Artificial intelligence physical training exercise system

Abstract

The invention relates to the field of artificial intelligence and computer vision, and discloses an artificial intelligence physical training exercise system which comprises a data acquisition module, a data preprocessing module, an action modeling module, an action evaluation module and a user feedback module. The data acquisition module acquires real-time video data of a moving target and extracts skeleton key point data; the data preprocessing module performs optimization processing on the skeleton key point data; the action modeling module is used for modeling and predicting the action state of the moving target by utilizing a target constraint dynamic optimization model based on the optimized skeleton data; the action evaluation module generates action scores and optimization suggestions through local action decomposition and action matching analysis; and the user feedback module provides real-time feedback information and generates a training summary report according to the action score. The method can achieve the efficient analysis and precise evaluation of the motion of the moving target in a complex scene, provides personalized optimization guidance, and is widely suitable for physical training and teaching scenes.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence technology and computer vision technology, and in particular to an intelligent system for motion monitoring, action evaluation, and feedback in sports teaching and training scenarios. Background Art

[0002] With the development of artificial intelligence technology and computer vision technology, intelligent sports training tools have gradually become an important part of modern sports teaching. However, the existing technologies still have significant limitations in meeting the requirements of real-time monitoring and action evaluation in sports teaching. Sports training scenarios usually involve multiple participants, complex dynamic actions, and wide spatial distribution, which pose high requirements for real-time monitoring and precise evaluation systems. However, most of the existing sports training equipment focuses on personal fitness tracking or simple health monitoring functions, lacking the ability of real-time monitoring and comprehensive evaluation for multi-person motion scenarios.

[0003] The existing technical means mainly rely on wearable sensors or camera devices to record the physiological or action data of moving targets. Although these methods can provide a certain degree of action analysis, they are often limited by costs, equipment complexity, and technical reliability in multi-person scenarios. In addition, the existing video analysis-based solutions usually lack the ability to accurately model and dynamically optimize skeletal data, making it difficult to meet the dual requirements of real-time and accuracy. Especially in complex environments such as occlusion and lighting changes, the accuracy of action recognition significantly decreases. More importantly, the existing technologies lack personalized support in the action evaluation and feedback links, usually only providing simple overall scores, and unable to conduct targeted analysis and guidance on specific defects in actions, resulting in limited training effects.

[0004] In summary, the existing technologies have obvious deficiencies in multi-target real-time monitoring, precise action modeling, personalized action evaluation, and efficient feedback mechanisms. There is an urgent need for an artificial intelligence sports training system that can adapt to complex motion scenarios, analyze multi-target actions in real time, and provide accurate feedback to improve the intelligent level and effect of sports teaching and training. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technologies, the present invention provides an artificial intelligence sports training motion system, which solves the problems of insufficient real-time monitoring and precise evaluation of actions in multi-person complex motion scenarios in the existing technologies, and at the same time provides technical problems of real-time feedback and personalized training guidance.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An artificial intelligence sports training motion system, including: A data acquisition module, configured to collect video data in a multi-target motion scenario through a camera device, and extract the skeletal key point data of each moving target; A data preprocessing module for optimizing the bone key point data; An action modeling module for establishing a target constraint dynamic optimization model based on the optimized bone key point data to model and predict the action state of each motion target in real time; An action evaluation module for comparing and analyzing the action of the target with the standard action through local action decomposition and action matching analysis, and generating an action score; A user feedback module for generating real-time feedback information and a training summary report according to the action score.

[0007] Preferably, the data acquisition module includes: A camera device for capturing a panoramic video stream of the motion target in real time; A bone key point extraction unit for extracting the bone key points of the motion target in the video stream through a deep learning algorithm, where the bone key points include joint positions, angles, and other motion parameters.

[0008] Preferably, the data preprocessing module includes: A light compensation unit for adjusting the brightness parameter of the captured video frame to make the brightness of the current frame close to the preset target brightness; An occlusion completion unit for predicting and completing the position data of the occluded joint based on the Kalman filtering algorithm; A noise filtering unit for eliminating invalid bone key points in the video stream caused by data jitter or error.

[0009] Preferably, the action modeling module performs real-time modeling of the action state of the motion target through a target constraint dynamic optimization model, and the target constraint dynamic optimization model includes: An action state variable for representing the joint angle, speed, and acceleration of the motion target; An action standard value for defining the preset reference standard for each action; An action optimization function for minimizing the deviation between the actual state of the motion target and the standard value, and imposing action range constraints and speed constraints.

[0010] Preferably, the action modeling module optimizes the action state through the following constraint optimization objective function: An action range constraint for restricting the joint angle of the motion target within a preset action range; An action speed constraint for restricting the joint action speed of the motion target not to exceed a preset threshold.

[0011] Preferably, the action evaluation module analyzes the action of the motion target through the local action matching theory, and the local action matching theory includes the following steps: Decompose the overall movement into several local key movements; Extract the characteristic parameters of each local key movement, including joint angles, speeds, and accelerations; Calculate the similarity between the extracted characteristic parameters and the corresponding parameters in the standard action model, and generate a local action score based on the similarity.

[0012] Preferably, the action evaluation module calculates the overall action score based on the local action scores, and the overall action score is the weighted average of all local action scores, which is used to reflect the overall quality of the movement target's action completion.

[0013] Preferably, the user feedback module includes: A real-time feedback unit for displaying the action status of the movement target through color markings, where green indicates a correct action, yellow indicates an action deviation, and red indicates an incorrect action; A report generation unit for generating a training summary report including action scores, local action analysis results, and improvement suggestions.

[0014] Preferably, the training summary report includes: An action score curve for displaying the trend change of the movement target's action score; Action improvement suggestions for prompting the optimization direction of the movement target in terms of joint angles, speeds, or other movement characteristics; Sports performance statistical data for demonstrating the comprehensive performance of the movement target in multiple trainings.

[0015] Preferably, the system is applicable to the sports teaching scenario, can simultaneously monitor the real-time movement states of multiple movement targets, and implement differential action optimization guidance for different targets according to the generated action scores.

[0016] The present invention provides an artificial intelligence sports training system. It has the following beneficial effects: 1. Through the collaborative work of the data acquisition module and the data preprocessing module, the present invention can capture high-quality video data of multi-target movement scenarios in real time, and optimize the skeleton key point data through technologies such as light compensation, occlusion completion, and noise filtering, making it adapt to complex lighting conditions, multi-target occlusion, and dynamic environments, ensuring the accuracy and integrity of the data, and providing reliable input for subsequent modeling and evaluation.

[0017] 2. The present invention adopts a target-constrained dynamic optimization model, which can perform real-time modeling and dynamic optimization on the action states (including joint angles, velocities, and accelerations) of a moving target and a standard action model. By constraining the action range, velocity, and acceleration, this model generates a continuous and smooth description of the action state and has real-time prediction ability, providing more accurate action feature data for the evaluation module.

[0018] 3. The present invention realizes the decomposition of the overall action of a moving target and the precise matching of local actions through local action decomposition technology and action matching analysis. The system can perform quantitative scoring on different action links, identify deviations, and generate optimization suggestions simultaneously, especially suitable for personalized action analysis in complex motion scenarios (such as running, shooting, etc.), improving the accuracy and pertinence of action evaluation.

[0019] 4. The present invention realizes an action feedback mechanism that combines real-time performance and visualization through a user feedback module. The system can intuitively present the user's action state in the form of color marking, text prompts, and voice announcements, and generate a structured report including an action scoring curve, local analysis, and training trend, helping the user quickly understand the training results and continuously optimize the action.

[0020] 5. The present invention can meet the sports training needs of multiple targets and multiple scenarios, supporting both individual training and team teaching scenarios. Through iterative learning of historical training data and continuous optimization of the standard action model, the system can provide targeted services according to the action characteristics of different users, with good adaptability and scalability, and is widely applicable to various sports training and teaching scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the system architecture of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] Please refer to the attached Figure 1, The present invention provides an artificial intelligence sports training system, aiming to achieve real-time monitoring, modeling, action analysis and feedback of moving targets in multi-target sports scenarios, and is applicable to sports training and teaching scenarios. The system collects video data in the sports scenario through a camera device, and combines a bone key point extraction algorithm, a dynamic optimization model and a local action matching technology to efficiently and accurately model and evaluate the action state of the moving target, and generate real-time feedback and training reports to assist sports teaching activities.

[0024] As Figure 1 shown, the artificial intelligence sports training system may include a data acquisition module, a data preprocessing module, an action modeling module, an action evaluation module, and a user feedback module. Each module of the system will be described in detail below.

[0025] Data acquisition module In this embodiment, the data acquisition module includes a camera device and a bone key point extraction unit, which are used to obtain real-time video data in the sports scenario and extract the bone key point data of the moving target. The video data is collected through the camera device, and the bone key points are extracted from the video frames through a deep learning algorithm.

[0026] As an option, the camera device can be a wide-angle camera or a camera with panoramic coverage function to ensure the complete capture of the sports scenario. In a possible implementation, the angle of the camera can be adjusted during installation so that the acquisition range covers all targets in the sports area and avoids blind spots.

[0027] It should be noted that the video data collected by the camera device is a continuous video stream, and its frame rate can be adjusted according to the dynamics of the sports scenario. For example, the frame rate range can be from 30 frames per second to 120 frames per second to ensure the capture of fast action details of the moving target.

[0028] Specifically, the bone key point extraction unit processes the video stream collected by the camera device through a deep learning algorithm to extract the bone key point data of the moving target.

[0029] In some embodiments, the algorithm can be OpenPose or HRNet. These algorithms are based on convolutional neural networks and identify the joint positions in the video by detecting human postures.

[0030] Exemplarily, the bone key point extraction includes the following steps: 1. First, preprocess the video frames, including image scaling, grayscale conversion, and noise elimination, to reduce the interference of the image background on the algorithm recognition.

[0031] Specifically, the image scaling can adjust the resolution of each frame to 256x256 pixels to reduce the computational amount while ensuring the accuracy.

[0032] 2. Then, the preprocessed image is subjected to human pose detection by a deep learning model to generate key point predictions.

[0033] In a possible implementation, the skeletal key points include the joint positions of the moving target, such as the shoulder joint, elbow joint, knee joint, and ankle joint, etc. The position of each skeletal point can be represented by three-dimensional coordinates, i.e., p ij =(x ij , y ij , z ij ), where p ij is the position of the j-th joint point of the i-th target.

[0034] It should be noted that in video data, the extraction of skeletal points is performed frame by frame, that is, the skeletal key point detection algorithm is independently run on each frame image in the video stream. After extraction, the data of each skeletal point is organized into a set S = {s 1 , s 2 ,..., s n}, where s i ={p i1 , p i2 ,..., p im} is the set of all key points of the i-th moving target.

[0035] In a possible implementation, in order to improve the accuracy of skeletal key point extraction, a post-processing step can be added to the deep learning model, such as key point confidence filtering. Specifically, for each detected key point, a confidence score c ij is calculated. If c ij is less than the preset threshold, then this key point will be filtered out to ensure data quality.

[0036] It can be understood that this post-processing step of skeletal key point extraction not only improves the accuracy of the data but also reduces the false detection rate of the algorithm in complex backgrounds. For example, in a scenario where multiple people are moving simultaneously, this step can effectively eliminate the false detections of background people or interfering targets.

[0037] In some embodiments, the data acquisition module can also mark the moving targets in the video frame through a multi-target tracking algorithm so that subsequent modules can accurately allocate and process the skeletal point data of each target.

[0038] Specifically, the multi-target tracking algorithm can use the optical flow method or a deep learning object detection network (such as YOLOv5 or DeepSORT). These algorithms can assign a unique identifier ID to each target according to the position changes of the moving target in consecutive frames.

[0039] As an option, the tracking results of moving targets can be represented as a set of target trajectories: T = {T 1 , T 2 ,..., T n}, T i = {(t 1 , p i1 ), (t 2 , p i2 ),...,(t k , p ik )} where T i represents the trajectory of the i-th moving target, t k is the timestamp, and p ik is the skeletal point data at that moment.

[0040] It should be noted that in different motion scenarios, the data acquisition module can adjust the acquisition parameters according to specific requirements. For example: In scenarios of fast actions (such as sprinting or jumping), the frame rate and resolution of the camera device can be increased to capture the detailed actions of moving targets.

[0041] In multi-person scenarios (such as group gymnastics training), the resolution can be appropriately reduced to improve the real-time performance of data acquisition.

[0042] It can be understood that this flexible parameter setting enables the data acquisition module to adapt to a variety of application scenarios and enhances the versatility and scalability of the system.

[0043] The operation process of the data acquisition module in this embodiment includes: The camera device captures the video stream in the motion scenario in real time and transmits it to the skeletal key point extraction unit through the interface; The skeletal key point extraction unit runs the deep learning algorithm to extract the skeletal point data of all moving targets frame by frame and performs confidence filtering and post-processing on the extraction results; Finally, the sorted skeletal point set and target trajectory set are output to the data preprocessing module to provide basic data for subsequent action modeling and analysis.

[0044] Data preprocessing module In this embodiment, the data preprocessing module mainly includes three core sub-functions: light compensation, occlusion completion, and noise filtering. Each function operates independently and works in coordination to ensure the high quality of the skeletal key point data in different environments and motion scenarios.

[0045] It should be noted that the data preprocessing module operates independently on each frame of data without relying on subsequent steps, and the optimized data after processing will be directly transmitted to the action modeling module.

[0046] As an option, the function of light compensation is used to adjust the brightness of the video frames output by the data acquisition module, especially in low-light or strong-light environments, and the image brightness tends to be balanced through an adaptive brightness adjustment algorithm.

[0047] Specifically, in this embodiment, light compensation can be achieved through the following formula: where I(x, y) represents the intensity value of the pixel point in the video frame, and L current represents the average brightness of the current frame, and L target represents the preset target brightness value.

[0048] In a possible implementation, the average brightness L of the current frame can be first calculated through global statistics current , and then the intensity of all pixel points is adjusted to make it close to the target brightness L target . It should be noted that this method is not only applicable to low-light environments, but also can be used for brightness suppression in overexposed scenes.

[0049] In some embodiments, light compensation can also be combined with histogram equalization technology to further enhance the image contrast, so as to improve the robustness of the skeleton key point extraction algorithm under complex lighting conditions.

[0050] The occlusion completion function is another core function of the data preprocessing module, which is used to handle the problem of missing skeleton key points caused by partial occlusion of moving objects. Exemplarily, in this embodiment, the position of the occluded joint points is predicted and completed through a Kalman filter.

[0051] In a possible implementation, the state update equation of the Kalman filter can be expressed as: where represents the predicted position of the occluded joint points in the current frame, A is the state transition matrix, B is the control matrix, and u t is the external input, and ω t is the process noise.

[0052] It can be understood that the Kalman filter predicts the position of the missing key points by combining the skeleton point trajectories of historical frames and partial skeleton point data of the current frame, and generates continuous skeleton data.

[0053] In some embodiments, in order to further improve the accuracy of occlusion completion, the predicted value Compare with the actual detected value, calculate the confidence score, and select whether to use the predicted value based on the confidence. If the confidence is lower than the set threshold, an approximate value can be generated through an interpolation algorithm.

[0054] It should be noted that the occlusion completion function is particularly applicable to scenarios of multiple people exercising, such as team sports training or multiplayer games, to solve the problem of missing skeletal points caused by mutual occlusion between targets.

[0055] The noise filtering function is used to eliminate outliers or instantaneous fluctuations in the skeletal key point data to ensure smooth and stable data. Specifically, in this embodiment, the moving average method is used to smooth the skeletal point data.

[0056] Exemplarily, the calculation formula of the moving average method is: where, x smootg,t represents the position of the smoothed skeletal point, n is the window size, and x i represents the position of the skeletal point in the i-th frame.

[0057] It can be understood that the moving average method effectively removes the instantaneous fluctuations caused by video jitter or data acquisition errors by calculating the average value of the positions of skeletal points within several adjacent frames.

[0058] In some embodiments, the noise filtering function can also be combined with the median filtering method to handle outliers with large deviations. For example, when the position of a joint point in a certain frame suddenly deviates from the normal trajectory range, the median filtering can replace the outlier with the median value within the window to avoid affecting the continuity of the data.

[0059] It should be noted that the parameters of the noise filtering function (such as the window size) can be adjusted according to the specific exercise scenario. For example, in a high-dynamic scenario (such as sprinting or jumping), a smaller window value can be selected to retain the fast-changing characteristics of the movement; while in a low-dynamic scenario (such as static standing posture training), a larger window value can be selected to further enhance the smoothing effect.

[0060] The operation process of the data preprocessing module in this embodiment is as follows: First, perform illumination compensation on the video frame data to adjust the image brightness to the target range; Then, use the Kalman filter to complete the occlusion of the missing skeletal key points and generate a complete set of skeletal points; Finally, perform noise filtering on the skeletal point data through the moving average method and smoothing filtering, and output the optimized set of skeletal points.

[0061] It should be noted that the three sub - functions of light compensation, occlusion completion, and noise filtering can run in parallel. The modular design ensures that each function can independently optimize the skeletal point data.

[0062] It can be understood that the data pre - processing module in this embodiment can significantly improve the quality of the skeletal key - point data, providing reliable basic data for the subsequent action modeling module. Especially in complex motion scenarios (such as multiple people, low light, occlusion, etc.), the data pre - processing module, through its functions of light compensation, occlusion completion, and noise filtering, ensures the accuracy and continuity of the skeletal data, thereby improving the stability and adaptability of the entire system.

[0063] Action modeling module In this embodiment, the function of the action modeling module is to model and real - time predict the action state of each moving target by establishing a target - constrained dynamic optimization model based on the optimized skeletal key - point data output by the data pre - processing module. This module is the core part of the entire system, used to describe the dynamic behavior of the moving target and provide high - quality modeling data for the subsequent action evaluation module.

[0064] In this embodiment, the main technical solutions of the action modeling module include action state modeling, target - constrained dynamic optimization, and action prediction. The input of action modeling is the set of skeletal points S output by the data pre - processing module clean , and the output is the optimized action state X * (t), which describes the motion characteristics of each moving target at time t.

[0065] As an option, the action state modeling adopts a parametric description method, representing the state variables of the moving target as a set of joint angles, joint velocities, and joint accelerations, covering the key physical characteristics of the action.

[0066] Specifically, in this embodiment, the action state modeling of the moving target can be expressed in the following form: X(t) = {x 1 (t), x 2 (t),..., x m (t)} where X(t) represents the overall state of the moving target at time t, and x i (t) is the state variable of the i - th joint, including the following: x i (t): the current position or angle of the joint; the velocity of the joint; the acceleration of the joint.

[0067] It should be noted that the joint angle is calculated from the three-dimensional positions of the skeletal key points. For example, the angle of the shoulder joint can be represented by the included angle between the direction vectors of the upper arm and the torso.

[0068] In a possible implementation, the motion modeling module optimizes the state of the motion target through the target constraint dynamic optimization model. The core idea of the target constraint dynamic optimization model is to minimize the deviation between the actual motion state X(t) and the standard motion state X ref (t), while satisfying a series of physical constraint conditions.

[0069] Exemplarily, the optimization objective function can be expressed as: Where: represents the deviation between the actual state and the standard state; w i is the weight parameter, which is used to represent the influence degree of different joints on the overall motion.

[0070] It can be understood that by adjusting the weight parameter w i , the importance of different joints can be differentiated. For example, in the running motion, the weights of the lower limb joints can be set higher, while the weights of the upper limb joints can be relatively lower.

[0071] As an option, the target constraint dynamic optimization model can also incorporate multiple physical constraint conditions to ensure that the generated motion state conforms to the physiological limits of the human body. Specifically, these constraint conditions include but are not limited to: Spatial constraint: used to limit the joint angle within a reasonable range to prevent abnormal motions that exceed the normal human motion range.

[0072] Exemplarily, the spatial constraint can be expressed as: x min,i ≤x i (t)≤x max,i where, x min,i and x max,i respectively represent the minimum and maximum allowable angles of the i-th joint.

[0073] Velocity constraint: used to limit the range of change of the joint velocity to ensure the smoothness and fluency of the motion.

[0074] Specifically, the velocity constraint can be expressed as: where, v max is the preset velocity threshold.

[0075] Acceleration constraint: used to limit the magnitude of joint acceleration and prevent actions from being too drastic.

[0076] Exemplarily, the acceleration constraint can be expressed as: where a max is a preset acceleration threshold.

[0077] In a possible implementation, the target constraint dynamic optimization problem is solved by the Lagrange multiplier method. The expression of the Lagrangian function is: where: J(t) is the optimization objective function; g j (X) represents the j-th constraint condition; λ j is the corresponding Lagrange multiplier.

[0078] It can be understood that the Lagrange multiplier method can integrate the constraint conditions into the optimization objective, solve it through the gradient descent method, and finally obtain the optimal action state X*(t).

[0079] It should be noted that in order to improve the efficiency of optimization calculation, sparse coding technology is adopted in this embodiment to reduce the dimensionality of the skeletal key point data. In some embodiments, sparse coding can compress the high-dimensional skeletal point data into low-dimensional feature vectors while retaining key motion information, thereby significantly reducing the computational complexity of the optimization problem.

[0080] The operation process of the action modeling module in this embodiment is as follows: First, receive the set of skeletal points S output by the data preprocessing module clean ; Then, calculate the state variables X(t) of each joint according to the skeletal point data, including joint angle, speed, and acceleration; Next, optimize the action state through the target constraint dynamic optimization model, minimize the deviation between the actual state and the standard state, and at the same time satisfy the constraint conditions such as space, speed, and acceleration; Finally, output the optimized action state X*(t) and transfer it to the action evaluation module.

[0081] It can be understood that the action modeling module can not only accurately describe the current action state of the motion target, but also predict its subsequent action trajectory through dynamic optimization, providing continuous and reliable data support for the action evaluation module. Especially in complex motion scenarios (such as multi-person motion or rapidly changing actions), this module effectively improves the robustness and adaptability of the system by introducing the target constraint dynamic optimization model.

[0082] Motion evaluation module In this embodiment, the function of the motion evaluation module is to compare and analyze the motion of the moving target with the standard motion by receiving the optimized motion state data X*(t) output by the motion modeling module, and generate a motion score and optimization suggestions by using local motion decomposition and motion matching analysis.

[0083] It should be noted that the motion evaluation module is the core module for realizing the evaluation of the quality of motion actions in the system of the present invention. Relying on dynamic analysis and mathematical models, it can objectively quantify the completion quality of each action and provide targeted improvement directions.

[0084] In this embodiment, the technical solution of the motion evaluation module includes local motion decomposition, motion matching analysis, and score generation. The module input is the optimized motion state X*(t) output by the motion modeling module and the preset standard motion model X ref (t). The outputs include local motion scores, overall motion scores, and optimization suggestions.

[0085] As an option, the motion evaluation module decomposes the overall motion of the moving target into several local key motions, and generates an overall motion score after performing matching analysis on each local motion one by one.

[0086] Specifically, in this embodiment, the implementation method of local motion decomposition is as follows: In a possible implementation manner, the overall motion A of the moving target is decomposed into several local motions A = {A 1 , A 2 ,..., A n}, where each local motion A i includes multiple joint motions.

[0087] Exemplarily, for a shooting motion A, it can be decomposed into the following local motions: A 1 : Takeoff; A 2 : Raise the hand; A 3 : Shoot.

[0088] The characteristics of each local motion A i are described by the joint state data x i (t), joint velocity and joint acceleration output by the motion modeling module.

[0089] It should be noted that the specific scheme of local motion decomposition depends on the standard motion model X ref(t) is designed, for example, based on a predefined action template or an action decomposition structure annotated by an expert.

[0090] In this embodiment, the implementation of action matching analysis is based on the cosine similarity algorithm. By comparing the local action feature F of the moving target and the standard action feature F ref the similarity, the quality of action completion is evaluated.

[0091] Specifically, the similarity calculation formula for local action matching is: Where: F represents the local action feature vector of the moving target; F ref represents the feature vector of the corresponding standard action; F·F ref is the dot product of the vectors, and ∥F∥·∥F ref ∥ are the magnitudes of the two vectors respectively.

[0092] It can be understood that the value range of the cosine similarity is [-1, 1], and the closer the value is to 1, the higher the matching degree between the local action and the standard action.

[0093] In some embodiments, in order to enhance the flexibility of action matching analysis, a weighting factor can be introduced in the similarity calculation to assign higher weights to the action features of specific joints. For example, for the running action, the features of the lower limb joints (such as the knee joint and the ankle joint) can be weighted to highlight the influence of these key joints on action completion.

[0094] As an option, in this embodiment, the scoring generation function calculates the overall action score M i by integrating the local action scores M to quantify the overall action quality of the moving target.

[0095] Specifically, the calculation formula for the local action score M i is: M i = similarity(F i , F i,ref ) Where: F i and F i,ref are the feature vector of the i-th local action and its standard feature vector respectively; The calculation method of similarity is the same as above.

[0096] The overall action score M is the weighted average of all local action scores, and the calculation formula is: Where: w i is the weight of the i-th local action; n is the total number of local actions.

[0097] It can be understood that the overall action score M provides a global evaluation of the action completion quality and can reflect the overall performance level of the motion target.

[0098] In a possible implementation, the action evaluation module can also generate optimization suggestions based on the deviation of the local action scores. The optimization suggestions include the joint actions that need to be improved and their specific adjustment directions. For example: "The shoulder joint is not fully extended. It is recommended to raise the shoulder joint angle to 90 degrees"; "The knee joint is not bent enough. It is recommended to increase the knee bending amplitude."

[0099] Exemplarily, the optimization suggestions are generated by analyzing the difference between the action state X*(t) of the motion target and the standard action state X ref (t). For example, when the state of a certain joint deviates from the standard value by more than a preset threshold, the system will mark this joint as an object to be optimized.

[0100] The operation process of the action evaluation module in this embodiment is as follows: First, receive the optimized action state X*(t) and the standard action model X ref (t) output by the action modeling module; Then, decompose the overall action A into local parts and extract the feature vectors of each local action; Next, use the cosine similarity algorithm to perform matching analysis on the feature vectors of each local action and the standard action feature vectors to generate local action scores; Finally, integrate the local action scores to calculate the overall action score and generate optimization suggestions according to the deviation.

[0101] It can be understood that the action evaluation module in this embodiment can accurately quantify the action quality of the motion target and provide targeted optimization suggestions by combining local action decomposition and action matching analysis, significantly improving the efficiency and accuracy of action training. Especially in complex motion scenarios (such as multi-person motion or non-standard actions), the action evaluation module can flexibly adapt to different action modes, providing important support for the application scalability of the system.

[0102] User feedback module In this embodiment, the function of the user feedback module is to receive the action scores, local action analysis results, and optimization suggestions output by the action evaluation module, and real-time feedback the evaluation information to the user in various forms. At the same time, it generates a complete training summary report to assist the user in improving the action and tracking the training effect.

[0103] It should be noted that the user feedback module is a key module for interaction and output in the system of the present invention. Its design ensures real-time, visualization, and personalization features, and can meet the needs of users for the effects of action training in different scenarios.

[0104] In this embodiment, the technical solution of the user feedback module includes two major functions: real-time feedback and generation of training summary reports. The module input is the action score M generated by the action evaluation module, the local action score M i and optimization suggestions. The output is real-time feedback information and a structured training report.

[0105] As an option, the user feedback module can display the action status and improvement suggestions of the motion target in real time through a graphical interface, and provide a comprehensive analysis result of the training in the form of a report.

[0106] Specifically, in this embodiment, the implementation method of the real-time feedback function is as follows: In a possible implementation manner, the real-time feedback function visually displays the action status of the motion target by means of color marking. For example: Green indicates that the action completion quality is relatively high, and the similarity to the standard action exceeds 90%; Yellow indicates that there are slight deviations in the action, and the similarity is between 60% and 90%; Red indicates that the action deviation is large, and the similarity is less than 60%.

[0107] Exemplarily, for the running action, the system can mark the overall action status and local action status of the target respectively according to the status scores of the knee joint and hip joint.

[0108] It should be noted that the thresholds for color marking can be adjusted according to different training scenarios. For example, in the training scenario for beginners, the ranges of yellow and green can be appropriately expanded to encourage users to gradually optimize their actions.

[0109] As an option, the real-time feedback function can also provide specific optimization suggestions in the form of text prompts. The text content is based on the analysis results generated by the action evaluation module. For example: "The knee joint is not fully bent. It is recommended to increase the bending angle to 45°"; "The actions of the shoulder joints are uncoordinated. It is recommended to adjust the speed of the shoulder joints to match the lower limb actions." It can be understood that such text prompts can help users quickly understand the action elements that need to be improved. Especially in complex actions (such as dance or martial arts actions), step-by-step prompts can significantly improve the optimizability of actions.

[0110] In this embodiment, the implementation method of the training summary report generation function is as follows: In one possible implementation, the training summary report presents the action scores, training trends, and optimization suggestions of the sports target in a structured manner, including the following: Action score curve: shows the changing trend of the overall action score M for each training. For example, you can draw a line graph of the score over time to intuitively reflect the user's progress.

[0111] Local action analysis: List the score M of each local action in detail i The deviations are marked, and the key movements that need to be improved are marked. For example, "the knee angle is too small during take-off, with a score of 65%; it is recommended to increase the flexion range of the knee joint."

[0112] Comparative Training Analysis: In some embodiments, the report can compare the scores of the current training with the scores of historical training, marking the points of improvement and stability. For example, "Compared to the last training, the shoulder joint movement score increased by 10%, but the knee joint score decreased by 5%." Understandably, this structured report can help users understand the training effect from both a global and local level, especially in long-term training, users can track their progress based on the report.

[0113] Exemplarily, the training summary report may also include a visual display of the data. For example: Use heat maps to show the score distribution of different joint actions. Darker colors indicate lower scores, prompting users to focus on improvement. Use pie charts or bar charts to display the components of the overall score, such as marking the weight of each local action score.

[0114] It should be noted that the specific form of the visualization can be adjusted according to the design of the user interface, such as using a simple line chart on a mobile device and providing a more complex interactive chart on a desktop device.

[0115] In one possible implementation, the user feedback module can also support real-time voice prompts. For example, for dynamic training scenarios (such as running or jumping), the system can use voice announcements to prompt the user's action status and improvement suggestions during exercise. Exemplary voice prompts include: "The knee is not bent enough. It is recommended to bend it more." "The shoulder movements are slow and not coordinated with the lower limbs." It should be noted that the content of the voice prompt can be personalized according to different training scenarios. For example, in team training, the voice prompt can include the average score of the overall action and key optimization suggestions for each member.

[0116] The operation process of the user feedback module in this embodiment is as follows: First, the action evaluation module generates an action score M, a local action score M i and optimization suggestions; Then, through the real-time feedback function, the scores and suggestions are presented to the user in the form of color marking, text prompts, or voice announcements; Next, the structured training summary report function integrates all training data and generates a comprehensive report including an action score curve, local analysis, and training comparison; Finally, the report is presented to the user in graphical and textual forms through the user interface, and at the same time, it supports being exported as a file for the user to save and refer to.

[0117] It can be understood that the user feedback module in this embodiment provides real-time and long-term training support for the user through diverse feedback methods. Especially in multi-person training or team teaching scenarios, through the score distribution and training comparison functions, it can help coaches and users discover common problems and individual differences in training, thereby optimizing the training plan.

[0118] Generally speaking, the present invention collects video data in a multi-target motion scenario, extracts the skeletal key point data of the motion target, optimizes it, and then based on the target constraint dynamic optimization model, realizes real-time modeling and prediction of the action state. Then, through local action decomposition and action matching analysis, it generates action scores and optimization suggestions, and intuitively presents the motion performance in the form of real-time feedback and training summary reports, which is widely applicable to the personalized and precise action guidance requirements in sports training and teaching scenarios.

[0119] The artificial intelligence sports training motion system of the present invention mainly completes real-time monitoring, action modeling, evaluation, and feedback of the motion target through the following steps, and can efficiently complete sports training tasks. The following is the complete workflow of the system: 1. Video data collection The system first performs real-time video collection of the motion scenario through the camera device in the data collection module.

[0120] The camera device is set at the best position in the venue to ensure complete coverage of the target motion area and avoid dead angles, and to adapt to the synchronous collection requirements in multi-target scenarios.

[0121] The video data is transmitted to the skeletal key point extraction unit through the interface, and the deep learning algorithm is started to extract the skeletal key point data corresponding to each frame from the video stream, forming a set of skeletal point data.

[0122] The specific position data of each skeletal point is represented by three-dimensional coordinates, and a unique motion target identifier is assigned for subsequent target tracking and processing.

[0123] 2. Optimization of skeletal point data The skeletal point data is transmitted to the data preprocessing module, which performs illumination compensation, occlusion completion, and noise filtering on the initial data to improve the accuracy and integrity of the data.

[0124] Illumination compensation adjusts the brightness and contrast of the video frames to adapt the data quality to the motion scenes under complex illumination conditions.

[0125] Occlusion completion predicts the positions of the occluded key points through historical frame data and supplements the missing data to ensure the continuity of the skeletal point trajectories of the moving targets.

[0126] Noise filtering further eliminates the abnormal skeletal point data caused by motion jitter or acquisition device errors, providing reliable input for subsequent modeling.

[0127] 3. Action state modeling The optimized skeletal point data is transmitted to the action modeling module. The system first extracts the action states of each moving target based on the skeletal point data, including motion parameters such as joint angles, velocities, and accelerations.

[0128] The action modeling module combines the standard action model to establish a target-constrained dynamic optimization model, calculates the state parameters of each joint, and optimizes and adjusts them to ensure that the action state is as close as possible to the requirements of the standard action.

[0129] During the operation of the module, the state trajectory of the moving target at the next moment is predicted based on its real-time action characteristics, generating continuous and smooth action description data to support subsequent action evaluation.

[0130] 4. Action quality assessment The action state data generated by the action modeling module is transmitted to the action evaluation module. The evaluation module disassembles the overall action into several local key actions through local action decomposition technology.

[0131] The characteristic parameters (such as angles and velocities) of each local action are matched with the corresponding parameters in the standard action model to calculate the completion quality of each local action.

[0132] The action evaluation module performs weighted integration on the scoring results of the local actions to generate the overall action score of the moving target, and records the action links that need to be optimized and the corresponding suggestions.

[0133] 5. Real-time feedback and training report generation The evaluation results are transmitted to the user feedback module, which first presents the real-time feedback information to the user through the user interface.

[0134] The system uses color marking, text prompts, or voice announcements to prompt the action completion status and optimization suggestions of the moving target.

[0135] Meanwhile, the user feedback module generates a training summary report that includes an overall action score curve, local action analysis, and training trend comparison. The report presents the user's training effects in a graphical and structured manner, facilitating user understanding and improvement.

[0136] Users can view the feedback information in real time through the interface or choose to export the training report for subsequent tracking analysis and recording.

[0137] 6. Data Storage and Iterative Analysis The system stores all training data in the background database for subsequent analysis and model optimization.

[0138] Through cumulative learning of historical training data, the system gradually optimizes the standard action model, improves the adaptability of the action modeling and evaluation module, and meets the requirements of different training scenarios and individual differences.

[0139] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Artificial intelligence sports training system, characterized by: include: The data acquisition module is used to collect video data in a multi-target motion scene through a camera device and extract the skeleton key point data of each moving target; A data preprocessing module, used for optimizing the skeleton key point data; The motion modeling module is used to establish a target constraint dynamic optimization model based on the optimized skeleton key point data, and to model and predict the motion state of each moving target in real time; The action evaluation module is used to compare and analyze the target's action with the standard action through local action decomposition and action matching analysis, and generate an action score; The user feedback module is used to generate real-time feedback information and a training summary report according to the action scores.

2. The artificial intelligence sports training system according to claim 1, characterized in that: The data acquisition module comprises: A camera device, used to capture a panoramic video stream of a moving target in real time; The skeleton key point extraction unit is used to extract the skeleton key points of the moving target in the video stream through a deep learning algorithm. The skeleton key points include joint positions, angles and other motion parameters.

3. The artificial intelligence sports training system according to claim 1, characterized in that: The data preprocessing module comprises: An illumination compensation unit, used to adjust the brightness parameters of the captured video frame so that the brightness of the current frame is close to the preset target brightness; The occlusion completion unit is used to predict the position data of the occluded joints and complete them based on the Kalman filter algorithm; The noise filtering unit is used to eliminate invalid skeleton key points in the video stream caused by data jitter or error.

4. The artificial intelligence sports training system according to claim 1, characterized in that: The action modeling module performs real-time modeling of the action state of the moving target through a target constraint dynamic optimization model, and the target constraint dynamic optimization model includes: Action state variables are used to represent the joint angles, velocities, and accelerations of the moving target; Action standard value, used to define the preset reference standard for each action; The motion optimization function is used to minimize the deviation between the actual state of the moving target and the standard value, and impose motion range constraints and speed constraints.

5. The artificial intelligence sports training system according to claim 4, characterized in that: The action modeling module optimizes the action state by optimizing the objective function through the following constraints: Motion range constraint: used to limit the joint angle of the moving target within the preset motion range; Action speed constraint: used to limit the joint action speed of the moving target to not exceed the preset threshold.

6. The artificial intelligence sports training system according to claim 1, characterized in that: The motion evaluation module analyzes the motion of the moving target through the local motion matching theory, and the local motion matching theory includes the following steps: Decompose the overall action into several local key actions; Extract the characteristic parameters of each local key action, including joint angle, velocity and acceleration; The similarity between the extracted feature parameters and the corresponding parameters in the standard action model is calculated, and a local action score is generated based on the similarity.

7. The artificial intelligence sports training system according to claim 6, characterized in that: The action evaluation module calculates the overall action score according to the local action scores. The overall action score is a weighted average of all local action scores and is used to reflect the overall quality of the movement target action.

8. The artificial intelligence sports training system according to claim 1, characterized in that: The user feedback module comprises: A real-time feedback unit is used to display the motion status of the moving target through color markings, where green indicates a correct motion, yellow indicates a deviation, and red indicates an incorrect motion; The report generation unit is used to generate a training summary report including action scores, local action analysis results and improvement suggestions.

9. The artificial intelligence sports training system according to claim 1, characterized in that: The training summary report includes: Action score curve, used to show the trend change of the action score of the sports target; Action improvement suggestions, which are used to indicate the optimization direction of the movement target in terms of joint angle, speed or other movement characteristics; Performance statistics that show how your sport goals stack up over multiple sessions.

10. The artificial intelligence sports training system according to any one of claims 1 to 9, characterized in that: The system is suitable for sports teaching scenarios, and can monitor the real-time motion status of multiple sports targets at the same time, and implement differentiated motion optimization guidance for different targets based on the generated motion scores.

Images