Motion posture evaluation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of crossing of artificial intelligence and sports science, in particular to a motion posture evaluation method and system based on artificial intelligence, and the method comprises the following steps: S1, multi-modal motion data collection: collecting a motion video of a user through a camera device, and synchronously obtaining basic information of the user; according to the motion posture evaluation method and system based on artificial intelligence, 9 types of core parameters are extracted by collecting 3-5 basic motions, dynamic mapping rules of the parameters and evaluation standards are established by combining age and injury history dynamic correction standards, individual differences such as height, body type, motion basis and rehabilitation progress are adapted, and the evaluation accuracy is improved. According to the method, users of different body types can obtain safe and efficient evaluation direction guidance according with own physiological structures, the body type related evaluation misjudgment rate is reduced, sports injuries caused by disjunction of the evaluation direction and the body type are avoided, meanwhile, the users can more easily reach the action standard matched with themselves, and the obtaining feeling and effect of training and rehabilitation are improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and sports science, specifically to a method and system for evaluating motion posture based on artificial intelligence. Background Technology

[0002] With increasing awareness of sports and health, the demand for sports injury prevention and performance optimization is becoming increasingly urgent. Problems such as muscle strains and joint wear caused by poor sports posture are frequent. Traditional subjective observation and assessment lacks precision and quantitative data support, making it difficult to meet professional needs. Sports posture assessment has emerged based on the development of multidisciplinary technologies. Its core is based on biomechanical principles, combined with sensor technology, computer vision and 3D motion capture technology, breaking through the limitations of traditional naked-eye observation, and achieving precise quantification of parameters such as joint angles, force sequence and center of gravity trajectory. At the same time, the integration of artificial intelligence algorithms further improves the efficiency and accuracy of assessment, providing scientific basis for sports training, rehabilitation medicine and public fitness, and promoting the transformation of sports assessment from experience-based judgment to data-driven.

[0003] Existing posture assessment methods suffer from adaptability issues due to individual differences, resulting in high misjudgment rates and poor user experience, thus limiting their large-scale application in scenarios such as mass fitness and rehabilitation medicine. To address this, an AI-based method and system for motion posture assessment are provided. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based method and system for evaluating movement postures, addressing the adaptability issues arising from individual differences in existing posture evaluation methods, as mentioned in the background section. These issues lead to high misjudgment rates, poor user experience, and limit the large-scale application of existing methods in scenarios such as mass fitness and rehabilitation medicine. To achieve the above objective, this invention provides the following technical solution: an artificial intelligence-based method for evaluating movement postures, comprising the following steps: S1: Multimodal motion data acquisition: Capture user motion videos through camera equipment, simultaneously obtain basic user information, and simultaneously collect user voice feedback and motion scene environment information. Output valid data after preprocessing. S2: Dynamic calibration: Performed upon first use or periodic recalibration, including: S21: Basic motion capture: Guide the user to complete 3-5 basic motions, such as standing naturally, raising arms, bending knees into a half squat, bending the torso laterally, and rotating the wrist. Capture each motion 2-3 times, select 20-30 effective frames that are unobstructed, have low jitter, and have standard posture, and perform time alignment and normalization on the effective frames. S22: Personalized parameter extraction: Calculate nine types of core parameters from valid frames, including limb proportion parameters, joint range of motion, basic motion parameters, body shape adaptation parameters, and joint movement speed thresholds. Use three-dimensional coordinate calculation to improve accuracy and output parameter labels and anomaly correction prompts. S23: Standard template adjustment: Based on personalized parameters, establish mapping rules, adjust the dynamic index thresholds and three-dimensional posture standards of the general template, and generate a dedicated standard template, which includes at least knee joint angle threshold, center of gravity shift threshold, movement symmetry threshold, joint movement speed threshold and three-dimensional posture deviation threshold. S24: Model transfer learning fine-tuning: Using the "MobileNetV3 + Transformer temporal branch" pre-trained model, freezing the first 80% of the feature layers, fine-tuning the evaluation decision layer, inputting valid frame samples containing personalized parameter labels and 3D pose labels, using the joint loss function L_total=0.6×L_param+0.4×L_acc+adding 0.2×L_3D for optimization, the time taken on mobile devices is ≤5 minutes; S3: Movement Posture Assessment: By calling a dedicated standard template and a fine-tuned model, the system analyzes the two-dimensional and three-dimensional joint angles, posture stability, movement symmetry, movement smoothness, joint mobility speed, and injury risk coefficient of the movement video, and generates a comprehensive score of 0-100. S4: Personalized Feedback and Iteration: Output video overlay annotations, voice prompts, VR / AR visualization feedback, and data visualization reports. Regular or triggered recalibration is performed, and templates and models are updated based on user feedback to form a closed-loop optimization.

[0005] More preferably, the basic action acquisition needs to meet the following requirements: Stand naturally with your feet shoulder-width apart, arms hanging naturally at your sides, and facing the camera at an angle of ≤5°. Key points: shoulders, hips, knees, and ankles should not be obstructed. Arms raised horizontally: Slowly raise both arms from a natural hanging position to horizontal, with palms facing down, elbow angle ≥175°, without shrugging shoulders; Half squat with knees bent: This can be performed in small movements during rehabilitation to avoid aggravating pain. Keep your knees aligned with your toes and your back straight. Wrist rotation: With hands hanging naturally, rotate wrists clockwise and counterclockwise once each, without any limbs obstructing the view; The effective frame selection process requires removing frames that are obscured by limbs, have jittery movements, or are subject to abnormal lighting, and retaining the frames with the highest stability.

[0006] More preferably, during the personalized parameter extraction process, if the parameter exceeds the physiological range during parameter anomaly detection, the user is automatically prompted to re-collect the corresponding basic action. Based on age and injury history, the parameters were adjusted as follows: the standard for joint range of motion for users over 60 years old was lowered by 10°, the stability threshold for children was relaxed by 20%, and the corresponding joint range of motion threshold for users with previous joint injuries was lowered by 5°-15° depending on the type of injury.

[0007] More preferably, the pre-trained model is trained based on COCO Keypoints, Human3.6M, and a rehabilitation exercise-specific dataset, with the backbone network being MobileNetV3 combined with a Transformer temporal branch; The fine-tuned samples were expanded to 40-60 frames by slight rotation (±3°), brightness adjustment (±5%), and depth data augmentation. The fine-tuning time on the mobile device was ≤5 minutes, and the evaluation accuracy after fine-tuning was ≥90%.

[0008] More preferably, the periodic recalibration triggering mechanism includes: Time-triggered: Automatic reminder every 3 months; Event triggers: User weight change ≥5kg, after sports injury rehabilitation, ≥3 assessment and error correction feedbacks, or when sports goals are changed; After recalibration, the custom template and fine-tuned model are automatically updated, and historical parameter change records and model fine-tuning logs are retained.

[0009] A further preferred embodiment of the system for performing the above-described AI-based motion posture assessment method includes: Multimodal data acquisition module: includes RGB and depth camera units, basic information input unit, and newly added: voice acquisition unit, environmental perception unit, data preprocessing unit, outputting motion video, effective frames of basic actions, and multi-dimensional user information; Dynamic calibration module: This is the core module, which includes a basic motion acquisition unit, a personalized parameter extraction unit, a standard template adjustment unit, a transfer learning fine-tuning unit, and a calibration management unit. The personalized parameter extraction unit calculates nine types of core parameters, and the standard template adjustment unit generates a custom template containing three-dimensional posture standards. AI assessment module: It calls up exclusive templates and personalized models to calculate multi-dimensional assessment indicators (including damage risk coefficients), and adopts model quantization and pruning techniques. The model size is ≤30MB and the inference latency on mobile devices is ≤80ms. Personalized feedback module: includes video overlay unit, voice prompt unit, VR / AR visualization unit and data visualization unit, adaptable to multiple scenarios and user groups; User profiling and big data module: includes user profiling unit, damage risk warning unit, group optimization unit and progress curve unit; Edge-Cloud Collaborative Architecture: The terminal performs data collection and real-time feedback, the edge nodes perform model inference, and the cloud performs template updates and big data analysis, supporting offline evaluation and cross-device synchronization.

[0010] Furthermore, preferably, the dynamic calibration module supports offline mode: valid frames and personalized parameters are stored locally with AES-256 encryption, and incrementally synchronized to the cloud after network recovery; The AI ​​evaluation module employs model quantization and pruning techniques, with a model size ≤30MB, mobile inference latency ≤80ms, and supports real-time evaluation.

[0011] Furthermore, the personalized feedback module is adapted to multiple scenarios: Fitness scenario: Output movement completion rate, muscle activation suggestions; Rehabilitation scenario: Provide suggestions for rehabilitation progress and training intensity adjustments; Competitive scenarios: Output action efficiency, tactical improvement suggestions; Feedback formats are tailored to different user groups: seniors use large fonts, voice prompts, and simplified charts; children use cartoon animations and NPC guidance; professional users use data reports, heatmaps, and 3D posture analysis reports, with support for VR / AR visualization comparison.

[0012] In a further preferred embodiment, the user profiling and big data module employs horizontal federated learning: users train and fine-tune model parameters locally, while parameters are aggregated in the cloud without transmitting the original data.

[0013] More preferably, the group optimization unit generates segmented group templates and rehabilitation scenario templates based on anonymous parameters to improve the temporary assessment accuracy of uncalibrated users.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: In this invention, 9 core parameters are extracted by collecting 3 to 5 basic movements and dynamically corrected by combining age and injury history. A dynamic mapping rule between parameters and assessment standards is established to adapt to individual differences such as height, body type, sports foundation, and rehabilitation progress. This allows users of different body types to obtain safe and efficient assessment guidance that conforms to their own physiological structure, reducing the misjudgment rate of body type-related assessments and avoiding sports injuries caused by the disconnect between assessment direction and body type. At the same time, it makes it easier for users to achieve movement standards that suit them, improving the sense of gain and effect of training and rehabilitation.

[0015] In this invention, RGB and depth data are integrated to calculate three-dimensional coordinates, and a new three-dimensional posture deviation threshold is added to solve the defect that two-dimensional images cannot distinguish between forward and backward displacements. At the same time, temporal analysis (Transformer branch) is introduced to evaluate the smoothness and rhythm of movements, covering all dimensions of static posture and dynamic process, and meeting the refined needs of competitive training and rehabilitation medicine.

[0016] In this invention, a dedicated standard skeleton is overlaid on the video, highlighting incorrect joints to solve the visual misleading problem of a general skeleton being out of sync with the user's body shape. Voice prompts are associated with personalized thresholds, large fonts and simplified charts are adapted for the elderly, and cartoon NPCs are used to guide children. Professional users can generate heat maps and 3D posture reports. At the data level, parameter change curves are generated, such as a knee joint range of motion rating trend chart from 60° to 72°, allowing users to intuitively perceive progress. Attached Figure Description

[0017] Figure 1 This is a flowchart of the motion posture evaluation method of the present invention; Figure 2 This is a schematic diagram of the motion posture evaluation system of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 - Figure 2 The present invention provides a technical solution: a motion posture evaluation method based on artificial intelligence, characterized by comprising the following steps: S1: Multimodal motion data acquisition: Capture user motion videos through camera equipment, simultaneously obtain basic user information, such as age, weight, height, exercise goals, basic exercise level, and history of injury. Simultaneously collect user voice feedback and motion scene environment information, such as voice feedback (e.g., difficulty in movement, joint pain) and motion scene environment information (e.g., light intensity, background complexity). After preprocessing, output effective data, and optimize data quality through noise reduction, alignment, and illumination compensation to provide multi-dimensional support for subsequent calibration and evaluation. S2: Dynamic calibration: Performed upon first use or periodic recalibration, forming the basis for personalized adaptation, including: S21: Basic Motion Acquisition: Guide the user to complete 3-5 basic motions: standing naturally, arms outstretched, knees bent in a half squat, trunk flexion, and wrist rotation. Acquire each motion 2-3 times, selecting 20-30 valid frames that are unobstructed, have low jitter, and conform to the correct posture. Frames with occlusion, jitter, inter-frame joint displacement ≥5px, and abnormal lighting are discarded. Perform temporal alignment and normalization on the valid frames. The basic motion acquisition must meet the following requirements: Stand naturally with your feet shoulder-width apart, arms hanging naturally at your sides, and facing the camera at an angle of ≤5°. Key points: shoulders, hips, knees, and ankles should not be obstructed. Arms raised horizontally: Slowly raise both arms from a natural hanging position to horizontal, with palms facing down, elbow angle ≥175°, without shrugging shoulders; Half squat with knees bent: This can be performed in small movements during rehabilitation to avoid aggravating pain. Keep your knees aligned with your toes and your back straight. Wrist rotation: With hands hanging naturally, rotate wrists clockwise and counterclockwise once each, without any limbs obstructing the view; Valid frame selection requires removing frames with limb occlusion, jittery movements, and abnormal lighting, and retaining the frames with the highest stability. S22: Personalized Parameter Extraction: Calculate nine core parameters from the valid frames, including: Limb proportion parameters: Height H = Head vertex Y coordinate - Ankle Y coordinate, Shoulder width S = | Left shoulder X coordinate - Right shoulder X coordinate|, Leg length L = Hip Y coordinate - Ankle Y coordinate, Arm length A = Shoulder point Y coordinate - Wrist point Y coordinate; Range of motion of joints: shoulder abduction angle α = angle between the shoulder-elbow-wrist vector and the vertical direction; knee flexion and extension limit β = angle between the hip-knee-ankle vector; lumbar lateral flexion angle γ = angle between the shoulder-hip-ankle vector and the vertical direction. Basic motion parameters: Movement stability S_stab = standard deviation of joint coordinates of 3 movements / mean; Movement symmetry Sym_basic = 1 - |left limb height - right limb height| / average height; Body type fit parameters: BMI = weight / (H²), trunk / leg length ratio T / L = (HL) / L; Joint mobility threshold: V_joint = the maximum speed of joint movement in a basic movement (such as the elbow joint speed when the arm is raised horizontally), reflecting the user's mobility. Three-dimensional coordinate calculation is used to improve accuracy, and output parameter labels and anomaly correction prompts are provided. When anomaly detection occurs during the personalized parameter extraction, if the parameter exceeds the physiological range, the user is automatically prompted to re-collect the corresponding basic action. Based on age and injury history, the parameters were adjusted as follows: the standard for joint range of motion for users over 60 years old was lowered by 10°, the stability threshold for children was relaxed by 20%, and the corresponding joint range of motion threshold for users with previous joint injuries was lowered by 5°-15° depending on the type of injury. S23: Standard Template Adjustment: Establish mapping rules based on personalized parameters, adjust the dynamic index thresholds of the general template, and add: 3D pose standards to generate a dedicated standard template, which includes at least: Knee joint angle threshold = 90° - max ((165cm - H) / 5cm × 2°, 140° - β × 0.5°); The threshold for shifting the center of gravity backward is calculated as 3cm + max(0, (BMI-24)×0.5cm). Action symmetry threshold = 5° + (1 - Sym_basic) × 3°; The joint movement speed threshold is set to V_joint × 1.2 to avoid misjudgment due to movements that are too fast or too slow. Three-dimensional posture deviation threshold: the deviation of the torso's forward tilt depth is ≤2cm to avoid misjudgment in two-dimensional images.

[0020] S24: Model Transfer Learning Fine-tuning: A pre-trained model using "MobileNetV3 + Transformer Temporal Branch" was employed, trained on COCO Keypoints, Human3.6M, and a rehabilitation exercise dataset. The first 80% of the feature layers were frozen, including the Transformer temporal feature layer. The evaluation decision layer was fine-tuned. Valid frame samples containing personalized parameter labels and 3D pose labels were input. A joint loss function L_total = 0.6 × L_param + 0.4 × L_acc + 0.2 × L_3D was added for optimization. Here, L_param is the mean squared error between predicted and actual parameters, L_acc is the cross-entropy loss between the evaluation result and manually labeled data, and L_3D is the mean squared error of the 3D pose prediction bias. Fine-tuned samples were expanded to 40-60 frames through slight rotation (±3°), brightness adjustment (±5%), and the addition of depth data augmentation (simulating shooting at different distances). The time taken on mobile devices was ≤5 minutes. After fine-tuning, the evaluation accuracy was ≥90%, and the time taken on mobile devices was ≤5 minutes. The pre-trained model was trained on COCO Keypoints, Human3.6M, and rehabilitation exercise datasets, with the backbone network being MobileNetV3 combined with the Transformer temporal branch. The fine-tuned samples were expanded to 40-60 frames by slight rotation (±3°), brightness adjustment (±5%), and depth data augmentation. The fine-tuning time on the mobile device was ≤5 minutes, and the evaluation accuracy after fine-tuning was ≥90%. S3: Movement Posture Assessment: By calling a dedicated standard template and a fine-tuned model, the system analyzes the two-dimensional and three-dimensional joint angles, posture stability, movement symmetry, movement smoothness, joint mobility speed, and injury risk coefficient of the movement video. A weighted scoring algorithm is used, and the weight of each indicator is dynamically adjusted based on the user's movement goals: fitness scenarios focus on the standardization of movements, while rehabilitation scenarios focus on the range of joint mobility and safety, generating a comprehensive score of 0-100. S4: Personalized Feedback and Iteration: Output video with overlaid annotations, including a custom standard skeleton highlighting faulty joints and 3D pose deviation annotations, voice prompts, VR / AR visualization feedback, and data visualization reports. Regular or triggered recalibration is implemented, and templates and models are updated based on user feedback to form a closed-loop optimization. The regular recalibration trigger mechanism includes: Time-triggered: Automatic reminder every 3 months; Event triggers: User weight change ≥5kg, after sports injury rehabilitation, ≥3 assessment and error correction feedbacks, or when sports goals are changed; After recalibration, the custom template and fine-tuned model are automatically updated, and historical parameter change records and model fine-tuning logs are retained.

[0021] In this embodiment, as Figure 2 Systems based on artificial intelligence-based motion posture assessment methods include: Multimodal data acquisition module: includes RGB and depth camera units, compatible with mobile phones / tablets / smart fitness mirrors, supports 1080P video acquisition and synchronous acquisition of depth data, basic information input unit, newly added: voice acquisition unit, environmental perception unit, data preprocessing unit, outputs motion video, effective frames of basic movements, and multi-dimensional user information; Dynamic calibration module: This is the core module and includes: Basic motion capture unit: Guides users to complete 3-5 basic motions through voice, text, and animation, captures motion videos, and outputs valid frames; Personalized parameter extraction unit: Calculates 9 types of core parameters from valid frames, improves parameter accuracy by combining 3D coordinate data, and outputs parameter labels and anomaly correction prompts; Standard template adjustment unit: Based on parameter labels and mapping rules, adjust the dynamic index thresholds and 3D pose standards of general templates to generate user-specific standard templates, which are then stored in the personalized template library; Transfer learning fine-tuning unit: Load the "MobileNetV3+Transformer temporal branch" pre-trained model, freeze the feature layer, fine-tune the decision layer, and output a personalized model adapted to the user; Calibration Management Unit: Triggers periodic / active recalibration, stores calibration history data (parameter changes, template update records, model fine-tuning logs), and supports calibration data backtracking and comparison; The personalized parameter extraction unit calculates nine types of core parameters, the standard template adjustment unit generates a dedicated template containing three-dimensional posture standards, and the dynamic calibration module supports offline mode: local storage of valid frames and personalized parameters encrypted with AES-256, and incremental synchronization to the cloud after network recovery. The AI ​​evaluation module employs model quantization and pruning techniques, with a model size ≤30MB, mobile inference latency ≤80ms, and supports real-time evaluation. AI assessment module: It calls up exclusive templates and personalized models to calculate multi-dimensional assessment indicators (including damage risk coefficients), and adopts model quantization and pruning techniques. The model size is ≤30MB and the inference latency on mobile devices is ≤80ms. Personalized feedback module: Includes video overlay unit, voice prompt unit, VR / AR visualization unit, and data visualization unit, adaptable to multiple scenarios and user groups. The personalized feedback module adapts to multiple scenarios: Fitness scenario: Output movement completion rate, muscle activation suggestions; Rehabilitation scenario: Provide suggestions for rehabilitation progress and training intensity adjustments; Competitive scenarios: Output action efficiency, tactical improvement suggestions; Feedback formats are tailored to different user groups: seniors use large fonts, voice prompts, and simplified charts; children use cartoon animations and NPC guidance; professional users use data reports, heatmaps, and 3D pose analysis reports, with support for VR / AR visualization comparison. The User Profile and Big Data Module includes a User Profile Unit, an Injury Risk Warning Unit, a Group Optimization Unit, and a Progress Curve Unit. The User Profile and Big Data Module adopts horizontal federated learning: users train and fine-tune model parameters locally, while parameters are aggregated in the cloud without transmitting the original data. The Group Optimization Unit generates subdivided group templates and rehabilitation scenario templates based on anonymous parameters, which in turn improves the temporary assessment accuracy of uncalibrated users. Edge-Cloud Collaborative Architecture: Terminal devices—mobile phones, tablets, and smart fitness mirrors—perform data acquisition, real-time feedback (video overlay, voice prompts, VR / AR visualization), and offline evaluation (local storage of valid frames and personalized models, protected by AES-256 encryption); edge nodes (routers / edge servers) perform model inference and data preprocessing, reducing the computing power pressure on the terminals; the cloud performs template updates, big data analysis, federated learning parameter aggregation, and generation of templates for segmented groups. After network recovery, the terminals and the cloud incrementally synchronize data, supporting cross-device synchronization of user data and personalized models.

[0022] The following is a detailed explanation using several examples: In this embodiment, 1. Multimodal data acquisition: The user uses the rear camera of the mobile phone, which needs to support depth data acquisition, to start the system and enter basic information: age 35 years old, weight 70kg, height 165cm, the selected exercise goal is muscle gain, the basic exercise level is moderate, and there is no history of injury; the system guides the user to shoot a squat exercise video, and simultaneously collects voice feedback and environmental information. The indoor lighting is sufficient and the background is simple; the data preprocessing unit removes 2 occluded frames and 3 jittery frames, and retains 25 valid frames.

[0023] 2. Dynamic calibration: Basic motion capture: The system guides users through five basic motions via animation: standing naturally, raising arms horizontally, bending knees into a half squat, lateral flexion of the torso, and wrist rotation. Each motion is captured twice, and 30 valid frames are selected. Personalized parameter extraction: Calculate 9 types of parameters: H=165cm, S=45cm, L=90cm, A=60cm, α=175°, β=135°, γ=30°, S_stab=0.08, Sym_basic=0.92, V_joint=0.3m / s, BMI=25.7, T / L=0.83; No abnormal parameters were found, no re-collection is required; Standard template adjustment: Generate exclusive thresholds: Knee joint angle threshold = 90° - max ((165-165) / 5×2°, 140-135×0.5°) = 87.5°, Center of gravity shift threshold = 3 + max (0, (25.7-24)×0.5) = 3.85cm, Movement symmetry threshold = 5 + (1-0.92)×3 = 5.24°, Joint movement speed threshold = 0.3×1.2 = 0.36m / s, Three-dimensional posture deviation threshold ≤ 2cm; Model fine-tuning: Load the "MobileNetV3+Transformer temporal branch" pre-trained model, freeze 80% of the feature layers, input 30 valid frame labeled samples, and optimize using L_total=0.6L_param+0.4L_acc+0.2L_3D. The time taken on the mobile device was 3.5 minutes, and the model accuracy after fine-tuning was 92.3%.

[0024] 3. Movement Posture Assessment Using a custom template and fine-tuned model, we analyzed squat videos: mean joint angle 85°, postural stability S_stab=0.09, movement symmetry Sym=0.91, movement smoothness 0.88, joint motion speed 0.32m / s (≤0.36m / s), three-dimensional posture deviation 1.5cm, injury risk coefficient 0.1, and overall score 91 points.

[0025] 4. Personalized feedback and iteration Output feedback: The video is overlaid with a custom standard frame, highlighting the right knee, which is slightly valgus (1.2° deviation). The voice prompt says, "Your squat posture is generally up to standard, but your right knee is slightly valgus. We suggest adjusting your toes to shoulder width." The system compares the user's posture with the custom template in real time in the VR visualization scene. If the user does not provide any feedback, the system records the evaluation data and sets a recalibration to be triggered after 3 months.

[0026] In this embodiment, 1. Multimodal data acquisition: The athlete uses a tablet device to enter basic information: age 22, weight 65kg, height 180cm, exercise goal "improve running efficiency", basic exercise level "high", no history of injury; 1000-meter running training video is collected, and voice feedback "no discomfort" is collected simultaneously, and the environment information is track and field; 30 valid frames are retained after preprocessing.

[0027] 2. Dynamic calibration Basic motion capture: Complete four basic motions: standing naturally, raising arms horizontally, lateral flexion of the torso, and wrist rotation, and select 25 valid frames; Personalized parameter extraction: Calculate 9 types of parameters: H=180cm, S=50cm, L=100cm, A=65cm, α=180°, β=140°, γ=35°, S_stab=0.05, Sym_basic=0.95, V_joint=0.4m / s, BMI=20.1, T / L=0.8; Standard template adjustment: Adapted to running scenarios, the cadence threshold is adjusted to 180 steps / minute, and personalized adaptations are made for hip flexion and extension and knee flexion and extension based on athlete goals and joint angle thresholds. Model fine-tuning: Input 30 valid frame labeled samples, fine-tuning took 2.8 minutes, and the model accuracy was 94.1%.

[0028] 3. Movement Posture Assessment Assessment results: Step frequency 172 steps / minute, average hip flexion-extension angle 38°, average knee flexion-extension angle 110°, movement fluency 0.92, overall score 88 points; the assessment suggests that "the step frequency is close to the target, but the left hip is not fully engaged, which affects running efficiency."

[0029] 4. Personalized feedback and iteration Output feedback: Data reports show cadence trends and joint angle distribution heatmaps. The VR scene simulates the optimized posture, with enhanced left hip power. A professional voice prompt suggests "Increase the left hip flexion and extension range and increase cadence to 180 steps / minute." When athletes adopt the suggestions, the system records training data, sets a recalibration after one month, and tracks the optimization effect.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A motion posture evaluation method based on artificial intelligence, characterized in that, Includes the following steps: S1: Multimodal motion data acquisition: Capture user motion videos through camera equipment, simultaneously obtain basic user information, and simultaneously collect user voice feedback and motion scene environment information. Output valid data after preprocessing. S2: Dynamic calibration: Performed upon first use or periodic recalibration, including: S21: Basic motion capture: Guide the user to complete 3-5 basic motions, such as standing naturally, raising arms, bending knees into a half squat, bending the torso laterally, and rotating the wrist. Capture each motion 2-3 times, and select 20-30 effective frames that are unobstructed, have low jitter, and have standard posture. Perform time alignment and normalization processing on the effective frames. S22: Personalized parameter extraction: Calculate nine types of core parameters from valid frames, including limb proportion parameters, joint range of motion, basic motion parameters, body shape adaptation parameters, and joint movement speed thresholds. Use three-dimensional coordinate calculation to improve accuracy and output parameter labels and anomaly correction prompts. S23: Standard template adjustment: Based on personalized parameters, establish mapping rules, adjust the dynamic index thresholds and three-dimensional posture standards of the general template, and generate a dedicated standard template, which includes at least knee joint angle threshold, center of gravity shift threshold, movement symmetry threshold, joint movement speed threshold and three-dimensional posture deviation threshold. S24: Model transfer learning fine-tuning: The "MobileNetV3 + Transformer temporal branch" pre-trained model is used. The first 80% of the feature layers are frozen, and the evaluation decision layer is fine-tuned. Valid frame samples with personalized parameter labels and 3D pose labels are input. The joint loss function L_total=0.6×L_param+0.4×L_acc+0.2×L_3D is used for optimization. The time taken on mobile devices is ≤5 minutes. S3: Movement Posture Assessment: By calling a dedicated standard template and a fine-tuned model, the system analyzes the two-dimensional and three-dimensional joint angles, posture stability, movement symmetry, movement smoothness, joint mobility speed, and injury risk coefficient of the movement video, and generates a comprehensive score of 0-100. S4: Personalized Feedback and Iteration: Output video overlay annotations, voice prompts, VR / AR visualization feedback, and data visualization reports. Regular or triggered recalibration is performed, and templates and models are updated based on user feedback to form a closed-loop optimization.

2. The motion posture evaluation method based on artificial intelligence according to claim 1, characterized in that, The basic action collection must meet the following requirements: Stand naturally with your feet shoulder-width apart, arms hanging naturally at your sides, and face the camera at an angle of ≤5°. Key points such as shoulders, hips, knees, and ankles should be unobstructed. Arms raised horizontally: Slowly raise both arms from a natural hanging position to horizontal, with palms facing down, elbow angle ≥175°, without shrugging shoulders; Half squat with knees bent: This can be performed in small movements during rehabilitation to avoid aggravating pain. Keep your knees aligned with your toes and your back straight. Wrist rotation: With hands hanging naturally, rotate wrists clockwise and counterclockwise once each, without any limbs obstructing the view; The effective frame selection process requires removing frames that are obscured by limbs, have jittery movements, or are subject to abnormal lighting, and retaining the frames with the highest stability.

3. The motion posture evaluation method based on artificial intelligence according to claim 1, characterized in that, When anomaly detection occurs during the personalized parameter extraction process, if a parameter exceeds the physiological range, the user will be automatically prompted to re-collect the corresponding basic action. Based on age and injury history, the parameters were adjusted as follows: the standard for joint range of motion for users over 60 years old was lowered by 10°, the stability threshold for children was relaxed by 20%, and the corresponding joint range of motion threshold for users with previous joint injuries was lowered by 5°-15° depending on the type of injury.

4. The motion posture evaluation method based on artificial intelligence according to claim 1, characterized in that, The pre-trained model is trained based on COCO Keypoints, Human3.6M, and a rehabilitation exercise-specific dataset, with the backbone network being MobileNetV3 combined with a Transformer temporal branch; The fine-tuned samples were expanded to 40-60 frames by slight rotation (±3°), brightness adjustment (±5%), and depth data augmentation. The fine-tuning time on the mobile device was ≤5 minutes, and the evaluation accuracy after fine-tuning was ≥90%.

5. The motion posture evaluation method based on artificial intelligence according to claim 1, characterized in that, The periodic recalibration triggering mechanism includes: Time-triggered: Automatic reminder every 3 months; Event triggers: Actively triggered when user weight changes by ≥5kg, after sports injury rehabilitation, after ≥3 assessment and error correction feedbacks, or when sports goals change; After recalibration, the custom template and fine-tuned model are automatically updated, and historical parameter change records and model fine-tuning logs are retained.

6. A system for evaluating motion posture based on artificial intelligence, characterized in that, The system for performing the method of any one of claims 1-5 comprises: Multimodal data acquisition module: includes RGB + depth camera unit, basic information input unit, voice acquisition unit, environmental perception unit, and data preprocessing unit, outputting motion video, effective frames of basic actions, and multi-dimensional user information; Dynamic calibration module: This is the core module, which includes a basic motion acquisition unit, a personalized parameter extraction unit, a standard template adjustment unit, a transfer learning fine-tuning unit, and a calibration management unit. The personalized parameter extraction unit calculates nine types of core parameters, and the standard template adjustment unit generates a custom template containing three-dimensional posture standards. AI assessment module: It calls up exclusive templates and personalized models to calculate multi-dimensional assessment indicators (including damage risk coefficients), and adopts model quantization and pruning techniques. The model size is ≤30MB and the inference latency on mobile devices is ≤80ms. Personalized feedback module: includes video overlay unit, voice prompt unit, VR / AR visualization unit and data visualization unit, adaptable to multiple scenarios and user groups; User profiling and big data module: includes user profiling unit, damage risk warning unit, group optimization unit and progress curve unit; Edge-Cloud Collaborative Architecture: The terminal performs data collection and real-time feedback, the edge nodes perform model inference, and the cloud performs template updates and big data analysis, supporting offline evaluation and cross-device synchronization.

7. The system for motion posture evaluation based on artificial intelligence according to claim 6, characterized in that, The dynamic calibration module supports offline mode: valid frames and personalized parameters are stored locally with AES-256 encryption, and incrementally synchronized to the cloud after the network is restored; The AI ​​evaluation module employs model quantization and pruning techniques, with a model size ≤30MB, mobile inference latency ≤80ms, and supports real-time evaluation.

8. The system for motion posture evaluation based on artificial intelligence according to claim 6, characterized in that, The personalized feedback module is adaptable to multiple scenarios: Fitness scenario: Output movement completion rate, muscle activation suggestions; Rehabilitation scenario: Provide suggestions for rehabilitation progress and training intensity adjustments; Competitive scenarios: Output action efficiency, tactical improvement suggestions; Feedback formats are tailored to the user group: elderly users should use large fonts, voice prompts, and simplified charts. Children use cartoon animation and NPC guidance; Professional users can use data reports, heatmaps, and 3D posture analysis reports to support VR / AR visualization and comparison.

9. The system for motion posture evaluation based on artificial intelligence according to claim 6, characterized in that, The user profiling and big data module adopts horizontal federated learning: users train and fine-tune model parameters locally, and aggregate parameters in the cloud without transmitting the original data.

10. The system of the motion posture evaluation method based on artificial intelligence according to claim 6, characterized in that, The group optimization unit generates segmented group templates and rehabilitation scenario templates based on anonymous parameters, which in turn improves the temporary assessment accuracy of uncalibrated users.

Images