An artificial intelligence-based sports teaching evaluation system

Abstract

The application relates to the technical field of physical education, and discloses a sports teaching evaluation system based on artificial intelligence, which comprises a data acquisition module, a multi-target detection module, a key point modeling module, an action trajectory analysis module, a multi-modal fusion module and a dynamic feedback module. The data acquisition module receives video streams of a camera device and motion physiological data of a wearable sensor; the multi-target detection module positions targets based on a YOLOv8 algorithm; the key point modeling module constructs a three-dimensional action posture model by using an OpenPose algorithm; the action trajectory analysis module evaluates action standardization; the multi-modal fusion module integrates data to construct a multi-dimensional feature matrix; and the dynamic feedback module generates real-time scores and individualized correction instructions through an optimized LSTM network. In addition, the system also comprises a model self-adaption module, an abnormal action recognition module and a distributed computing module. The system realizes accurate evaluation of sports teaching, guarantees sports safety, promotes individualized teaching, and effectively improves the quality of sports teaching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of physical education, in particular to a physical education evaluation system based on artificial intelligence. BACKGROUND

[0002] In traditional physical education, there are many limitations in teaching evaluation methods, which are difficult to meet the needs of modern physical education. From the perspective of teaching quality evaluation, in the past, teachers mainly relied on subjective observation and experience to evaluate students' movement performance. In the teaching of gymnastics, martial arts and other projects, teachers need to pay attention to multiple students at the same time, and it is difficult to accurately capture the subtle deviations of each student's movement. Taking aerobics as an example, the differences in the stretching angle of students' arms and the amplitude of body twisting, teachers may miss due to observation angle or attention dispersion, which leads to the inability to accurately assess students' mastery of movement specifications, and restricts the improvement of teaching quality.

[0003] In terms of sports safety, there is a lack of effective real-time monitoring mechanism. In some competitive sports projects, such as basketball, football matches or training, students are prone to injury risks due to excessive exercise intensity, improper movements, etc. Due to the inability to obtain students' exercise physiological data and movement state in real time, teachers cannot timely discover and stop dangerous movements, and it is difficult to ensure the safety of students during exercise. For example, if students are excessively tired in basketball games, leading to high heart rate, if not monitored in time, it may cause more serious health problems.

[0004] From the perspective of teaching individualization, under the traditional teaching mode, it is difficult for teachers to provide individualized teaching guidance for each student's characteristics. Each student has different physical fitness, sports talent and learning ability, but teachers usually adopt unified teaching standards and methods, which cannot meet the differentiated needs of students. For example, in track and field teaching, it is difficult to give targeted training plans for students with strong explosive power but poor endurance, and students with good endurance but lack of explosive power.

[0005] With the rapid development of artificial intelligence technology, its application in the field of education has gradually deepened. The progress of computer vision technology, such as continuous optimization of target detection and posture recognition algorithms, provides the possibility for accurate capture of human movement; the innovation of sensor technology makes it possible to obtain high-precision exercise physiological data. However, in the field of physical education evaluation, how to organically integrate these advanced technologies to form a complete, efficient and intelligent evaluation system still faces many challenges. The present application is based on such background, and is committed to filling this technical gap and promoting the development of intelligent and precise physical education evaluation. SUMMARY

[0006] The present application aims to provide a physical education evaluation system based on artificial intelligence to solve the problems raised in the background art.

[0007] To achieve the above object, the present application provides the following technical scheme: a sports teaching evaluation system based on artificial intelligence, comprising:

[0008] A data acquisition module for receiving real-time video stream from a camera device and motion physiological data from a wearable sensor;

[0009] A multi-target detection module for target positioning of teachers, students, exercise equipment and balls in the real-time video stream based on YOLOv8 algorithm, and outputting detection frame coordinates and target categories;

[0010] A key point modeling module for extracting skeletal key points of the detected target human body using OpenPose algorithm to generate a three-dimensional action posture model;

[0011] An action trajectory analysis module for calculating the Euclidean distance deviation of the motion trajectory from the preset standard action based on the three-dimensional action posture model, and generating an action standard evaluation result;

[0012] A multi-modal fusion module for time sequence alignment of the motion physiological data and the action standard evaluation result, and construction of a multi-dimensional feature matrix;

[0013] A dynamic feedback module for generating real-time scores and individualized action correction instructions through an LSTM network optimized by transfer learning based on the multi-dimensional feature matrix.

[0014] Preferably, the execution steps of the multi-target detection module include:

[0015] Inputting the real-time video stream into a pre-trained YOLOv8 model and extracting multi-scale feature maps through a convolutional neural network;

[0016] Performing anchor frame matching and non-maximum suppression processing on the feature maps to output target detection results.

[0017] Preferably, the execution steps of the key point modeling module include:

[0018] Performing local image cropping on the detected human target and inputting it into the OpenPose network for heat map prediction;

[0019] Extracting two-dimensional coordinates of skeletal key points through a Gaussian filtering and peak detection algorithm;

[0020] Mapping the two-dimensional coordinates to a three-dimensional space in combination with depth information of a binocular camera to generate an action posture sequence with a timestamp.

[0021] Preferably, the execution steps of the action trajectory analysis module include:

[0022] The three-dimensional motion posture model is matched with a standard motion template by dynamic time warping, and joint angle deviation is calculated;

[0023] The motion trajectory noise is filtered based on a Laplace smoothing algorithm;

[0024] The trajectory deviation of the continuous motion frame is integrated, and a motion continuity score is output;

[0025] The Euclidean distance deviation calculation formula is:

[0026]

[0027] Where (x i ,y i ,z i ) represents the three-dimensional coordinates of the i-th joint, (x std ,y std ,z std ) represents the corresponding coordinates of the standard motion, and n is the total number of joints.

[0028] Preferably, the calculation formula of the dynamic time warping matching is:

[0029]

[0030] Where A=(a1,…,a m ) and B=(b1,…,b n ) are the joint angle sequences of the actual motion and the standard motion, respectively, P is the set of all warping paths, (p,q) is a pair of matching points in the path π, satisfying p∈A,q∈B, d(a p ,b q ) is the difference measure function of the two angle values.

[0031] Preferably, the execution steps of the multi-modal fusion module include:

[0032] The heart rate, acceleration and angular velocity in the motion physiological data are wavelet transformed and denoised;

[0033] The denoised physiological data and the motion standard evaluation result are feature-level fused to generate a fusion feature vector containing spatio-temporal correlation;

[0034] The weight ratio of visual data and sensor data is dynamically allocated through an attention mechanism.

[0035] Preferably, the execution steps of the dynamic feedback module include:

[0036] The LSTM network is trained in stages using a curriculum learning strategy, with the convolutional layer parameters frozen in the first stage and the fully connected layer optimized jointly in the second stage;

[0037] The multi-dimensional feature matrix is segmented by a sliding window, and the LSTM network is input to predict the action score of the next 3 frames;

[0038] According to the prediction result, an instruction set containing joint angle correction amount and movement rhythm adjustment suggestion is generated;

[0039] The loss function used in the LSTM training is:

[0040]

[0041] Wherein, is the total loss value of model training, is the predicted output at time t, y t is the true value, T is the time step, θ is the model parameter, and λ is the regularization coefficient.

[0042] Preferably, the system further comprises:

[0043] A model adaptive module is configured to dynamically adjust the confidence threshold of the YOLOv8 model according to real-time detection results.

[0044] The execution steps of the model adaptive module include:

[0045] Statistically determine the false positive rate and the missed detection rate of target detection in the current scene;

[0046] When the false positive rate exceeds a first threshold, an exponential decay algorithm is used to reduce the confidence threshold;

[0047] When the missed detection rate exceeds a second threshold, the model weights are updated locally by gradient backpropagation;

[0048] The formula of the exponential decay algorithm is:

[0049] c t = c0·e -kt

[0050] Wherein, c t is the confidence threshold at time t, c0 is the initial threshold, k is the decay rate coefficient, and t is the iteration number.

[0051] Preferably, the multi-target detection module further comprises:

[0052] An abnormal action recognition submodule is configured to detect the abnormal trajectory of the exercise equipment.

[0053] The execution steps of the submodule include:

[0054] The trajectory of the ball game is fitted by a polynomial, and the fitting residual is calculated.

[0055] When the residual exceeds a preset threshold, triggering the high-speed camera device to perform 1000fps slow motion playback analysis;

[0056] The polynomial fitting residual calculation formula is:

[0057]

[0058] wherein, is the trajectory fitting residual, p i is the i-th trajectory point coordinate, t i is the timestamp, a j is the polynomial coefficient, d is the polynomial degree, and N is the total number of trajectory points.

[0059] Preferably, the system further comprises:

[0060] A distributed computing module is used to realize parallel data processing of multiple cameras; the execution steps of the distributed computing module include:

[0061] Apache Kafka is used to build a message queue to synchronize timestamps of multiple video streams;

[0062] Dask framework is used to distribute detection tasks to a GPU cluster to realize sub-second delay real-time feedback.

[0063] Compared with the prior art, the present application has the following advantages:

[0064] In terms of precise teaching evaluation, the system obtains real-time video streams and motion physiological data through the data acquisition module, locates teachers and students, sports equipment and balls based on the YOLOv8 algorithm through the multi-target detection module, constructs a three-dimensional action posture model through the OpenPose algorithm through the key point modeling module, and calculates the Euclidean distance deviation from the standard action through the action trajectory analysis module to generate an action standard evaluation result. The use of this series of technical means makes the evaluation of student actions more accurate and objective. Taking swimming teaching as an example, the system can accurately analyze the amplitude and frequency of the student's arm stroke, the force and angle of the leg kick, and other details, and after comparison with the standard action, a detailed and accurate evaluation report is obtained, helping teachers to fully understand the learning situation of students and providing a strong basis for teaching improvement.

[0065] From the perspective of improving sports safety, the multi-modal fusion module combines motion physiological data and action standard evaluation results, and if the student's exercise intensity is too high, causing abnormal increase of heart rate, and the action is deformed, the dynamic feedback module will timely issue a warning. Teachers can adjust the teaching arrangement in a timely manner to avoid students from being injured due to excessive fatigue or incorrect actions. For example, in long-distance running training, the system monitors the heart rate and running posture of students in real time, and once an abnormality is found, the teacher can quickly make the student stop exercising or adjust the running method to protect the health of the students.

[0066] In terms of personalized teaching, the system generates personalized action correction instructions through the dynamic feedback module based on the action assessment results and physiological data of each student. For students with different physical fitness and learning progress, targeted training recommendations are provided. For example, in basketball shooting teaching, for students with insufficient arm strength, the system recommends strengthening arm strength training and adjusting the power mode during shooting; for students with poor coordination, a special coordination training program is given to meet the differentiated learning needs of students and improve learning effectiveness.

[0067] In terms of improving teaching efficiency, the distributed computing module uses Apache Kafka and Dask framework to realize multi-camera parallel data processing, achieving sub-second level real-time feedback. Teachers can obtain students' evaluation information in real time and adjust teaching strategies in a timely manner without spending a lot of time on manual observation and recording. At the same time, the model adaptive module can dynamically adjust the confidence threshold of the YOLOv8 model according to the detection results, and the abnormal action recognition submodule can timely find the abnormal trajectory of the exercise equipment to ensure the smooth progress of teaching and comprehensively improve the overall efficiency and quality of physical education. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 The working principle diagram of the sports teaching evaluation system described in the present application;

[0069] Figure 2 The flowchart of the key point modeling module;

[0070] Figure 3 The flowchart of multi-modal data fusion processing;

[0071] Figure 4 The flowchart of the dynamic feedback module score and instruction generation. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0073] Please refer to Figures 1-4 The present application provides a technical solution: a sports teaching evaluation system based on artificial intelligence, which aims to use advanced technical means to comprehensively and accurately evaluate the sports teaching process and provide strong support for teaching. The system includes:

[0074] Data acquisition module: This module is responsible for receiving real-time video streams from the camera equipment and motion physiological data collected by wearable sensors. The camera equipment captures the activities of teachers, sports equipment and ball movement in the sports teaching scene in all directions; the wearable sensors monitor the heart rate, acceleration, angular velocity and other physiological information of students in real time during exercise, providing a multi-dimensional data basis for subsequent comprehensive analysis.

[0075] Multi-target detection module: Based on the YOLOv8 algorithm, this module accurately locates teachers, students, sports equipment and balls in real-time video streams. Through powerful target positioning function, it can quickly determine the position of each target in the video picture and output the detection box coordinates and target category, such as distinguishing students, teachers, basketballs, footballs and other different targets, providing a prerequisite for further analysis of actions and movement trajectories.

[0076] Key point modeling module: Using OpenPose algorithm, the key points of the detected target human body are extracted. Starting from the joints of the human body, a three-dimensional action posture model is constructed to accurately record the posture changes of the human body during exercise, providing key data support for action analysis.

[0077] Action trajectory analysis module: According to the generated three-dimensional action posture model, the Euclidean distance deviation of the movement trajectory and the preset standard action is calculated. By comparing the differences between the actual action and the standard action, the action standard evaluation result is generated to accurately judge the standardization of student action and provide quantitative basis for teaching improvement.

[0078] Multi-modal fusion module: Time sequence alignment is performed on the motion physiological data and the action standard evaluation result to construct a multi-dimensional feature matrix. Different types of data are organically combined to fully explore the potential association between data, providing richer and more comprehensive information for subsequent dynamic feedback.

[0079] Dynamic feedback module: Based on the multi-dimensional feature matrix, the LSTM network optimized by transfer learning generates real-time scores and individualized action correction instructions. Real-time scores allow students and teachers to promptly understand the learning and teaching effects; individualized action correction instructions provide targeted improvement suggestions for each student's specific situation to help students improve their action level faster.

[0080] The present application will be further described in conjunction with Examples 1 to 6:

[0081] Example 1:

[0082] In this embodiment, the specific execution process of the multi-target detection module is described in detail. The multi-target detection module is based on the YOLOv8 algorithm and is the key link for the whole system to realize accurate target positioning.

[0083] First, the real-time video stream obtained from the camera equipment is input into the pre-trained YOLOv8 model. YOLOv8 model is an advanced convolutional neural network that has strong feature extraction capability. In this process, the model processes the image data in the video stream through multiple convolutional layers, extracting multi-scale feature maps. These feature maps contain image information at different levels. Small-scale feature maps can capture the details of the target, and large-scale feature maps can help identify the overall contour and position of the target. For example, when identifying a basketball, small-scale feature maps can clearly present the texture details of the basketball surface, and large-scale feature maps can quickly locate the approximate position of the basketball in the entire scene.

[0084] Next, the extracted feature maps are subjected to anchor box matching and non-maximum suppression processing. Anchor box matching refers to matching the bounding box predicted by the model with the pre-set anchor box, and selecting the bounding box that best matches the target object by calculating indicators such as the Intersection over Union (IoU) between the two. Non-maximum suppression processing is to remove redundant detection boxes. In actual detection, multiple detection boxes may point to the same target object. Non-maximum suppression will sort the detection boxes according to their confidence levels, keeping the detection box with the highest confidence and removing those with high overlap and low confidence with the highest confidence detection box, thus ensuring that each target object is detected only once.

[0085] After this series of processing, the target detection result is output, including the detection box coordinates and target category. The detection box coordinates accurately describe the position of the target in the video picture, such as (x1, y1, x2, y2) representing the coordinates of the left upper corner and the right lower corner of the detection box; the target category clearly identifies the detected object, such as "student" "basketball" etc. In this way, the multi-target detection module can accurately and efficiently locate various targets in the sports teaching scene, providing a reliable data basis for subsequent analysis work.

[0086] Example 2:

[0087] When the multi-target detection module detects a human target, the key point modeling module first performs local image cropping on the detected human target. The entire human body is cropped from the video picture, and only the image area related to the human body is retained, which can reduce the data volume of subsequent processing and improve processing efficiency. Then, the cropped local image is input into the OpenPose network for heat map prediction. The OpenPose network analyzes the input image and predicts the heat map of each skeletal key point of the human body. The value of each pixel point in the heat map represents the probability of the position appearing as a skeletal key point, and the higher the value, the more likely it is that the position is a skeletal key point.

[0088] Next, the two-dimensional coordinates of the skeletal key points are extracted through a Gaussian filter and a peak detection algorithm. The Gaussian filter is a commonly used image smoothing method that can smooth the heat map and remove noise interference, making the heat map clearer and more stable. After Gaussian filtering, a peak detection algorithm is used to find the peak points in the heat map, and the coordinates corresponding to these peak points are the two-dimensional coordinates of the skeletal key points. For example, when detecting the key points of the human wrist, the two-dimensional coordinate position of the wrist in the image can be accurately found through the peak detection algorithm.

[0089] Finally, the two-dimensional coordinates are mapped to the three-dimensional space in combination with the depth information of the binocular camera to generate a timestamped action posture sequence. The binocular camera can obtain the depth information of the scene, and using these depth information, the two-dimensional coordinates obtained previously can be converted into three-dimensional coordinates, thereby constructing a three-dimensional action posture model of the human body. At the same time, a timestamp is added to each action posture to form an action posture sequence, recording the posture changes of the human body at different time points. In this way, through a series of operations of the key point modeling module, a three-dimensional action posture model of the human body can be accurately generated, providing accurate data support for subsequent action analysis.

[0090] Embodiment 3:

[0091] The action trajectory analysis module is based on the three-dimensional action posture model to comprehensively evaluate the student's actions. First, the three-dimensional action posture model is matched with the standard action template through dynamic time warping. In physical education, different sports have corresponding standard actions, and these standard action templates are pre-stored in the system. Dynamic time warping (DTW) can solve the problem of inconsistent time series data on the time axis, and it finds an optimal warping path to align the joint angle sequence of the actual action with the joint angle sequence of the standard action in time, thereby making a more accurate comparison. The calculation formula is:

[0092]

[0093] where A = (a1, …, a m ) and B = (b1, …, b n ) are the joint angle sequences of the actual action and the standard action, P is the set of all warping paths, (p, q) is a pair of matching points in the path π, which satisfies p ∈ A, q ∈ B, d(a p ,b q ) is the difference measure function of the two angle values. Through this formula, the difference in joint angles between the actual action and the standard action is calculated.

[0094] After calculating the joint angle deviation, the motion trajectory noise is filtered based on the Laplace smoothing algorithm. During the movement, due to various factors, the collected motion data may have noise, and the Laplace smoothing algorithm can effectively remove these noises, making the motion trajectory smoother and more accurate.

[0095] Then, the trajectory deviation of the continuous action frame is integrated, and the action continuity score is output. Action continuity is an important indicator to measure the quality of students' actions. By integrating the trajectory deviation, the continuity of the action can be quantified. For example, if the student is doing a shooting action, the transition between each action link is smooth, and the trajectory deviation integral value will be small, and the corresponding action continuity score will be high. Conversely, if the action is stuck and not smooth, the trajectory deviation integral value will be large, and the action continuity score will be low.

[0096] At the same time, when calculating the action standardization, the Euclidean distance deviation calculation formula will also be used:

[0097]

[0098] Where (x i ,y i ,z i ) represents the three-dimensional coordinates of the i-th joint, (x std ,y std ,z std ) represents the corresponding coordinates of the standard action, and n is the total number of joints. This formula is used to calculate the distance deviation of each joint in the actual action from the corresponding joint in the standard action in three-dimensional space, and the distance deviation of all joints is integrated to obtain the overall action standardization evaluation result. Through these calculations and analyses, the action trajectory analysis module can comprehensively and accurately evaluate the standardization and continuity of students' actions, providing valuable references for teaching.

[0099] Example 4:

[0100] The multi-modal fusion module aims to deeply integrate the motion physiological data and the action standardization evaluation results to obtain more comprehensive and accurate information. The heart rate, acceleration and angular velocity in the motion physiological data are denoised by wavelet transform. In the actual collection process, the motion physiological data is easily disturbed by the external environment and the device itself, producing noise. Wavelet transform is an effective signal processing method, which can decompose the signal into different frequency bands. By processing different frequency bands, the noise components are removed and the useful signal features are retained. For example, when processing heart rate data, wavelet transform can remove noise caused by motion artifacts and other factors, making the heart rate data more accurately reflect the student's true physiological state.

[0101] The de-noised physiological data and the action standard evaluation results are fused at a feature level to generate a fusion feature vector containing spatio-temporal correlations. Feature-level fusion is a direct fusion of features from different data sources, fully exploiting the potential connections between the data. In a sports teaching scenario, the student's physical function state is reflected by the physiological data, and the student's sports skill level is reflected by the action standard evaluation results, and there is a close correlation between the two. Through feature-level fusion, these information is integrated together to form a fusion feature vector containing spatio-temporal correlations. For example, in basketball teaching, if the heart rate is too high and the action standard is poor when the student is fast dribbling, the fusion feature vector can reflect the information of both aspects at the same time, providing a more comprehensive basis for subsequent analysis.

[0102] The attention mechanism dynamically allocates the weight ratio of visual data and sensor data. In multi-modal data fusion, different types of data may have different importance for the final analysis result. The attention mechanism can dynamically adjust the weight of visual data (such as image information in video stream) and sensor data (such as physiological data) according to the characteristics of the data and the current analysis requirements. In some cases, such as when analyzing a student's shooting action, the information about hand movement and basketball trajectory in the visual data may be more important, and the attention mechanism will increase the weight of this part of the visual data; while in the evaluation of the student's physical consumption, the heart rate and other information in the physiological data are more critical, and the attention mechanism will increase the weight of the sensor data. In this way, the multi-modal fusion module can more flexibly and effectively integrate multi-source data to provide better data support for subsequent dynamic feedback.

[0103] Embodiment 5:

[0104] This embodiment details the specific implementation process of the dynamic feedback module, which provides real-time scoring and personalized action correction instructions for students.

[0105] The dynamic feedback module is one of the core modules of the system to realize the teaching auxiliary function, which generates real-time scores and personalized action correction instructions based on multi-dimensional feature matrices. The course learning strategy is used to train the LSTM network in stages. The course learning strategy simulates the process of human learning new knowledge, learning from simple to complex step by step. In the early stage of training, the convolutional layer parameters are frozen in the first stage, because the convolutional layer is mainly responsible for extracting image features, and these features are relatively stable in the early stage of training. Freezing the convolutional layer parameters can speed up the training and avoid excessive adjustment of the convolutional layer in the early stage of training, which can cause the model to be unstable. In the second stage, the fully connected layer is optimized jointly, and the fully connected layer is responsible for classifying and predicting the extracted features. By jointly optimizing the fully connected layer, the parameters of the model can be better adjusted, and the prediction accuracy of the model can be improved.

[0106] The multi-dimensional feature matrix is segmented by a sliding window, and the LSTM network is input to predict the action score of the next 3 frames. The sliding window segmentation is to divide the multi-dimensional feature matrix according to a certain time step, and each window contains feature information within a period of time. These window data are input into the trained LSTM network in turn, and the LSTM network can capture the long-term dependence in the time series data. Through the learning of historical data, the action score of the next 3 frames is predicted.

[0107] According to the prediction result, an instruction set containing joint angle correction amount and movement rhythm adjustment suggestion is generated. For example, if it is predicted that the student's shooting action score in the next 3 frames is low, and it is found through analysis that it is caused by incorrect wrist joint angle, the system will generate the corresponding joint angle correction amount to guide the student to adjust the angle of the wrist; if it is also found that the student's shooting rhythm is too fast, the system will give a movement rhythm adjustment suggestion to let the student slow down the rhythm.

[0108] In the LSTM training process, the loss function used is:

[0109]

[0110] Wherein, is the total loss value of model training, is the predicted output at time t, y t is the true value, T is the time step, θ is the model parameter, and λ is the regularization coefficient. This loss function calculates the difference between the predicted value and the true value, and combines the regularization term to optimize the parameters of the model, so that the model can better fit the data and improve the prediction accuracy, thereby providing more reliable real-time scores and personalized action correction instructions for students.

[0111] Embodiment 6:

[0112] This embodiment covers the specific implementation of the model adaptive module, the abnormal action recognition submodule in the multi-target detection module, and the distributed computing module.

[0113] The model adaptive module dynamically adjusts the confidence threshold of the YOLOv8 model according to the real-time detection results. First, the false positive rate and the missed detection rate of target detection in the current scene are counted. The false positive rate refers to the proportion of non-target objects that are misjudged as target objects, and the missed detection rate refers to the proportion of target objects that are not detected. By statistically analyzing the detection results within a period of time, these two key indicators are obtained.

[0114] When the false positive rate exceeds the first threshold, the confidence threshold is reduced using an exponential decay algorithm. The formula of the exponential decay algorithm is:

[0115] c t = c0·e -kt

[0116] wherein c t is the confidence threshold at time t, c0is the initial threshold, k is the decay rate coefficient, and t is the iteration number. By reducing the confidence threshold, the false positive situation can be reduced. For example, in a football teaching scene, if the system frequently misjudges the trash on the sidelines as a football, when the false positive rate exceeds the first threshold set, the confidence threshold is reduced using the exponential decay algorithm, so that the system is more strict in judging the target in the subsequent detection, thereby reducing the false positives.

[0117] When the missed detection rate exceeds the second threshold, the model weights are updated locally by gradient backpropagation. Gradient backpropagation is a commonly used neural network training optimization method, which calculates the gradient of the loss function with respect to the model parameters, and updates the model parameters in the opposite direction of the gradient, so that the model is optimized in the direction of reducing the loss function during the training process. In this case, updating the model weights locally can improve the detection ability of the model for target objects and reduce the missed detection rate.

[0118] In the multi-target detection module, the abnormal motion recognition submodule is used to detect the abnormal trajectory of the sports equipment. The submodule first performs polynomial fitting on the ball movement trajectory and calculates the fitting residual. Polynomial fitting is to approximate the actual movement trajectory with a polynomial function, and the fitting residual calculation formula is:

[0119]

[0120] wherein, is the trajectory fitting residual, p i is the i-th trajectory point coordinate, t i is the timestamp, a j is the polynomial coefficient, d is the polynomial degree, and N is the total number of trajectory points. When the residual exceeds the preset threshold, it indicates that the ball movement trajectory is abnormal, at which time the high-speed camera device is triggered to perform 1000fps slow motion playback analysis, so as to more clearly observe and analyze the abnormal situation.

[0121] The distributed computing module is used to realize parallel data processing of multiple cameras. First, Apache Kafka is used to build a message queue to synchronize the timestamps of multiple video streams. Apache Kafka is a high-throughput distributed message system that can effectively manage and transmit large amounts of data streams. By adding timestamps to multiple video streams and synchronizing them, it can ensure that the data collected by different cameras is consistent in time, providing accurate time series data for subsequent analysis.

[0122] Then, the detection task is distributed to the GPU cluster through the Dask framework to achieve real-time feedback with sub-second delay. The Dask framework is an open-source framework for distributed computing, which can decompose and distribute computing tasks to multiple computing nodes for execution. The GPU cluster has strong parallel computing capability, and by deploying the detection task to the GPU cluster through the Dask framework, the computing resources of the GPU can be fully utilized, the data processing speed can be greatly improved, and real-time feedback with sub-second delay can be achieved, so that the system can timely monitor and evaluate the sports teaching process. Through the cooperative work of these modules, the system can run more intelligently and efficiently, providing more comprehensive and accurate support for sports teaching evaluation.

[0123] It should be noted that, in this paper, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying that there is any such actual relationship or order between such entities or actions. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed, or inherent to such a process, method, article or apparatus.

[0124] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, alternatives and variations can be made thereto without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A physical education teaching and assessment system based on artificial intelligence, characterized in that, include: The data acquisition module is used to receive real-time video streams from the camera device and motion physiological data from wearable sensors; The multi-target detection module, based on the YOLOv8 algorithm, locates teachers, students, sports equipment, and balls in the real-time video stream, and outputs the coordinates of the detection box and the target category. The key point modeling module uses the OpenPose algorithm to extract the skeletal key points of the detected target human body and generate a 3D motion pose model. The motion trajectory analysis module calculates the Euclidean distance deviation between the motion trajectory and the preset standard motion based on the three-dimensional motion posture model, and generates a motion standardization evaluation result. The multimodal fusion module aligns the motion physiological data with the action standardization evaluation results in time sequence to construct a multidimensional feature matrix; The dynamic feedback module generates real-time scores and personalized action correction instructions based on the multidimensional feature matrix and through a transfer learning-optimized LSTM network. The execution steps of the key point modeling module include: The detected human target is locally cropped and input into the OpenPose network for heatmap prediction; Two-dimensional coordinates of key skeletal points are extracted using Gaussian filtering and peak detection algorithms. By combining depth information from binocular cameras, two-dimensional coordinates are mapped to three-dimensional space to generate a sequence of action poses with timestamps; The execution steps of the motion trajectory analysis module include: The three-dimensional motion posture model is dynamically time-warped and matched with the standard motion template to calculate the joint angle deviation. Motion trajectory noise is filtered based on the Laplace smoothing algorithm; Integrate the trajectory deviation of consecutive action frames and output an action continuity score; The formula for calculating the Euclidean distance deviation is: Among them, (x i ,y i ,z i (x) represents the three-dimensional coordinates of the i-th joint. std ,y std ,z std () represents the coordinates corresponding to the standard motion, and n is the total number of joints; The calculation formula for the dynamic time warping matching is: Where A = (a1, ..., a m ) and B = (b1,…,b n Let π be the joint angle sequence of the actual movement and the standard movement, respectively; let P be the set of all regular paths; and let (p, q) be a pair of matching points in path π such that p∈A, q∈B, and d(a p ,b q () is a function that measures the difference between two angle values; The execution steps of the dynamic feedback module include: A course-based learning strategy is adopted to train the LSTM network in stages. The first stage freezes the parameters of the convolutional layers, and the second stage jointly optimizes the fully connected layers. The multidimensional feature matrix is ​​segmented using a sliding window, and the result is input into the LSTM network to predict the action scores for the next 3 frames. Generate an instruction set that includes joint angle corrections and motion rhythm adjustment suggestions based on the prediction results; The loss function used in the LSTM training is: in, This represents the total loss value during model training. For the predicted output at time t, y t λ represents the true value, T represents the time step, θ represents the model parameters, and λ represents the regularization coefficient.

2. The system as described in claim 1, characterized in that, The execution steps of the multi-target detection module include: The real-time video stream is input into a pre-trained YOLOv8 model, and multi-scale feature maps are extracted through a convolutional neural network. Anchor box matching and non-maximum suppression are performed on the feature map to output the target detection result.

3. The system as described in claim 1, characterized in that, The execution steps of the multimodal fusion module include: Wavelet transform was used to denoise the heart rate, acceleration, and angular velocity in the exercise physiological data. The denoised physiological data is fused with the action standardization evaluation results at the feature level to generate a fused feature vector containing spatiotemporal correlation; The weight ratio between visual data and sensor data is dynamically allocated through an attention mechanism.

4. The system as described in claim 1, characterized in that, Also includes: The model adaptation module is used to dynamically adjust the confidence threshold of the YOLOv8 model based on real-time detection results; The execution steps of the model adaptation module include: Statistically analyze the false alarm rate and false negative rate of target detection in the current scenario; When the false alarm rate exceeds the first threshold, the confidence threshold is reduced by an exponential decay algorithm. When the false negative rate exceeds the second threshold, the model weights are locally updated through gradient backpropagation. The formula for the exponential decay algorithm is: c t =c0·e -kt Among them, c t Let t be the confidence threshold, c0 be the initial threshold, k be the decay rate coefficient, and t be the number of iterations.

5. The system as described in claim 1, characterized in that, The multi-target detection module further includes: The abnormal motion recognition submodule is used to detect abnormal trajectories of exercise equipment; The execution steps of the submodule include: Perform polynomial fitting on the trajectory of the ball and calculate the fitting residuals; When the residual exceeds a preset threshold, the high-speed camera is triggered to perform slow-motion playback analysis at 1000fps. The formula for calculating the polynomial fitting residual is as follows: Where ∈ represents the trajectory fitting residual, p i Let t be the coordinates of the i-th trajectory point. i For timestamps, a j denoted as the polynomial coefficient, d as the polynomial degree, and N as the total number of trajectory points.

6. The system as described in claim 1, characterized in that, Also includes: A distributed computing module is used to implement parallel data processing from multiple cameras; The execution steps of the distributed computing module include: A message queue is built using Apache Kafka to synchronize timestamps of multiple video streams; By distributing detection tasks to a GPU cluster using the Dask framework, real-time feedback with sub-second latency is achieved.

Images