Multi-target physical training action evaluation feedback system based on artificial intelligence

Abstract

The invention discloses a multi-target physical training action evaluation feedback system based on artificial intelligence, and the system obtains video data and physiological data in a multi-target motion scene through a multi-dimensional data collection module, and optimizes skeleton key point data through a data preprocessing module; and the continuous identity association of the moving target is realized by the multi-target tracking and binding module. The motion modeling module constructs motion state variables based on a target constraint dynamic optimization model and applies physiological and physical constraints, the fusion evaluation module combines local motion decomposition and physiological data correlation analysis to generate a comprehensive score, and the personalized feedback module outputs hierarchical improvement suggestions. And the adaptive standard library module iteratively optimizes standard action parameters according to the user type and historical data. The system solves the problems that in the prior art, multi-target tracking is confused, action evaluation and physiological states are disjointed, and feedback is lack of accuracy, and efficient analysis, accurate evaluation and personalized guidance of actions in a multi-target scene are achieved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence computer vision and sports training evaluation technology, specifically to a multi-objective sports training motion evaluation feedback system based on artificial intelligence. Background Technology

[0002] With the application of artificial intelligence and computer vision technology in the sports field, intelligent training and evaluation tools are gradually becoming an important support for sports teaching and training. While some existing systems can perform video analysis and evaluation of single-target movements, in multi-target scenarios such as multi-person team training, there are issues with motion target tracking easily becoming confused and identity binding being interrupted, making it difficult to accurately distinguish the motion data of different targets.

[0003] Meanwhile, existing motion assessment technologies mostly focus only on the matching degree between movement posture and standard movements, ignoring the correlation between physiological state and movement. For example, while some systems can identify movement deviations, they cannot determine whether the deviations are caused by physiological fatigue, leading to one-sided assessment results and a lack of targeted feedback. In addition, existing feedback mechanisms are mostly overall scores or simple deviation prompts, without hierarchical classification of deviations, making it difficult for users to quickly locate the root cause of the problem; and standard movement libraries are mostly based on fixed parameters, which cannot adapt to the training needs of users of different levels, nor can they be dynamically optimized based on historical training data, resulting in poor adaptability.

[0004] In summary, existing technologies have significant shortcomings in terms of multi-target tracking accuracy, the refinement of action and physiological fusion assessment feedback levels, and the adaptive standard library. There is an urgent need for an intelligent system that can adapt to complex multi-target scenarios, achieve collaborative assessment of action and physiology, provide accurate hierarchical feedback, and dynamically optimize the standard library, so as to improve the intelligence level and training effect of sports training. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a multi-objective sports training movement evaluation and feedback system based on artificial intelligence to solve the problems mentioned in the background technology.

[0006] According to one aspect of this application, an artificial intelligence-based multi-objective sports training movement evaluation and feedback system includes: The multi-dimensional data acquisition module is used to collect video and physiological data in multi-target motion scenes and extract the skeletal key point data of each moving target. The data preprocessing module is used to perform illumination compensation, occlusion completion, and noise filtering on the skeletal key point data, and output optimized skeletal key point data. The multi-target tracking and binding module is used to achieve real-time tracking and continuous identity binding of multiple moving targets based on video data and optimized skeletal key point data. The motion modeling module is used to build a dynamic optimization model of target constraints based on optimized skeletal keypoint data, and to model and predict the motion state of each moving target in real time. The fusion assessment module is used to perform local decomposition of the overall movement of each motion target and generate a comprehensive score by combining movement matching analysis and physiological data correlation analysis. The personalized feedback module is used to generate real-time feedback information and training summary reports for each motion target based on the comprehensive score and the level of motion deviation. The adaptive standard library module stores standard motion parameters for different motion types and iteratively optimizes these parameters based on user type and historical training data.

[0007] Preferably, the multi-dimensional data acquisition module includes a camera unit and a physiological data acquisition unit; the camera unit uses four wide-angle cameras respectively arranged on the upper, lower, left, and right sides of the sports field to capture panoramic video streams in real time; the physiological data acquisition unit collects heart rate, blood oxygen, lactate concentration, and respiratory rate data of each sports target through wearable devices; the skeletal key point data is extracted from the video stream through a deep learning algorithm, and the extracted skeletal key points include joint position, joint angle, joint velocity, and joint acceleration.

[0008] Preferably, the data preprocessing module's processing procedure includes: Illumination compensation: Calculate the average brightness of the current video frame, and adjust the intensity value of each pixel in the video frame according to the preset target brightness to make the brightness of the current frame approach the preset target brightness; Occlusion completion: Based on the Kalman filter algorithm, combined with the skeletal trajectory data of moving targets in historical frames, the position data of occluded joints are predicted and completed; Noise filtering: The moving average method is used to calculate the average value of the skeletal key point positions within a preset number of adjacent frames, and the average value is used to replace the abnormally fluctuating skeletal key point position data in the current frame.

[0009] Preferably, the implementation process of the multi-target tracking binding module includes: The DeepSORT algorithm is used to perform preliminary tracking of multiple moving targets in the video stream, generating a temporary tracking trajectory for each target; Extract the skeletal feature parameters for each motion target, including the ratio of shoulder width to height, the ratio of leg length to height, and the ratio of inter-joint distance. By associating temporary tracking trajectories with skeletal feature parameters, when tracking is interrupted due to occlusion of the moving target, the target identity is re-matched based on the skeletal feature parameters to achieve continuous identity binding.

[0010] Preferably, the target constraint dynamic optimization model in the motion modeling module includes motion state variables, motion standard values, and motion optimization functions; the motion state variables include the joint angles, joint velocities, and joint accelerations of the moving target; the motion standard values ​​are preset reference standards for the corresponding motion types stored in the adaptive standard library module; the motion optimization function is used to minimize the deviation between the actual motion state of the moving target and the motion standard values, while applying motion range constraints and motion speed constraints; the motion range constraints limit the joint angles to within the physiological limits of the human body; the motion speed constraints limit the joint motion speed to not exceed a preset speed threshold.

[0011] Preferably, the evaluation process of the fusion evaluation module includes: The overall motion of each moving target is decomposed into several local key motions, and the feature parameters of each local key motion are extracted. The feature parameters include joint angle, joint velocity and joint acceleration. Calculate the similarity between the feature parameters of each local key action and the corresponding parameters in the action standard value, generate a score for each local key action based on the similarity, and calculate the basic action score by weighting the scores of all local key actions. The correlation between physiological data and movement is analyzed. When the basic movement score reaches a preset threshold but the physiological data exceeds the normal physiological range of the corresponding movement, the physiological fit score is reduced; when the basic movement score does not reach the preset threshold but the physiological data is within the normal physiological range, the physiological fit score remains unchanged. The basic movement score and the physiological adaptation score are weighted according to preset weights to obtain the comprehensive movement score for each exercise goal. The preset weights are set according to the type of exercise: the basic movement score for limb movement exercise has a weight of 0.6 and the physiological adaptation score has a weight of 0.4, while the basic movement score for endurance exercise has a weight of 0.3 and the physiological adaptation score has a weight of 0.7.

[0012] Preferably, the feedback process of the personalized feedback module includes: Based on the comprehensive score and the results of local motion decomposition, the motion deviation level is determined. The motion deviation level includes joint level, motion segment level and overall level. Joint-level deviation feedback generates adjustment suggestions for angular and velocity deviations of individual joints; motion segment-level deviation feedback generates adjustment suggestions for rhythmic force application timing deviations of key local movements; and overall-level deviation feedback generates adjustment suggestions for coordination and rhythm deviations of the overall movement. The training summary report includes the overall score change curve for each exercise target, the distribution of local movement scores, and the results of physiological data and movement adaptability analysis.

[0013] Preferably, the optimization process of the adaptive standard library module includes: Different standard action parameters are preset according to user type, which includes novice, intermediate, and professional users. The standard action parameter deviation tolerance range for novice users is larger than that for intermediate users, and the standard action parameter deviation tolerance range for intermediate users is larger than that for professional users. Regularly analyze the overall score distribution and deviation types in historical training data. When the frequency of a certain type of deviation exceeds the preset threshold, adjust the deviation tolerance range of the corresponding standard action parameters or optimize the action standard value. Based on historical training data for different sports goals, a personalized subset of standard movement parameters is generated for each sports goal, making the standard movement parameters more suitable for individual physiological characteristics.

[0014] An AI-based multi-objective sports training movement evaluation and feedback method includes the following steps: Step 1: Collect video and physiological data from multi-target motion scenes using a multi-dimensional data acquisition module, and extract skeletal key point data for each moving target; Step 2: The data preprocessing module performs illumination compensation, occlusion completion, and noise filtering on the skeletal keypoint data, and outputs optimized skeletal keypoint data. Step 3: Based on video data and optimized skeletal keypoint data, the multi-target tracking and binding module enables real-time tracking and continuous identity binding of multiple moving targets. Step 4: Based on the optimized skeletal keypoint data, the motion modeling module establishes a target constraint dynamic optimization model to model and predict the motion state of each moving target in real time. Step 5: The overall motion of each motion target is locally decomposed through the fusion evaluation module, and a comprehensive score is generated by combining motion matching analysis and physiological data correlation analysis. Step 6: Based on the comprehensive score and the level of movement deviation, the personalized feedback module generates real-time feedback information and training summary report for each movement target. Step 7: Iteratively optimize the standard action parameters based on user type and historical training data using the adaptive standard library module.

[0015] Preferably, The advantages of this application compared to existing technologies are: 1. High accuracy in multi-target tracking and binding: By combining the DeepSORT algorithm with skeletal feature parameters, real-time tracking and continuous identity binding of multiple moving targets are achieved, solving the identity confusion problem caused by occlusion and intersection, and adapting to multi-target scenarios such as team training.

[0016] 2. Comprehensive assessment integrating movement and physiology: It not only focuses on the matching degree between movement and standard, but also combines physiological data correlation analysis to generate a comprehensive score, avoiding the one-sidedness of single movement assessment, and making the assessment results more objective.

[0017] 3. Refined and precise feedback hierarchy: Based on deviation hierarchy at the joint level, motion segment level, and overall level, targeted feedback is generated to help users quickly locate the root cause of the problem and improve improvement efficiency.

[0018] 4. Strong adaptability of the standard library: It dynamically optimizes the standard action parameters based on user type and historical training data to meet the needs of users of different skill levels, and generates personalized standard subsets to improve the system's adaptability and usability. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the module connections of an artificial intelligence-based multi-objective sports training motion evaluation and feedback system according to an embodiment of this application.

[0020] Figure 2 This is a flowchart illustrating an artificial intelligence-based multi-objective sports training motion evaluation and feedback method according to an embodiment of this application. Detailed Implementation

[0021] To make the content of this application easier to understand, the appendices in the embodiments of this application will be described below. Figure 1-2 The technical solutions in the embodiments of this application will be clearly and completely described.

[0022] like Figure 1 As shown, the AI-based multi-objective sports training motion evaluation and feedback system includes a multi-dimensional data acquisition module, a data preprocessing module, a multi-objective tracking and binding module, a motion modeling module, a fusion evaluation module, a personalized feedback module, and an adaptive standard library module.

[0023] The multi-dimensional data acquisition module is used to collect video and physiological data from multi-target motion scenes and extract skeletal keypoint data for each moving target. This module includes a camera unit and a physiological data acquisition unit.

[0024] The camera unit employs four wide-angle cameras, positioned on the top, bottom, left, and right sides of the sports field. Each camera's capture range covers one-quarter of the field, forming a panoramic video capture network to ensure seamless capture of the dynamic movements of multiple moving targets. The camera frame rate can be adjusted according to the type of movement: 120 frames per second for fast-moving scenes such as sprinting and jumping, and 30 frames per second for slow-motion scenes such as yoga, balancing data accuracy and computational efficiency.

[0025] The physiological data acquisition unit collects heart rate, blood oxygen, lactate concentration, and respiratory rate data for each moving target using wearable devices such as smart wristbands and chest straps. The wearable devices communicate with the system in real time via Bluetooth, with a data sampling frequency of once per second to ensure time synchronization between physiological and video data.

[0026] Skeletal keypoint data was extracted from the video stream using a deep learning algorithm, employing the HRNet algorithm, which can accurately identify the positions of 24 key joints in the human body, including the shoulder, elbow, hip, knee, and ankle joints. During the extraction process, the video stream was processed frame by frame, calculating the three-dimensional coordinates of each joint and deriving joint angle, joint velocity, and joint acceleration data to form a skeletal keypoint dataset for each moving target.

[0027] The data preprocessing module is used to optimize the skeletal key point data, solve the data quality problems caused by target occlusion due to changes in lighting and equipment errors during the acquisition process, and output the optimized skeletal key point data.

[0028] Lighting compensation processing addresses the issue of uneven brightness in video frames under low-light or high-light conditions. First, the average brightness of the current video frame is calculated by statistically analyzing the average intensity values ​​of all pixels in the frame. Then, based on a preset target brightness, the intensity value of each pixel in the video frame is adjusted using the formula Inewxy = Ixy × Ltarget / Lcurrent, where Ixy is the original pixel intensity value, Ltarget is the preset target brightness, and Lcurrent is the average brightness of the current frame. This processing ensures that the brightness of video frames under different lighting conditions is more consistent, improving the stability of skeletal keypoint extraction.

[0029] Occlusion completion processing addresses the issue of missing joints caused by mutual occlusion or obstruction by site facilities in moving targets, employing the Kalman filter algorithm. This algorithm establishes a joint position prediction model based on the motion trajectories of joints in historical frames. The state update equation is x̂t = Ax̂t-1 + But + ωt, where x̂t is the predicted position of the occluded joint in the current frame, A is the state transition matrix, B is the control matrix, ut is the external input, and ωt is the process noise. Simultaneously, the predicted positions are corrected by incorporating the skeletal structure features of adjacent unoccluded frames, ensuring that the completed joint positions conform to human motion patterns.

[0030] Noise filtering addresses abnormal fluctuations in skeletal keypoints caused by device jitter or data transmission errors by employing a moving average method. The average of the same keypoint position data from the current frame and the four preceding frames (five frames in total) is calculated and used as the optimized position for that keypoint in the current frame. The formula is xsmootht = (xt + xt-1 + xt-2 + xt-3 + xt-4) / 5, where xsmootht is the optimized keypoint position, and xt to xt-4 are the original keypoint positions from the previous five frames. This process eliminates transient abnormal data, ensuring the smoothness and continuity of the skeletal keypoint data.

[0031] The multi-target tracking and binding module is used to realize real-time tracking and continuous identity binding of multiple moving targets, avoid identity confusion caused by target occlusion and crossover, and ensure that the motion data and physiological data of each target can be accurately associated.

[0032] This module uses the DeepSORT algorithm as the basic tracking framework. First, it detects each moving target in the video stream and generates a target bounding box. Then, based on the target's appearance and motion features, it calculates the matching degree between targets in adjacent frames, achieving initial tracking. To improve the stability of identity binding, skeletal feature parameters are introduced as a supplement. These parameters include the ratio of shoulder width to height, the ratio of leg length to height, and the ratio of joint spacing. These parameters are unique to each individual and can effectively distinguish different moving targets.

[0033] When tracking is interrupted due to occlusion of a moving target, the system stores the target's skeletal feature parameters before the interruption. After the occlusion is removed, the system extracts the currently detected target's skeletal feature parameters and performs a similarity match with the stored feature parameters. If the similarity is greater than a preset threshold of 0.85, the target is identified as the same target, and its identity binding is restored. Simultaneously, a unique identifier is assigned to each moving target, and this identifier is associated with and stored in conjunction with skeletal keypoint data and physiological data to ensure the accuracy of data attribution during subsequent modeling, evaluation, and feedback processes.

[0034] The motion modeling module is used to establish a target constraint dynamic optimization model based on optimized skeletal keypoint data, model and predict the motion state of each moving target in real time, and provide accurate motion feature descriptions for subsequent evaluation.

[0035] The target-constrained dynamic optimization model consists of three parts: action state variables, action standard values, and action optimization functions. Action state variables quantify the motion characteristics of the moving target, including the angular velocity and acceleration of each joint. Joint angles are calculated using the three-dimensional coordinates of adjacent joints; for example, the elbow joint angle is calculated using the vector angle between the upper arm joint, forearm joint, and wrist joint. Joint velocity is the ratio of the difference in joint position between adjacent frames to the time interval; joint acceleration is the ratio of the difference in joint velocity between adjacent frames to the time interval.

[0036] The motion standard value is a preset reference standard for the corresponding motion type stored in the adaptive standard library module. For example, the standard angle of the elbow joint in a basketball shooting motion is 90 to 110 degrees, and the standard speed of the knee joint is 0.5 to 0.8 meters per second.

[0037] The motion optimization function minimizes the deviation between the actual motion state and the standard motion value of the target. The objective function is Jt = Σi=1mwi × 0.5 × ||xit - xireft||², where Jt is the deviation value, wi is the weight of the i-th joint, xit is the actual joint state, xireft is the standard joint state, and m is the number of joints. Simultaneously, the model imposes motion range constraints and motion speed constraints. The motion range constraint limits joint angles to within the physiological limits of the human body; for example, the knee flexion angle is no less than 0 degrees and no more than 170 degrees. The motion speed constraint limits joint motion speeds to no more than a preset speed threshold; for example, the hip joint movement speed is no more than 1.2 meters per second. The optimization problem is solved using the Lagrange multiplier method to obtain the optimal motion state description, providing high-quality motion data for the evaluation module.

[0038] The fusion assessment module is used to comprehensively evaluate the movement quality of each movement target. It combines movement matching analysis and physiological data correlation analysis to generate an objective and comprehensive score, avoiding the one-sidedness of single movement assessment.

[0039] The evaluation process consists of three steps: local motion decomposition and basic motion score calculation. The overall motion is decomposed into several key local motions based on the motion logic. For example, running is decomposed into four key local motions: arm swing, push-off, take-off, and landing; rope skipping is decomposed into three key local motions: rope swing, jump, and landing. Feature parameters of each key local motion are extracted, including joint angles, joint velocities, and joint accelerations. The similarity between these feature parameters and corresponding parameters in the standard motion values ​​is calculated using a cosine similarity algorithm, with the formula: similarity FiFref = Fi・Fref / ||Fi||・||Fref||, where Fi is the actual local motion feature vector and Fref is the standard local motion feature vector. A score is generated for each key local motion based on the similarity. A similarity of 1 results in a score of 100, and a similarity of 0 results in a score of 0, with intermediate values ​​linearly mapped. The scores of all key local motions are weighted to obtain the basic motion score. The weights are set according to the importance of the local motion to the overall motion. For example, in running, the weight of the push-off motion is 0.3, and the weight of the arm swing motion is 0.2.

[0040] Physiological data correlation analysis is used to determine the fit between physiological state and movement. First, the normal physiological range corresponding to each key local movement is determined. For example, the normal heart rate range for the running push-off movement is 120 to 150 beats per minute, and the normal respiratory rate range is 18 to 25 breaths per minute. When the baseline movement score reaches a preset threshold of 80 points, if the physiological data exceeds the normal physiological range (e.g., heart rate higher than 150 beats per minute or lower than 120 beats per minute), the physiological fit score is reduced; the greater the deviation, the greater the reduction. If the physiological data is within the normal physiological range, the physiological fit score is 100 points. When the baseline movement score does not reach the preset threshold of 80 points, regardless of whether the physiological data is normal, the physiological fit score remains at 80% of the baseline movement score to avoid masking physiological problems due to movement deviations.

[0041] The comprehensive score calculation combines the basic movement score and the physiological adaptation score with preset weights. These preset weights are set according to the type of exercise: for physical activities such as basketball and gymnastics, the basic movement score has a weight of 0.6, and the physiological adaptation score has a weight of 0.4; for endurance activities such as long-distance running and cycling, the basic movement score has a weight of 0.3, and the physiological adaptation score has a weight of 0.7. The calculation formula is: Comprehensive Score = Basic Movement Score × Movement Weight + Physiological Adaptation Score × Physiological Weight. The comprehensive score ranges from 0 to 100 points, with higher scores indicating better movement quality and better physiological adaptation.

[0042] The personalized feedback module generates real-time feedback information and training summary reports for each exercise target based on comprehensive scores and movement deviation analysis results. The feedback content is hierarchically precise, helping users quickly identify problems and make improvements.

[0043] Real-time feedback is generated based on a hierarchical structure of movement deviations, including joint-level, segment-level, and overall levels. Joint-level deviation feedback addresses angular and velocity deviations in individual joints, such as "Insufficient knee flexion angle; current angle is 120 degrees, standard angle is 140 to 160 degrees; it is recommended to increase the knee flexion range"; "Elbow extension speed is too fast; current speed is 1.0 m / s, standard speed is 0.6 to 0.8 m / s; it is recommended to slow down the extension rhythm." Segment-level deviation feedback addresses timing and rhythmic deviations in key local movements, such as "Delayed force exertion during the running push-off phase, 0.2 seconds later than the standard timing; it is recommended to advance the push-off point"; "Excessive speed during the shooting arm raise phase, 0.3 seconds shorter than the standard time; it is recommended to slow down the arm raise speed." Overall deviation feedback addresses coordination and rhythm deviations in the overall movement. For example, "In running, the arm swing and push-off rhythms are not coordinated; the arm swing frequency is higher than the push-off frequency. It is recommended to adjust the arm swing speed to match the push-off frequency." Similarly, "The overall rhythm of the jump rope movement is uneven, with some cycles having a time difference exceeding 0.1 seconds. It is recommended to maintain a constant jump rope speed." Real-time feedback is displayed through a graphical interface, using color coding to indicate movement status: green indicates an overall score above 90 points, yellow indicates an overall score between 60 and 90 points, and red indicates an overall score below 60 points. Text-based deviation adjustment suggestions are also displayed.

[0044] The training summary report is generated after each training session and includes three parts: a comprehensive score change curve, plotted on the horizontal axis with the comprehensive score on the vertical axis, showing the trend of the comprehensive score change for each exercise target during the training process, intuitively reflecting the fluctuation of training effect; a local movement score distribution, using a bar chart to display the score of each key local movement, marking the three local movements with the lowest scores and indicating key areas for improvement; and physiological data and movement adaptability analysis results, which statistically analyzes the number of times physiological data exceeded the normal range during training and the corresponding movement segments, analyzing the impact of physiological state on movement quality, for example, "There were 5 instances of heart rate exceeding the normal range during training, all occurring during the running push-off phase, it is recommended to reduce the force intensity of this phase to adjust the heart rate." The training summary report can be exported as a document for easy archiving and subsequent analysis.

[0045] The adaptive standard library module stores standard motion parameters for different types of sports and dynamically optimizes these parameters based on user type and historical training data, thereby improving the system's adaptability to different users and different training stages.

[0046] The standard movement parameters are stored using a categorized storage method, divided into three main categories according to the type of exercise: limb movement exercises, endurance exercises, and comprehensive exercises. Each category contains the standard movement parameters for specific sports. For example, limb movement exercises include basketball, gymnastics, and dance, while endurance exercises include long-distance running, cycling, and swimming. The standard movement parameters for each sport include the standard angle range, standard velocity range, and standard acceleration range of each joint, as well as the normal physiological range of the corresponding local movements.

[0047] Different standard motion parameters are preset based on user type. User types are divided into beginner, intermediate, and professional. The tolerance range for deviation of standard motion parameters is larger for beginners than for intermediate users. For example, the tolerance range for elbow joint angle deviation in basketball shooting motion is ±20 degrees for beginners, ±15 degrees for intermediate users, and ±10 degrees for professional users. The normal range of physiological data is also wider for beginners. For example, the normal range of heart rate for running is 100 to 160 beats per minute for beginners, 110 to 150 beats per minute for intermediate users, and 120 to 140 beats per minute for professional users. When users use the system for the first time, they select their own type, and the system automatically loads the standard motion parameters for that type.

[0048] The system periodically optimizes standard movement parameters, initiating an optimization process every 100 training data sessions. First, it analyzes the overall score distribution from historical training data, calculating the frequency of different deviation types. When a deviation, such as insufficient knee angle, occurs more than 30% of the time, the cause is analyzed. If the standard parameter is set too strictly, the tolerance range for that parameter is expanded. If it indicates a common movement defect among users, targeted basic training suggestions are generated and added to the feedback. Simultaneously, based on historical training data for different movement goals, a personalized subset of standard movement parameters is generated for each goal. For example, if a user has a limited shoulder joint range of motion, the system adjusts the standard angle range of the user's shoulder joint to better suit their physiological characteristics, improving the objectivity of the assessment and the feasibility of the feedback.

[0049] A multi-objective sports training movement evaluation and feedback method based on artificial intelligence includes the following steps: Step 1: Collect video and physiological data from multi-target motion scenes using a multi-dimensional data acquisition module, and extract skeletal key point data for each moving target; Step 2: The data preprocessing module performs illumination compensation, occlusion completion, and noise filtering on the skeletal keypoint data, and outputs optimized skeletal keypoint data. Step 3: Based on video data and optimized skeletal keypoint data, the multi-target tracking and binding module enables real-time tracking and continuous identity binding of multiple moving targets. Step 4: Based on the optimized skeletal keypoint data, the motion modeling module establishes a target constraint dynamic optimization model to model and predict the motion state of each moving target in real time. Step 5: The overall motion of each motion target is locally decomposed through the fusion evaluation module, and a comprehensive score is generated by combining motion matching analysis and physiological data correlation analysis. Step 6: Based on the comprehensive score and the level of movement deviation, the personalized feedback module generates real-time feedback information and training summary report for each movement target. Step 7: Iteratively optimize the standard action parameters based on user type and historical training data using the adaptive standard library module.

[0050] The above embodiments are only used to illustrate the technical solutions of the embodiments of this application, and are not intended to limit them. Although the embodiments of this application have been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features, without departing from the spirit and scope defined by the claims of this application.

Claims

1. A multi-objective sports training movement evaluation and feedback system based on artificial intelligence, characterized in that, include: The multi-dimensional data acquisition module is used to collect video and physiological data in multi-target motion scenes and extract the skeletal key point data of each moving target. The data preprocessing module is used to perform illumination compensation, occlusion completion, and noise filtering on the skeletal key point data, and output optimized skeletal key point data. The multi-target tracking and binding module is used to achieve real-time tracking and continuous identity binding of multiple moving targets based on video data and optimized skeletal key point data. The motion modeling module is used to build a dynamic optimization model of target constraints based on optimized skeletal keypoint data, and to model and predict the motion state of each moving target in real time. The fusion assessment module is used to perform local decomposition of the overall movement of each motion target and generate a comprehensive score by combining movement matching analysis and physiological data correlation analysis. The personalized feedback module is used to generate real-time feedback information and training summary reports for each motion target based on the comprehensive score and the level of motion deviation. The adaptive standard library module stores standard motion parameters for different motion types and iteratively optimizes these parameters based on user type and historical training data.

2. The multi-objective sports training movement evaluation and feedback system based on artificial intelligence according to claim 1, characterized in that, The multi-dimensional data acquisition module includes a camera unit and a physiological data acquisition unit; the camera unit uses four wide-angle cameras respectively arranged on the upper, lower, left and right sides of the sports field to capture panoramic video streams in real time; The physiological data acquisition unit collects heart rate, blood oxygen, lactate concentration, and respiratory rate data for each moving target through wearable devices; the skeletal key point data is extracted from the video stream using a deep learning algorithm, and the extracted skeletal key points include joint position, joint angle, joint velocity, and joint acceleration.

3. The multi-objective sports training movement evaluation and feedback system based on artificial intelligence according to claim 1, characterized in that, The data preprocessing module's processing steps include: Illumination compensation: Calculate the average brightness of the current video frame, and adjust the intensity value of each pixel in the video frame according to the preset target brightness to make the brightness of the current frame approach the preset target brightness; Occlusion completion: Based on the Kalman filter algorithm, combined with the skeletal trajectory data of moving targets in historical frames, the position data of occluded joints are predicted and completed; Noise filtering: The moving average method is used to calculate the average value of the skeletal key point positions within a preset number of adjacent frames, and the average value is used to replace the abnormally fluctuating skeletal key point position data in the current frame.

4. The multi-objective sports training movement evaluation and feedback system based on artificial intelligence according to claim 1, characterized in that, The implementation process of the multi-target tracking binding module includes: The DeepSORT algorithm is used to perform preliminary tracking of multiple moving targets in the video stream, generating a temporary tracking trajectory for each target; Extract the skeletal feature parameters for each motion target, including the ratio of shoulder width to height, the ratio of leg length to height, and the ratio of inter-joint distance. By associating temporary tracking trajectories with skeletal feature parameters, when tracking is interrupted due to occlusion of the moving target, the target identity is re-matched based on the skeletal feature parameters to achieve continuous identity binding.

5. The multi-objective sports training movement evaluation and feedback system based on artificial intelligence according to claim 1, characterized in that, The target constraint dynamic optimization model in the motion modeling module includes motion state variables, motion standard values, and motion optimization functions. Action state variables include joint angles, joint velocities, and joint accelerations of the moving target; action standard values ​​are preset reference standards for the corresponding motion type stored in the adaptive standard library module; The motion optimization function is used to minimize the deviation between the actual motion state of the moving target and the standard motion value, while applying motion range constraints and motion speed constraints; the motion range constraint limits the joint angles to the range of human physiological angle limits. Movement speed constraints limit the joint movement speed to a preset speed threshold.

6. The multi-objective sports training movement evaluation and feedback system based on artificial intelligence according to claim 1, characterized in that, The evaluation process of the fusion evaluation module includes: The overall motion of each moving target is decomposed into several local key motions, and the feature parameters of each local key motion are extracted. The feature parameters include joint angle, joint velocity and joint acceleration. Calculate the similarity between the feature parameters of each local key action and the corresponding parameters in the action standard value, generate a score for each local key action based on the similarity, and calculate the basic action score by weighting the scores of all local key actions. The correlation between physiological data and movement is analyzed. When the basic movement score reaches a preset threshold but the physiological data exceeds the normal physiological range of the corresponding movement, the physiological fit score is reduced; when the basic movement score does not reach the preset threshold but the physiological data is within the normal physiological range, the physiological fit score remains unchanged. The basic movement score and the physiological adaptation score are weighted according to preset weights to obtain the comprehensive movement score for each exercise goal. The preset weights are set according to the type of exercise: the basic movement score for limb movement exercise has a weight of 0.6 and the physiological adaptation score has a weight of 0.4, while the basic movement score for endurance exercise has a weight of 0.3 and the physiological adaptation score has a weight of 0.

7.

7. The multi-objective sports training movement evaluation and feedback system based on artificial intelligence according to claim 1, characterized in that, The feedback process of the personalized feedback module includes: Based on the comprehensive score and the results of local motion decomposition, the motion deviation level is determined. The motion deviation level includes joint level, motion segment level and overall level. Joint-level deviation feedback generates adjustment suggestions for angular and velocity deviations of individual joints; motion segment-level deviation feedback generates adjustment suggestions for rhythmic force application timing deviations of key local movements; and overall-level deviation feedback generates adjustment suggestions for coordination and rhythm deviations of the overall movement. The training summary report includes the overall score change curve for each exercise target, the distribution of local movement scores, and the results of physiological data and movement adaptability analysis.

8. The multi-objective sports training movement evaluation and feedback system based on artificial intelligence according to claim 1, characterized in that, The optimization process of the adaptive standard library module includes: Different standard action parameters are preset according to user type, which includes novice, intermediate, and professional users. The standard action parameter deviation tolerance range for novice users is larger than that for intermediate users, and the standard action parameter deviation tolerance range for intermediate users is larger than that for professional users. Regularly analyze the overall score distribution and deviation types in historical training data. When the frequency of a certain type of deviation exceeds the preset threshold, adjust the deviation tolerance range of the corresponding standard action parameters or optimize the action standard value. Based on historical training data for different sports goals, a personalized subset of standard movement parameters is generated for each sports goal, making the standard movement parameters more suitable for individual physiological characteristics.

9. A multi-objective sports training movement evaluation and feedback method based on artificial intelligence, characterized in that, The application of the AI-based multi-objective sports training motion evaluation and feedback system according to any one of claims 1 to 8 includes the following steps: Step 1: Collect video and physiological data from multi-target motion scenes using a multi-dimensional data acquisition module, and extract skeletal key point data for each moving target; Step 2: The data preprocessing module performs illumination compensation, occlusion completion, and noise filtering on the skeletal keypoint data, and outputs optimized skeletal keypoint data. Step 3: Based on video data and optimized skeletal keypoint data, the multi-target tracking and binding module enables real-time tracking and continuous identity binding of multiple moving targets. Step 4: Based on the optimized skeletal keypoint data, the motion modeling module establishes a target constraint dynamic optimization model to model and predict the motion state of each moving target in real time. Step 5: The overall motion of each motion target is locally decomposed through the fusion evaluation module, and a comprehensive score is generated by combining motion matching analysis and physiological data correlation analysis. Step 6: Based on the comprehensive score and the level of movement deviation, the personalized feedback module generates real-time feedback information and training summary report for each movement target. Step 7: Iteratively optimize the standard action parameters based on user type and historical training data using the adaptive standard library module.

Images