correction instruction generation and feedback module. By establishing a personalized physiological model and dynamically adapting it to standard movements, a unique movement reference is generated for each student. Simultaneously, relative assessment and incentives are provided by combining group data. Finally, specific and operable real-time correction instructions are generated through hierarchical difference analysis, thereby achieving scientific, personalized, and interactive physical education movement training guidance. By introducing individual physiological parameter modeling and standard movement dynamic adaptation, this system fundamentally changes the benchmark for movement assessment, transforming it from a unified static model into a dynamic reference that varies from person to person.

Description

Technical Field

[0001] This invention belongs to the field of computer image recognition technology, and specifically relates to a student physical exercise correction system based on image recognition. Background Technology

[0002] The rapid development of artificial intelligence and computer vision technologies has brought profound changes to the field of sports teaching and training. Capturing human postures through image or video data and using deep learning models for motion analysis and evaluation has become an important technical means to improve teaching efficiency and the scientific nature of training.

[0003] Image recognition-based sports movement correction technology aims to assist students or athletes in standardizing their technical movements, reducing the risk of sports injuries, and improving training effectiveness through automated and intelligent methods.

[0004] Existing technologies typically rely on comparing captured images of student movements with pre-set standard movement models, generating corrective suggestions by calculating the angle and positional deviations of key skeletal nodes. However, this approach has significant limitations: its evaluation criteria are often static and uniform, failing to adequately consider individual differences among students in terms of body shape, flexibility, and strength. This can easily lead to "fitting the mold" correction schemes, requiring students to forcibly match a "perfect" model that may not suit their own conditions, potentially increasing the risk of injury or undermining their learning confidence. Furthermore, existing systems mostly focus on isolated analysis and feedback for individual students, lacking mechanisms for relative evaluation within a group setting. They fail to provide a reference point based on peer performance, making the correction process tedious, failing to stimulate students' competitive spirit and long-term practice motivation, and challenging the sustainability of learning outcomes.

[0005] Therefore, how to construct an intelligent sports movement correction system that can take into account both individual physiological differences and group interaction incentives has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This invention provides a student physical education movement correction system based on image recognition, which solves the problems in the prior art that ignore individual differences due to the use of a uniform static standard model for movement evaluation, and the lack of a group interaction mechanism that leads to a boring correction process and insufficient learning motivation.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is a student physical education movement correction system based on image recognition. The system includes an individual physiological parameter modeling module, a standard movement dynamic adaptation module, a multimodal data acquisition and processing module, a group relative assessment and incentive module, and a personalized correction instruction generation and feedback module.

[0008] The individual physiological parameter modeling module is used to establish and maintain a dynamically updated individual physiological parameter model for each registered student. This module is implemented through the following steps: First, during system initialization, students are guided to complete a set of standardized static posture capture and dynamic basic movement tests. The multimodal data acquisition and processing module collects data on students' height, limb length, trunk proportions, joint range of motion, and basic strength performance.

[0009] Furthermore, the individual physiological parameter modeling module incorporates a parametric human model engine, mapping the collected raw data to the student's personalized bone length parameters, joint rotation limit parameters, and muscle strength level parameters. These parameters collectively constitute the core data layer of the individual physiological parameter model. This individual physiological parameter model is not static; its joint range of motion and strength level parameters are periodically fine-tuned based on subsequent training data recorded by the system.

[0010] The standard movement dynamic adaptation module is used to dynamically adapt each standard movement in the general standard movement library to the specific student model output by the individual physiological parameter modeling module in real time, generating a personalized standard movement reference for that student. The workflow of the standard movement dynamic adaptation module is as follows: The module receives a general movement description file from the standard movement library. This general movement description file defines the ideal spatial trajectory, joint angle sequence, and force application timing that each body part should achieve when performing the movement.

[0011] Furthermore, the module calls upon the individual physiological parameter model of the target student, proportionally scales the general spatial trajectory based on the skeletal length parameter in the model, performs compliance-compliant cropping and smooth transition on the general joint angle sequence based on the joint rotation limit parameter, and adaptively adjusts the intensity curve of the force exertion sequence based on the muscle strength level parameter. After the above adaptation process, a personalized standard movement reference sequence is output that matches the student's physiological conditions in terms of spatial form, angle range, and force exertion characteristics.

[0012] A multimodal data acquisition and processing module is used to collect students' movement data in real time during sports training and to perform preprocessing and feature extraction. This module includes a visual acquisition unit, an inertial measurement unit (IMU) integration interface, and a data processing core. The visual acquisition unit consists of multiple high-definition cameras deployed around the training area to simultaneously capture video streams of students' movements from different perspectives. The IMU integration interface receives acceleration, angular velocity, and attitude data sent by wearable devices worn by students. The data processing core first performs time synchronization and spatial calibration on the multiple video streams, then runs a lightweight human pose estimation algorithm to extract the students' two-dimensional or three-dimensional skeletal keypoint sequences from the videos in real time. Simultaneously, the data processing core filters and fuses the inertial measurement data to compensate for information loss in visual data during rapid rotation or occlusion. Finally, the multimodal data acquisition and processing module outputs a time-aligned multimodal motion data stream that fuses the visual skeletal point sequences and inertial measurement data.

[0013] The group relative assessment and incentive module is used to introduce a group reference system based on individual assessments, generating relative assessment indicators and incentive signals based on peer performance. This module includes a group data pool, a relative assessment calculator, and a dynamic incentive strategy engine. The group data pool stores and indexes the historical best performance data of all students when completing the same training project, including final scores, key action angles, and completion times. Upon receiving the current student's real-time assessment result, the relative assessment calculator performs the following operations: First, based on the current student's physiological parameter model, it selects a peer group from the group data pool whose physiological similarity to the student is higher than a preset threshold, such as 80%, as a reference group.

[0014] The system calculates the percentile ranking of each student's current performance indicators within the historical data of the reference group. The dynamic incentive strategy engine then generates incentive signals based on the percentile ranking: when the ranking enters the top 20%, a competitive challenge task is generated, such as challenging a more difficult variation of the movement; when the ranking is in the middle 60%, encouraging feedback based on small improvements and a visual progress bar are generated; when the ranking is in the bottom 20%, guiding prompts focusing on basic consolidation are generated, including specific peer comparison cases.

[0015] The personalized correction instruction generation and feedback module serves as the system's decision-making and output terminal. This module receives real-time student action data streams from the multimodal data acquisition and processing module, as well as personalized standard action reference sequences from the standard action dynamic adaptation module. The core of this module is a hierarchical action difference analyzer. This analyzer first dynamically time-aligns the real-time action sequence with the reference sequence in the time dimension.

[0016] In the spatial dimension, the analyzer decomposes motion differences into three levels: the first level is joint angle deviation, calculating the difference between the actual angle of the main joints in the keyframe and the personalized reference angle; the second level is motion trajectory deviation, calculating the similarity between the actual motion trajectory of the extremities, such as hands and feet, and the reference trajectory; the third level is force application pattern deviation, evaluating the coordination of force application timing and intensity by analyzing acceleration characteristics in inertial measurement data. For each type of deviation identified, the correction instruction generator associates with a predefined correction knowledge base containing multiple expression methods, selects the most explicit instruction text based on the severity and type of the deviation, and combines augmented reality technology or mobile terminal interface to provide real-time feedback in the form of highlighted skeletons, arrow guidance, text overlay, or voice prompts. Simultaneously, the personalized correction instruction generation and feedback module updates the evaluation results of this training, including absolute deviation values ​​and group relative rankings, to the individual physiological parameter modeling module and the group data pool, achieving iterative optimization of the model.

[0017] As one embodiment of the present invention, the periodic fine-tuning process of the joint range of motion parameters in the individual physiological parameter modeling module is as follows: The system records the maximum joint angles that students actually achieve in multiple training sessions without triggering system alarms, and these data points form a sample set; every preset period, for example, every 10 effective training sessions, the system uses the statistical value of the sample set, such as the 95th percentile, to update the corresponding joint rotation limit parameters in the model, but the updated values ​​must not exceed the anatomical safety limit.

[0018] As one embodiment of the present invention, the adaptive adjustment of the force exertion timing intensity curve by the standard movement dynamic adaptation module follows the following rules: for students whose muscle strength level parameter is marked as "beginner", 80% of the peak intensity in the general force exertion curve is used as their personalized reference peak, and the duration of force exertion is appropriately extended; for students whose level is marked as "advanced", 110% of the general curve is used as the reference peak, and the preparation period for force exertion may be compressed.

[0019] In one embodiment of the present invention, the process of selecting peer groups with similar physiological conditions in the group relative assessment and incentive module adopts multi-dimensional feature weighted Euclidean distance calculation. Specifically, height, lower limb length, and trunk length in the individual physiological parameter model are selected as key feature dimensions, and weights are assigned to each dimension based on expert experience. The distance between the current student and all students in the group data pool in these weighted feature spaces is calculated, and individuals whose distance is less than a set threshold are included in the reference group.

[0020] In one embodiment of the present invention, the correction knowledge base in the personalized correction instruction generation and feedback module is organized using a conditional triggering mechanism. Each correction instruction is associated with a conditional label defined by the action type, deviation level, and deviation value range. When the deviation data output by the hierarchical action difference analyzer matches a certain conditional label, the system randomly selects one of the two to three candidate instructions associated with that label for feedback, in order to avoid repetition and monotony of instructions.

[0021] In one embodiment of the present invention, the system operates within a three-stage cyclical framework comprising offline modeling, online assessment, and long-term tracking. The offline modeling stage completes student registration and the initial construction of individual physiological parameter models. The online assessment stage performs real-time motion capture, personalized comparison, group reference, and corrective feedback during each training session. The long-term tracking stage is responsible for periodically integrating training data from previous stages, such as a semester, reassessing the trends in students' physiological parameters and skill progress curves, and generating periodic teaching reports and personalized training plan suggestions for teachers.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention fundamentally changes the benchmark for movement assessment by introducing individual physiological parameter modeling and dynamic adaptation to standard movements, transforming it from a uniform static model to a dynamic reference tailored to each individual. This design means that corrective suggestions no longer require students to forcibly match potentially unrealistic "ideal" models, but rather guide them towards the optimal movement pattern that suits their own physiological conditions. This significantly improves the scientific rigor and safety of correction, effectively avoids the potential risk of sports injuries caused by "forcing a fit," and protects students' learning confidence.

[0023] 2. This invention creatively constructs a group relative assessment and incentive module, embedding the student's training process from isolated individual behavior into a positive peer reference system. Through peer selection and percentile ranking based on physiological similarity, the system provides students with intuitive and meaningful horizontal comparison coordinates. Combined with dynamic incentive strategies, the system can automatically generate diverse feedback ranging from competitive challenges to consolidation guidance, greatly enriching the interactivity and fun of the training process, effectively stimulating students' intrinsic competitive awareness and continuous practice motivation, thereby solving the technical problem of traditional correction systems being tedious and difficult to maintain long-term learning interest.

[0024] 3. This invention employs a hierarchical movement difference analysis mechanism, decomposing complex movement deviations into multiple quantifiable levels such as joint angles, movement trajectories, and force application patterns for refined diagnosis. This structured analysis method, combined with a diverse corrective instruction library triggered by conditions, ensures that the system-generated feedback is not only timely but also highly specific and actionable. Students can clearly understand "what's wrong" and "how to improve," thereby significantly improving the efficiency and effectiveness of corrective training. Simultaneously, the system's three-stage cyclical framework ensures a closed loop from short-term real-time feedback to long-term ability tracking, providing a complete technical solution for the personalization and scientification of physical education. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical architecture of the image recognition-based student physical activity correction system proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the standard action dynamic adaptation module in this invention; Figure 3 This is a logical flow diagram of the multimodal data acquisition and processing module in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the group relative evaluation and incentive module in this invention; Figure 5 This is a schematic diagram of the hierarchical action difference analysis principle of the personalized correction instruction generation and feedback module in this invention. Detailed Implementation

[0026] This invention provides a student physical education movement correction system based on image recognition, the overall technical architecture of which is as follows: Figure 1 As shown, the system comprises five core functional units: an individual physiological parameter modeling module, a standard movement dynamic adaptation module, a multimodal data acquisition and processing module, a group relative assessment and incentive module, and a personalized correction instruction generation and feedback module. These modules are tightly coupled through predefined data interfaces and event-driven mechanisms, forming a complete closed loop from individual modeling, dynamic adaptation, real-time perception, group reference, to precise feedback. The specific implementation methods of each module will be described in detail below with reference to the accompanying drawings.

[0027] The individual physiological parameter modeling module is the fundamental supporting unit of the entire system. Its core task is to build and continuously maintain a high-fidelity individualized physiological parameter model for each registered student. The individual physiological parameter modeling module initiates its workflow during the system initialization phase: when a new student registers for the first time, the system guides them into a dedicated calibration area where they complete a set of standardized static posture capture and dynamic basic movement tests. Static postures include fixed positions such as standing naturally, arms outstretched, and squatting to 90 degrees; dynamic basic movements cover low-intensity full-body exercises such as high knees, lunges, and shoulder circles.

[0028] During this process, the multimodal data acquisition and processing module is activated simultaneously. Multiple high-definition cameras deployed around the field capture the student's three-dimensional spatial posture from different perspectives, and receive acceleration, angular velocity, and quaternion posture data transmitted from the student's lightweight wearable device via the inertial measurement unit (IMU) integrated interface. This raw data is then fed into the individual physiological parameter modeling module, where it undergoes noise filtering and time alignment before being input into the built-in parametric human model engine. This parametric human model engine employs an improved SMPL (Skinned Multi-PersonLinear) human modeling framework, but abandons its general average template. Instead, it uses the collected actual skeletal keypoint coordinates as constraints, employing a nonlinear optimization algorithm to deduce the student's personalized skeletal length parameter set, including 12 core indicators such as upper arm length, forearm length, thigh length, lower leg length, trunk height, and shoulder width. Simultaneously, the system analyzes the maximum joint flexion and extension angles that the student can safely complete during dynamic testing, preliminarily estimating the range of motion of the hip, knee, ankle, shoulder, elbow, and wrist joints, and recording these as initial joint rotation limit parameters. In addition, based on the peak acceleration and rate of change of angular momentum in the inertial measurement data when students complete specific weight-bearing or explosive movements (such as standing long jump and push-ups), the system quantifies and classifies their muscle strength level into three levels: primary, intermediate, and advanced, and stores them in the model as muscle strength level parameters.

[0029] The three types of parameters mentioned above—bone length parameters, joint rotation limit parameters, and muscle strength level parameters—constitute the core data layer of the individual physiological parameter model. This model is not a one-time creation but rather possesses dynamic evolution capabilities. In each subsequent training session, the system continuously records joint angle data during the student's actual movements, paying particular attention to effective training segments that do not trigger safety alarms and whose movement quality scores exceed a threshold (e.g., 85 points). After accumulating 10 effective training sessions, the system automatically initiates a fine-tuning process: for each major joint, it extracts the maximum flexion and extension angles actually achieved during effective training, forming an angle set containing at least 30 effective samples; then, it calculates the 95th percentile value of this angle set and uses it as a new candidate value for the joint rotation limit parameter. If the candidate value for the joint rotation limit parameter does not exceed the preset anatomical safety upper limit (e.g., maximum knee flexion not exceeding 135 degrees), the corresponding parameter in the model is updated; otherwise, the original value is retained. Through this mechanism, the individual physiological parameter model can adaptively evolve as the student's physical fitness improves, ensuring that subsequent movement assessments are always based on actual physiological capabilities.

[0030] The function of the standard motion dynamic adaptation module is to transform the abstract motion specifications in the general standard motion library into personalized reference benchmarks that match the specific physiological conditions of students. Please refer to the appendix. Figure 2 The workflow of the standard motion dynamic adaptation module begins with receiving a motion description file from the standard motion library. This motion description file uses a structured XML format and clearly defines the spatial trajectory, joint angle sequence, and timing force characteristics that a standard motion (such as basketball shooting, standing long jump, or pull-up) should ideally satisfy. For example, for the standing long jump, the file specifies that the ideal angles of the hip, knee, and ankle joints at the moment of takeoff are 110 degrees, 90 degrees, and 100 degrees, respectively; the ideal parabolic trajectory equation of the body's center of gravity during the flight phase; and the ideal timing curve of the ground reaction force during the lower limb extension phase.

[0031] Upon receiving such a description, the standard motion dynamic adaptation module immediately invokes the individual physiological parameter model of the target student. First, based on the limb length and torso proportion parameters in the model, the general spatial trajectory is geometrically scaled. Specifically, the system multiplies all spatial coordinate points in the general trajectory by a scaling factor consisting of the ratio of the student's actual lower limb length to the standard model's lower limb length, thereby generating a unique trajectory that conforms to the student's limb dimensions. Second, for the general joint angle sequence, the module performs compliance pruning: if the target angle of a joint in the general sequence exceeds the rotational limit recorded in the student's model, the system forcibly restricts the angle to within the limit value and inserts a smooth transition function (such as cubic spline interpolation) between adjacent frames to avoid dynamic discontinuities caused by abrupt angle changes. Finally, for the force exertion timing intensity curve, the module adaptively adjusts it according to the student's muscle strength level parameters.

[0032] If a student is labeled "Beginner," the system multiplies the peak intensity in the general curve by a coefficient of 0.8 and extends the duration of the main force phase by 1.2 times. If the student is labeled "Advanced," the peak intensity is multiplied by a coefficient of 1.1, and the force preparation period is compressed to 0.9 times the original duration. After the above three-fold adaptation processing, the module outputs a complete personalized standard movement reference sequence. This personalized standard movement reference sequence is highly consistent with the student's physiological reality in three dimensions: spatial morphology, angular feasible domain, and force intensity, providing a scientific and feasible reference benchmark for subsequent real-time comparison.

[0033] The multimodal data acquisition and processing module is responsible for acquiring high-precision data streams of student movements in real-time within realistic training scenarios. Please refer to the attached document. Figure 3 The multimodal data acquisition and processing module consists of three collaborative parts: a visual acquisition unit, an inertial measurement unit (IMU) integrated interface, and a data processing core. The visual acquisition unit comprises 4 to 6 high-definition network cameras deployed around the training area. Each camera has a resolution of 1920×1080 and a frame rate of at least 30 frames per second. All cameras undergo rigorous extrinsic parameter calibration to ensure their imaging planes have a known spatial position and orientation within a unified world coordinate system. All cameras achieve millisecond-level time synchronization via hardware trigger signals, ensuring strict alignment of the multi-view video streams on the timeline. The IMU integrated interface establishes a low-latency communication link with a miniature wearable device worn by students via Bluetooth 5.0 or Wi-Fi Direct protocol. This device, typically integrated into the wrist, ankle, or belt, contains a three-axis accelerometer, a three-axis gyroscope, and a magnetometer, with a sampling frequency of 100 Hz.

[0034] The data processing core, serving as the central hub of the module, first performs background modeling and foreground segmentation on multiple video streams to remove irrelevant environmental interference. Then, it runs a lightweight yet robust 3D human pose estimation algorithm. This algorithm employs a two-stage strategy: the first stage detects 17 2D keypoints in a single frame image; the second stage utilizes multi-view geometric constraints and temporal continuity priors to elevate these 2D keypoints to 3D spatial coordinates. Simultaneously, inertial measurement data is fed into a Kalman filter for denoising and pose fusion, generating high-precision local limb orientation and angular velocity information. In cases of rapid rotation or multiple people obstructing visual keypoints, causing loss of visual data, the system automatically switches to an inertial data-driven pose estimation mode and optimally fuses visual and inertial data in the state space through complementary filtering. Finally, the multimodal data acquisition and processing module outputs a multimodal motion data stream with precisely aligned timestamps. Each frame contains 3D skeletal keypoint coordinates (in millimeters), joint angles (in degrees), limb end-effector velocity vectors (in meters per second), and a fused pose confidence score (between 0 and 1). The multimodal motion data stream is pushed to downstream modules in real time at a rate of 30 frames per second, providing raw data for motion evaluation.

[0035] The group relative assessment and motivation module aims to break through the limitations of traditional isolated assessments and introduce social reference mechanisms to enhance training motivation. Please refer to the appendix. Figure 4 The relative assessment and incentive module comprises three sub-units: a group data pool, a relative assessment calculator, and a dynamic incentive strategy engine. The group data pool is a distributed time-series database that organizes and stores all historical training records by training item, student ID, and training date. Each record contains structured indicators such as movement completion score, key joint angles, completion time, and force coordination index. When a student is performing a training exercise (e.g., a 50-meter sprint), the relative assessment calculator first obtains the student's current physiological feature vector from the individual physiological parameter modeling module. This vector consists of three core parameters: height, lower limb length, and trunk length, each assigned a weight set by an expert (e.g., 0.4, 0.4, 0.2). Subsequently, the calculator iterates through all records of peers who have completed the same exercise in the group data pool, calculating the weighted Euclidean distance between the current student and each historical student. in, These represent height, lower limb length, and torso length, respectively. For the corresponding weights, For current students, For historical students, all students whose distance is less than a preset threshold (e.g., 15 cm) are grouped into a physiologically similar reference group. Next, the calculator extracts various indicators from the current student's training (e.g., 7.2 seconds for the 50-meter dash) and calculates their percentile ranking within the historical best performance distribution of the reference group. If the ranking is in the top 20%, the dynamic incentive strategy engine activates a competitive challenge mode, pushing tasks such as "You have surpassed 80% of your similar peers! Try the 'accelerated start' training" to the student; if the ranking is in the middle 60%, encouraging feedback such as "You improved by 0.3 seconds compared to last week! Just 0.2 seconds more and you'll be in the front" is displayed, accompanied by a progress bar visualization; if the ranking is in the bottom 20%, the system generates guiding prompts, such as "Pay attention to insufficient forward shift of the center of gravity at the start; refer to Zhang's (85% physiological similarity) starting video," along with anonymized peer action clips. All incentive content is presented in real-time via mobile terminals or sideline displays, forming a positive behavior reinforcement loop.

[0036] The personalized correction instruction generation and feedback module serves as the system's decision output terminal, performing refined action deviation diagnosis and multi-channel feedback generation. Please refer to the attached document. Figure 5 The core of this personalized correction instruction generation and feedback module is a hierarchical motion difference analyzer. Its input consists of real-time student motion data streams provided by the multimodal data acquisition and processing module, and personalized standard motion reference sequences output by the standard motion dynamic adaptation module. The analyzer first performs a dynamic time warping algorithm on the two sequences in the time dimension to solve the frame misalignment problem caused by different motion rhythms, ensuring that key motion phases (such as take-off and landing) correspond accurately.

[0037] Subsequently, in the spatial dimension, the analyzer decomposes the motion differences into three levels for quantitative evaluation. The first level is joint angle deviation: the system iterates through major joints such as the hip, knee, and ankle, calculating the absolute difference between the actual angle and the reference angle in each keyframe. If the deviation of a joint exceeds a preset tolerance (e.g., ±8 degrees for the knee joint) for three consecutive frames, it is marked as a significant angle deviation. The second level is motion trajectory deviation: for extremities such as the hands and feet, the system calculates the Fréchet distance between the actual motion trajectory and the reference trajectory. If this Fréchet distance exceeds 10% of the limb length, it is determined to be a trajectory deviation. The third level is force pattern deviation: by analyzing the acceleration timing data provided by the inertial measurement unit, the system detects whether the peak force occurrence time lags behind the reference timing by more than 0.1 seconds, or whether the force intensity is lower than 70% of the reference value, thereby determining a force coordination defect. For each type of identified deviation, the correction instruction generator queries a preset correction knowledge base. The correction knowledge base is organized using a conditional triggering mechanism, where each instruction is associated with a composite condition tag defined by the action type, deviation level, and deviation value range.

[0038] For example, under the tag "Standing Long Jump - Joint Angle Deviation - Knee Insufficiency (<80 degrees)", there are three alternative instructions: "Fully flex the knee to 90 degrees before takeoff", "Imagine sitting in a chair to increase the depth of knee flexion", and "Push the knee forward, do not lock it". The system randomly selects one option from the alternatives based on the severity of the current deviation (e.g., the larger the deviation, the more direct the instruction will be selected) and the historical usage frequency (to avoid repetition). The final feedback is presented in a multimodal form through augmented reality glasses, a tablet, or a field-side projection device: the deviated joint is highlighted on the video screen, a red arrow is superimposed to indicate the correct direction of movement, and a synthesized voice prompt is played. In addition, all evaluation results of this training, including deviation values ​​at each level, group percentile ranking, and final comprehensive score, are encrypted and uploaded to the cloud to update the strength level parameters in the individual physiological parameter model and expand the historical records in the group data pool, thereby achieving continuous evolution of the system's capabilities.

[0039] The system operates within a three-stage cyclical framework. The offline modeling stage initializes individual physiological parameter models for all students at the beginning of the semester. The online assessment stage performs real-time motion data collection, comparison, evaluation, and feedback during each physical education class. The long-term tracking stage automatically summarizes training data every four weeks or at the end of the semester, analyzing trends in students' physiological parameters (such as joint range of motion improvement rates and strength level leaps) and skill progress curves (such as exponential decay fitting of movement completion time). It also generates PDF teaching reports for physical education teachers, including overall class performance distribution, individual progress radar charts, and targeted training suggestions. This three-stage cyclical framework ensures that the system provides both millisecond-level real-time guidance and long-term teaching decisions, achieving an organic unity between micro-intervention and macro-planning.

Claims

1. A student physical education movement correction system based on image recognition, characterized in that, include: The individual physiological parameter modeling module is used to create and maintain a dynamically updated individual physiological parameter model for each registered student. The standard movement dynamic adaptation module is used to dynamically adapt the standard movements in the general standard movement library to the specific student model output by the individual physiological parameter modeling module in real time, so as to generate a personalized standard movement reference sequence for the student. The multimodal data acquisition and processing module is used to collect students' movement data during physical training in real time and perform preprocessing and feature extraction. The group relative assessment and incentive module is used to introduce a group reference system on the basis of individual assessment to generate relative assessment indicators and incentive signals based on peer performance. The personalized correction instruction generation and feedback module is used to receive the real-time student action data stream output by the multimodal data acquisition and processing module and the personalized standard action reference sequence output by the standard action dynamic adaptation module, and to perform action difference analysis and correction feedback.

2. The student physical education movement correction system based on image recognition according to claim 1, characterized in that, The periodic fine-tuning process of the joint range of motion parameters in the individual physiological parameter modeling module is as follows: The system records the maximum joint angles that students actually achieve in multiple training sessions without triggering system alarms, forming a sample set; every preset period, the system uses the statistical values ​​of this sample set to update the corresponding joint rotation limit parameters in the model, and the updated values ​​must not exceed the anatomical safety limit. The module has a built-in parametric human body model engine, which maps the collected static posture and dynamic basic movement test data of students into personalized bone length parameters, joint rotation limit parameters and muscle strength level parameters, so as to form the core data layer of the individual physiological parameter model. The module scales the general spatial trajectory proportionally based on the bone length parameters in the model, performs compliance trimming and smooth transition on the general joint angle sequence based on the joint rotation limit parameters, and adaptively adjusts the intensity curve of the force exertion sequence based on the muscle strength level parameters.

3. The student physical education movement correction system based on image recognition according to claim 2, characterized in that, The statistical value is the 95th percentile of the sample set; the preset period is every 10 effective training iterations. The module includes a visual acquisition unit, an inertial measurement unit integrated interface, and a data processing core; The visual acquisition unit consists of multiple high-definition cameras deployed around the training ground to simultaneously capture motion video streams from different perspectives; The inertial measurement unit integrated interface is used to receive acceleration, angular velocity, and attitude data sent by the wearable device worn by the student; The data processing core is used to synchronize and spatially calibrate multiple video streams, run human pose estimation algorithms to extract skeletal keypoint sequences, and filter and fuse inertial measurement data to output a time-aligned multimodal motion data stream.

4. The student physical education movement correction system based on image recognition according to claim 1, characterized in that, The standard movement dynamic adaptation module follows these rules to adaptively adjust the force exertion timing intensity curve: for students whose muscle strength level parameter is marked as beginner, 80% of the peak intensity in the general force exertion curve is used as their personalized reference peak, and the duration of force exertion is extended. For students whose level is marked as advanced, 110% of the general curve is used as the reference peak, and the preparation period for exertion is compressed.

5. A student physical education movement correction system based on image recognition according to claim 1, characterized in that, The process of screening peer groups with similar physiological conditions in the relative group assessment and incentive module uses multidimensional feature weighted Euclidean distance calculation. Specifically, the following steps are taken: height, lower limb length and trunk length in the individual physiological parameter model are selected as key feature dimensions, and weights are assigned to each dimension based on expert experience. The distance between the current student and all students in the group data pool in these weighted feature spaces is calculated, and individuals whose distance is less than the set threshold are included in the reference group. The module includes a group data pool, a relative evaluation calculator, and a dynamic incentive strategy engine; The group data pool stores and indexes the historical best performance data of all students when completing the same training project; The relative assessment calculator is used to select a peer group with a physiological similarity higher than a preset threshold from the group data pool based on the current student's physiological parameter model, and to calculate the percentile ranking of the current student's various action indicators in the historical data of the reference group. The dynamic incentive strategy engine is used to generate incentive signals based on percentile ranking.

6. The student physical education movement correction system based on image recognition according to claim 5, characterized in that, The formula for calculating the weighted Euclidean distance is as follows: The distance is equal to the height dimension weight multiplied by the square of the difference between the current student's height and the historical student's height, plus the lower limb length dimension weight multiplied by the square of the corresponding difference, plus the trunk length dimension weight multiplied by the square of the corresponding difference, and finally the square root of the sum is taken.

7. A student physical education movement correction system based on image recognition according to claim 1, characterized in that, The correction knowledge base in the personalized correction instruction generation and feedback module is organized using a condition triggering mechanism. Each correction instruction is associated with a condition label defined by the action type, deviation level, and deviation value range; When the deviation data output by the hierarchical action difference analyzer matches a certain condition label, the system randomly selects one of the two to three alternative instructions associated with the condition label for feedback. The module contains a hierarchical motion difference analyzer, which is used to dynamically time-align the real-time motion sequence with the reference sequence in the time dimension, and to decompose the motion difference into three levels in the spatial dimension: joint angle deviation, motion trajectory deviation and force pattern deviation for quantitative evaluation. The module also includes a correction instruction generator, which is used to associate with a predefined correction knowledge base, select instruction text according to the identified deviation type and severity, and provide real-time feedback in conjunction with augmented reality technology or a mobile terminal interface. Meanwhile, the module updates the evaluation results of this training to the individual physiological parameter modeling module and the population data pool.

8. A student physical education movement correction system based on image recognition according to claim 1, characterized in that, The dynamic incentive strategy engine generates incentive signals based on percentile ranking according to the following rules: when the ranking enters the top 20%, a competitive challenge task is generated. When the ranking is in the middle 60%, generate encouraging feedback and a visual progress bar based on small improvements; when the ranking is in the bottom 20%, generate guiding prompts with specific peer comparison cases.

9. A student physical education movement correction system based on image recognition according to claim 1, characterized in that, The system operates within a three-stage cyclical framework that includes offline modeling, online evaluation, and long-term tracking. The offline modeling phase completes the initial construction of student registration and individual physiological parameter models; During the online assessment phase, motion capture, personalized comparison, group reference, and corrective feedback are performed in real time during each training session. By periodically integrating training data during long-term tracking phases, the changing trends of students' physiological parameters and skill progress curves are reassessed, and periodic teaching reports and personalized training plan suggestions are generated for teachers.

10. A student physical education movement correction system based on image recognition according to claim 1, characterized in that, When visual key points are lost due to rapid rotation or occlusion, the data processing core of the multimodal data acquisition and processing module automatically switches to an attitude estimation mode dominated by inertial measurement data, and fuses visual data and inertial data in the state space through complementary filtering.

Images