Real-time feedback method and computer device for students' running posture based on motion recognition

Through video data analysis and clustering processing, abnormal areas and typical problems of running postures are determined, and feedback system is deployed, which solves the problem of lack of targetedness and timeliness of running posture feedback in the existing technology, and realizes personalized and real-time running postures Feedback improves students' running experience and safety.

CN118865503BActive Publication Date: 2024-12-06ZHUHAI XINWEI TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411312571.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-20
Publication Date
2024-12-06
Estimated Expiration
2044-09-20

AI Technical Summary

Technical Problem

The existing technology is difficult to provide real-time and effective feedback on abnormal problems in students' running postures, especially in consideration of different areas of the track and individual differences, which lacks targetedness and timeliness, affecting students' running experience and safety.

Method used

By collecting video data, building a running pose data set, clustering and area division, determining typical pose problems and data density distribution, selecting appropriate feedback locations and ranges, and deploying feedback systems to provide targeted feedback in real time.

Benefits of technology

It realizes dynamic three-dimensional analysis of students' running postures, provides personalized and real-time feedback, improves the pertinence and effectiveness of posture correction, and enhances students' running experience and safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118865503B_ABST
    Figure CN118865503B_ABST
Patent Text Reader

Abstract

The present application relates to a method and computer device for real-time feedback of students' running posture based on motion recognition. The method comprises: constructing a first running posture data set; determining a high-incidence area of ​​posture abnormality; dividing the target runway area into M initial sub-areas according to the first clustering processing result; constructing a second running posture data set, and dividing the initial sub-area into N fine sub-areas according to the second clustering processing result; determining the main problem area according to typical posture problems and posture data density distribution; determining candidate feedback positions within the fine sub-areas; selecting the candidate feedback position with the highest comprehensive evaluation score and covering the main problem area as the feedback system layout position; determining the effective feedback range of the feedback system layout position; triggering the feedback system to provide running posture feedback to students based on typical posture problems. Analyzing a specific runway area in combination with the running process makes the running posture analysis more targeted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of machine vision technology, and in particular to a method and computer device for real-time feedback of students' running posture based on motion recognition. Background Art

[0002] Standardized running posture can not only improve running efficiency, but also prevent sports injuries. However, many students lack a clear understanding of the correct running posture, and often have posture problems such as leaning forward, leaning back, and unstable center of gravity during running. If these problems are not corrected in time, they are likely to cause harm to the body and affect students' running experience and enthusiasm. In traditional physical education teaching, it is difficult for teachers to observe and provide feedback on each student's running posture in real time. After class, students lack effective supervision and guidance when exercising independently, which makes it difficult to correct many posture problems.

[0003] In recent years, computer vision-based action recognition technology has made significant progress in the field of motion analysis. Some studies have begun to try to apply these technologies to running posture analysis. For example, deep learning models such as OpenPose and AlphaPose are used to identify and locate key points of the human body, providing a basis for quantitative analysis of running posture. However, these methods still have some limitations.

[0004] First, most existing studies focus on the analysis of a single perspective, which makes it difficult to fully capture the three-dimensional motion changes during running. Second, existing methods often ignore the impact of different areas of the track (such as straights and curves) on running posture and lack targeted analysis. Furthermore, many systems focus mainly on the static characteristics of posture, while ignoring the impact of dynamic factors such as speed and acceleration on running efficiency.

[0005] In addition, most existing feedback systems use a unified standard to judge the postures of all students, lacking consideration for individual differences. At the same time, the feedback is often not provided in a timely or intuitive manner, making it difficult to be effectively received and applied during exercise.

[0006] Finally, the existing technology lacks comprehensive analysis of posture changes during running and precise positioning of problem areas, which limits the pertinence and effectiveness of feedback. Summary of the invention

[0007] Based on this, it is necessary to provide a real-time feedback method and computer equipment for students' running posture based on motion recognition, which can effectively feedback posture problems during running, in response to the above technical problems.

[0008] A method for real-time feedback of students' running posture based on motion recognition comprises the following steps:

[0009] Collecting first video data of a student running in a target track area, and constructing a first running posture data set according to the first video data;

[0010] Analyze the first running posture data set to determine high incidence areas of abnormal posture;

[0011] Based on the high-incidence areas of abnormal posture, a first clustering process is performed on the first running posture data set, and the target runway area is divided into M initial sub-areas according to the first clustering process result;

[0012] Collecting second video data of running in the initial sub-region, extracting coordinates of key points of the student's body from the second video data, constructing a second running posture data set, performing a second clustering process on the first running posture data set and the second running posture data, and dividing the initial sub-region into N fine sub-regions according to the second clustering process result;

[0013] According to the refined sub-region, typical posture problems and posture data density distribution corresponding to the initial sub-region are obtained, and the main problem area is determined according to the typical posture problems and posture data density distribution;

[0014] Determine candidate feedback locations in the fine sub-region based on the posture data density distribution and the problem severity index corresponding to the main problem area;

[0015] Conduct a comprehensive evaluation of candidate feedback locations, and select the candidate feedback location with the highest comprehensive evaluation score and covering the main problem area as the location for the feedback system deployment;

[0016] Taking the feedback system layout location as the center and combining the attitude data density distribution, determine the effective feedback range of the feedback system layout location;

[0017] The feedback system is deployed at the feedback system layout location, and a corresponding trigger area is set for the feedback system layout location according to the effective feedback range. When the student enters the trigger area, the feedback system is triggered to provide the student with running posture feedback based on typical posture problems.

[0018] In one embodiment, the step of analyzing the first running posture data set to determine the high-incidence area of ​​abnormal posture includes the steps of:

[0019] Divide the target runway area into N intervals evenly;

[0020] Calculate the posture abnormality score corresponding to the interval;

[0021] The sum of the mean and standard deviation of the posture abnormality score was calculated, and the interval where the posture abnormality score was greater than the sum was defined as the high-incidence area of ​​posture abnormality.

[0022] In one embodiment, based on the high incidence area of ​​abnormal posture, the first clustering process is performed on the first running posture data set, and the target runway area is divided into M initial sub-areas according to the first clustering process result, including the steps of:

[0023] Define a first-class feature vector corresponding to a data point in the first running posture data set;

[0024] Performing a first clustering process on the first type of feature vectors using a first clustering algorithm to obtain a first type of cluster set;

[0025] The first cluster set is mapped to the target runway area to obtain an initial sub-area.

[0026] In one embodiment, the steps of collecting second video data of running in the initial sub-region, constructing a second running posture data set according to the second video data, performing a second clustering process on the first running posture data set and the second running posture data set, and dividing the initial sub-region into N fine sub-regions according to the second clustering process result include the steps of:

[0027] Merging the first running posture data set and the second running posture data set to obtain a comprehensive data set;

[0028] Define the corresponding second-class eigenvectors for the data points in the synthetic dataset;

[0029] Using a second clustering algorithm to perform a second clustering process on the second type of feature vectors corresponding to the data points in the initial sub-region to obtain a second type of cluster set;

[0030] The second cluster set is mapped to the initial sub-region to obtain a refined sub-region.

[0031] In one embodiment, the steps of obtaining typical posture problems and posture data density distribution corresponding to the initial sub-region according to the refined sub-region, and determining the main problem region according to the typical posture problems and posture data density distribution, include the steps of:

[0032] Using the posture score function, calculate the posture score corresponding to the data point in the fine sub-region;

[0033] Calculate the mean and standard deviation of the posture features in the fine sub-region, and obtain the abnormal threshold based on the mean and standard deviation;

[0034] The posture features corresponding to the posture scores greater than the abnormal threshold are regarded as typical posture problems;

[0035] The kernel density estimation method is used to calculate the density distribution of posture data in the fine sub-region;

[0036] According to the density distribution of posture data and the weights of typical posture problems, the spatial distribution of the severity of the problem corresponding to the fine sub-region is calculated;

[0037] Sort the spatial distribution of problem severity from large to small, and take the first part of the sorting results as the main problem area.

[0038] In one embodiment, the step of determining candidate feedback positions in the fine sub-region according to the posture data density distribution and the problem severity index corresponding to the main problem region includes the steps of:

[0039] The reaction time function, the density distribution of posture data and the problem severity index corresponding to the main problem area are weighted to obtain the feedback position scoring function; the reaction time function indicates how much time and space the student has to adjust the posture after providing feedback at a certain position;

[0040] Using the feedback position scoring function, calculate the feedback position score corresponding to the position in the fine sub-region;

[0041] Based on the sliding window algorithm, the local maximum value in the feedback position score is obtained;

[0042] The position corresponding to the local maximum value greater than the preset threshold is taken as the primary candidate feedback position; the preset threshold is set according to the average value and standard deviation of the feedback position score;

[0043] Sort the primary candidate feedback positions from large to small, and calculate the distance between any two primary candidate feedback positions in the sorting result. When the distance is less than a preset interval, delete the primary candidate feedback position with the smallest local maximum value among the two primary candidate feedback positions to obtain a secondary candidate feedback position.

[0044] According to the length of the fine sub-region and the preset maximum number of feedback positions, the first W secondary candidate feedback positions from largest to smallest are used as candidate feedback positions.

[0045] In one embodiment, the step of comprehensively evaluating candidate feedback locations and selecting a candidate feedback location with the highest comprehensive evaluation score and covering a major problem area as a location for deploying the feedback system includes the following steps:

[0046] Calculate the coverage effect score, problem targeting score, feedback timeliness score and environmental adaptability score of the candidate feedback location;

[0047] Weighted processing is performed on the coverage effect score, problem targeting score, feedback timeliness score, and environmental adaptability score to obtain a comprehensive evaluation score for the candidate feedback location;

[0048] Sort the candidate feedback positions according to the comprehensive evaluation scores from large to small to obtain a sorted list;

[0049] According to the coverage function and the sorting list, the feedback system layout position is obtained; the coverage function represents the coverage degree of the selected position set to the main problem area.

[0050] In one embodiment, the step of determining the effective feedback range of the feedback system layout position with the feedback system layout position as the center and combining the posture data density distribution includes the steps of:

[0051] Smoothing the density distribution of attitude data;

[0052] Setting an initial search range corresponding to the feedback system placement location;

[0053] In the initial search range, the primary effective feedback range is determined using the smoothed attitude data density distribution;

[0054] It is determined whether there is an overlapping area between adjacent primary effective feedback ranges. If there is an overlapping area, the adjacent effective feedback ranges are adjusted based on the overlapping area to obtain the effective feedback range of the feedback system deployment position.

[0055] In one embodiment, the feedback system is deployed at the feedback system deployment position, a trigger area is set for the feedback system deployment position according to the effective feedback range, and when the student enters the trigger area, the feedback system is triggered to provide the student with running posture feedback based on typical posture problems, including the steps of:

[0056] Score the student's current posture based on the real-time posture evaluation function to obtain the current posture score;

[0057] When the current posture score exceeds the trigger threshold corresponding to the trigger area, running posture feedback is provided to the student based on the feedback strategy corresponding to the typical posture problem.

[0058] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0059] One of the above technical solutions has the following advantages and beneficial effects:

[0060] The present application comprises the following steps: collecting first video data of a student running in a target track area, extracting coordinates of key body points of the student from the first video data, and constructing a first running posture data set; analyzing the first running posture data set to determine high-incidence areas of posture abnormalities; performing a first clustering process on the first running posture data set based on the high-incidence areas of posture abnormalities, and dividing the target track area into M initial sub-areas according to the first clustering process result; collecting second video data of a student running in the initial sub-area, extracting coordinates of key body points of the student from the second video data, and constructing a second running posture data set; performing a second clustering process on the first running posture data set and the second running posture data set together, and dividing the initial sub-area into N refined sub-areas according to the second clustering process result; and obtaining, according to the refined sub-areas, Typical posture problems and posture data density distribution corresponding to the initial sub-area, and the main problem area is determined based on the typical posture problems and posture data density distribution; candidate feedback positions are determined in the fine sub-area based on the posture data density distribution and the problem severity index corresponding to the main problem area; the candidate feedback positions are comprehensively evaluated, and the candidate feedback positions with the highest comprehensive evaluation score and covering the main problem area are selected as the feedback system layout position; the feedback system layout position is taken as the center, and the effective feedback range of the feedback system layout position is determined in combination with the posture data density distribution; the feedback system is deployed at the feedback system layout position, and the corresponding trigger area is set for the feedback system layout position according to the effective feedback range. When the student enters the trigger area, the feedback system is triggered to provide running posture feedback to the student based on the typical posture problem. This application analyzes the dynamic three-dimensional movements of students running based on video data. In the process of video analysis, a specific runway area is analyzed in combination with the running process, so that the running posture analysis is more targeted. In addition, in the process of analyzing the running posture, the differences of the students themselves are considered, so that the analysis results are closer to the actual situation of each student. Finally, the student's running posture can be reminded and adjusted in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the flow of the method for real-time feedback of students' running posture based on motion recognition in an embodiment of the present application.

[0062] Figure 2 Schematic diagram of the internal structure of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0064] To achieve the above object, in one embodiment, Figure 1As shown, a method for real-time feedback of students' running posture based on motion recognition is provided, comprising the following steps:

[0065] Step S110, collecting first video data of a student running in a target track area, and constructing a first running posture data set based on the first video data.

[0066] First, high-precision camera equipment is deployed in the target runway area. According to the length and shape of the runway, multiple high-speed cameras are evenly distributed around the runway. These cameras must meet the following requirements: resolution not less than 1920x1080 pixels, frame rate at least 120fps, and field of view not less than 120 degrees. A 30% overlap area is maintained between adjacent cameras to ensure seamless coverage of the runway and continuity of subsequent data processing. The installation height and angle of the camera are carefully calculated to minimize occlusion and distortion.

[0067] The first video data collection process is arranged in different time periods, including students' physical education classes and extracurricular activities. This ensures that data is collected at different times and under different lighting conditions, increasing the diversity and representativeness of the data. During each collection period, students of different ages, genders and athletic abilities are selected to participate in running, and each student must complete at least one full lap of the track. The total sample size collected is adjusted according to the number of students in the school and research needs. It is usually recommended to collect the first video data of no less than 200 students to ensure the validity of statistical analysis.

[0068] After the acquisition is completed, the video processing and key point extraction phase begins. Each frame of video is processed using advanced real-time action recognition algorithms such as OpenPose, AlphaPose, or custom deep learning models. These algorithms can accurately identify and locate the key points of the human body, which usually include but are not limited to the following 17 points: top of the head, tip of the nose, left and right eyes, left and right ears, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles. For each key point, its two-dimensional coordinates (x, y) and confidence score s are extracted.

[0069] After the key points are extracted, the first video data is preprocessed and standardized. First, the coordinates from different camera perspectives are converted to a unified runway coordinate system using fixed markers on the runway. This ensures that data collected from different areas can be compared and analyzed in the same coordinate system. Then, all data are time-aligned to ensure that data from different students have the same starting point, which facilitates subsequent horizontal comparisons.

[0070] In order to eliminate the influence of individual height differences on data analysis, the scale of the posture data of each frame is normalized. This process is achieved by dividing the coordinates of all key points by the torso length (usually defined as the distance from the midpoint of the shoulder to the midpoint of the hip). The formula for calculating the normalized coordinates is:

[0071] x'=x / L

[0072] y'=y / L

[0073] Where L represents the torso length. This normalization process makes the posture data of students of different heights comparable. The normalized key point coordinate set is denoted as P.

[0074] In order to capture the dynamic characteristics of running, the velocity v and acceleration a of key points between adjacent frames are calculated. The calculation formulas for velocity and acceleration are as follows:

[0075] v(t)=(P(t)-P(t-1)) / Δt

[0076] a(t)=(v(t)-v(t-1)) / Δt

[0077] Where Δt is the time interval between adjacent frames, usually 1 / 30 second (assuming the resampled frame rate is 30fps). P(t) represents the set of key point coordinates at time t.

[0078] During data processing, confidence thresholds are set to filter out low-quality key point data. For missing data due to occlusion or other reasons, interpolation or prediction methods are used to repair them to ensure the continuity and integrity of the data. For obviously abnormal data points, such as coordinate jumps caused by recognition errors, median filtering or Kalman filtering and other technologies are used for smoothing.

[0079] At the same time, the timestamp t of each frame of data and the position l of the student on the track are recorded. The position l is represented by a continuous value starting from 0, which can be calculated from the marked points on the track and the student's motion trajectory.

[0080] The first running posture dataset D finally constructed contains the following information:

[0081] D={(P1,v1,a1,t1,l1),(P2,v2,a2,t2,l2),...,(Pn,vn,an,tn,ln)}

[0082] Where Pi is the normalized key point coordinate set of the i-th frame, vi and ai are the corresponding velocity and acceleration vector sets, ti is the timestamp, and li is the current runway position (a continuous value with the starting point as 0).

[0083] Step S120, analyzing the first running posture data set to determine a high incidence area of ​​abnormal posture.

[0084] In one example, the step of analyzing a first running posture data set to determine a high-incidence area of ​​posture abnormalities includes the steps of: dividing the target runway area into N intervals on average; calculating the posture abnormality score corresponding to the interval; calculating the sum of the mean and standard deviation of the posture abnormality score, and defining the interval with a posture abnormality score greater than the sum as a high-incidence area of ​​posture abnormalities.

[0085] First, the entire target runway area is divided into N intervals according to the runway length, and the length of each interval is L meters. This division method can help accurately locate the location where the abnormal posture occurs. The selection of the number of intervals N and the length L requires a trade-off between accuracy and computational complexity, and can usually be determined based on the total length of the runway and the desired analysis accuracy. For example, for a standard 400-meter runway, N=40 and L=10 meters can be selected, so that more detailed spatial distribution information can be obtained.

[0086] For each interval i, a posture anomaly score S(i) is calculated. This score reflects the degree of posture anomaly in the interval. The calculation formula is as follows:

[0087] S(i)=Σ(A(Pj,vj,aj)) / Ni,forall(Pj,vj,aj,tj,lj)inDwherei*L≤lj<(i+1)*L

[0088] In this formula, A(Pj,vj,aj) is the posture anomaly scoring function, which is used to evaluate the degree of anomaly of a single data point j. Ni represents the number of data points in interval i. This formula actually calculates the average degree of anomaly of all data points in interval i. In this way, the continuous runway space can be discretized into a series of scoring values, which is convenient for subsequent analysis and judgment. Among them, Pj is the coordinate set of data point j in the i-th frame, vj and aj are the velocity and acceleration vector sets corresponding to data point j, tj is the timestamp corresponding to data point j, and lj is the runway position corresponding to data point j.

[0089] The design of the posture anomaly scoring function A is the key, which determines how to judge whether a posture is abnormal. This function takes into account the deviation of key point position, velocity and acceleration:

[0090] A(P,v,a)=w1*D(P,P_std)+w2*|v-v_avg| / v_avg+w3*max(0,|a|-a_max) / a_max

[0091] In this function, D(P, P_std) represents the normalized distance between the current posture and the standard posture template, which is used to measure the degree of deviation of the posture. Specifically, the Euclidean distance between each key point of the current posture and the corresponding point of the standard template can be calculated, and then summed or maximized. v_avg is the average speed, which is used to judge speed abnormalities. By comparing the difference between the current speed and the average speed, abnormal acceleration or deceleration can be detected. a_max is the expected maximum acceleration, which is used to detect acceleration abnormalities. When the acceleration exceeds this threshold, it may indicate a sharp speed change, which is usually associated with a bad posture. w1, w2, and w3 are weight coefficients, and the importance of each factor can be adjusted according to actual needs. For example, if you are more concerned about the posture itself rather than the speed change, you can increase the value of w1.

[0092] After calculating S(i) for each interval, the next step is to determine which intervals can be considered as high-incidence areas. Here, a statistical threshold method is used to define intervals where S(i) is greater than the mean plus the standard deviation as high-incidence areas:

[0093] High-incidence area = {i|S(i)>μS+σS}

[0094] Where μS is the mean value of S(i) and σS is the standard deviation of S(i). This method is based on the characteristics of normal distribution and can effectively identify areas that are significantly higher than the average level. Selecting μS+σS as the threshold means that about 15.9% of the intervals will be marked as high-incidence areas. This ratio can be adjusted according to actual needs.

[0095] Therefore, the high incidence area of ​​posture abnormality can be expressed as a set of intervals:

[0096] R_high={[i*L,(i+1)*L)|i∈high incidence area}

[0097] Where [i*L,(i+1)*L) represents the half-open interval from i*L to (i+1)*L. This set intuitively describes which positions on the runway are most likely to have abnormal postures.

[0098] Step S130: Based on the high-incidence areas of abnormal posture, a first clustering process is performed on the first running posture data set, and the target runway area is divided into M initial sub-areas according to the first clustering process result.

[0099] In one example, based on the high-incidence areas of posture abnormalities, a first clustering process is performed on a first running posture data set, and a target runway area is divided into M initial sub-areas according to the first clustering process results, including the steps of: defining corresponding first-category feature vectors for data points in the first running posture data set; performing a first clustering process on the first-category feature vectors using a first clustering algorithm to obtain a first-category cluster set; and mapping the first-category cluster set to the target runway area to obtain an initial sub-area.

[0100] For each data point (Pi, vi, ai, ti, li) in the first running posture data set D, a feature vector Fi (i.e., the first type of feature vector) is defined to describe its characteristics:

[0101] Fi = [Pi, vi, ai, I (li ∈ R_high)]

[0102] This feature vector contains four parts. The first three parts Pi, vi and ai come directly from the original data, representing the posture, velocity and acceleration information respectively. The second part I(li∈R_high) is an indicator function, which is used to mark whether the data point is located in a high-incidence area. Among them, I(·) is an indicator function, which takes the value of 1 when the condition is met and 0 otherwise.

[0103] Next, use the K-means clustering algorithm to cluster these feature vectors and obtain the first cluster set. The core idea of ​​the K-means algorithm is to divide n data points into K clusters so that each data point belongs to the cluster center closest to it. The goal of the algorithm is to minimize the sum of the squares of the distances from all data points to the cluster center to which they belong. Mathematically, this can be expressed as minimizing the following objective function:

[0104] J=Σ(i=1ton)Σ(j=1toK)||Fi-μj||^2*I(Fi∈Cj)

[0105] In this objective function, n represents the total number of data points, and K is the preset number of clusters. μj represents the center point of the jth cluster, which is the average position of all data points in the cluster. Cj represents the jth cluster, which is a collection of data points. ||Fi-μj|| represents the Euclidean distance from the feature vector Fi to the cluster center μj. I(Fi∈Cj) is another indicator function, which is 1 when Fi belongs to cluster Cj and 0 otherwise. This objective function actually calculates the sum of the squares of the distances from all data points to their respective cluster centers.

[0106] The execution process of the K-means algorithm is as follows: First, randomly select K points as the initial cluster centers, and then assign each data point to the cluster center closest to it to form K initial clusters. Next, recalculate the center point of each cluster (that is, the average position of all points in the cluster). Repeat this assignment and recalculation process until the cluster center no longer changes significantly or reaches the preset maximum number of iterations.

[0107] The selection of the number of clusters K is determined using the Elbow Method. The specific approach is to try different K values, usually starting from 2 and gradually increasing to a reasonable upper limit, such as 10. For each K value, calculate the corresponding objective function value J. As K increases, the J value will gradually decrease, but the rate of decrease will gradually slow down. When the decrease in the J value caused by increasing the K value is no longer significant, it is considered that a suitable K value has been found. In practical applications, the K value usually corresponds to the main functional areas of the runway, such as the starting area, straight area, curve area, and finish area.

[0108] After executing the K-means algorithm, K clusters C1, C2, ..., CK (i.e., the first cluster set) are obtained. Each cluster Cj corresponds to an initial sub-region on the runway. In order to map these clustering results back to the physical space of the runway, the initial sub-region Sj is defined as follows:

[0109]

[0110] The meaning of this definition is that if there is a data point belonging to cluster Cj and its position coordinate is l, then l belongs to the initial sub-region Sj. In this way, the correspondence between the clustering result and the actual runway space is established.

[0111] By this method, the final initial sub-region division results S1, S2, ..., SK are obtained. Each initial sub-region Sj is a continuous interval or the union of several continuous intervals on the runway. These initial sub-regions can be identified as different functional areas according to their positions and characteristics on the runway. For example:

[0112] 1. Starting area (S1): Usually located at the beginning of the track, characterized by large acceleration and gradually increasing speed.

[0113] 2. The first straight area (S2): Immediately following the starting area, the speed is relatively stable and the posture is more standardized.

[0114] 3. First turn area (S3): The first turning part of the runway, which may include two stages: turning in and turning out.

[0115] 4. Second straight area (S4): The straight road after the first turn, with characteristics similar to the first straight area.

[0116] 5. Second turning area (S5): The second turning section has similar characteristics to the first turning area.

[0117] 6. Final straight section (S6): The last straight section, which may include the sprint stage.

[0118] 7. Finish Zone (S7): The last part of the track where deceleration or the final sprint may occur.

[0119] The purpose of the clustering method based on R_high is: First, R_high represents the areas on the track that are most prone to posture problems. Including it in the clustering process ensures that the initial sub-region division pays more attention to these problem-prone areas. This method not only captures the physical characteristics of the track (such as straights and curves), but also reflects the specific posture challenges that students may face in different areas. The I(li∈R_high) term in the feature vector Fi allows the clustering algorithm to pay special attention to data points located in the problem-prone areas, thereby giving these areas more importance in the initial sub-region division.

[0120] Step S140, collecting second video data of running in the initial sub-region, constructing a second running posture data set based on the second video data, performing a second clustering process on the first running posture data set and the second running posture data set, and dividing the initial sub-region into N fine sub-regions based on the second clustering process result.

[0121] It should be noted that the steps for constructing the second running posture data set are the same as those for constructing the first running posture data set, and will not be repeated here.

[0122] In one example, a second video data of running in an initial sub-region is collected, a second running posture data set is constructed based on the second video data, a second running posture data set is jointly subjected to a second clustering process on the first running posture data set and the second running posture data set, and the initial sub-region is divided into N fine sub-regions based on the second clustering process result, including the steps of: merging the first running posture data set and the second running posture data set to obtain a comprehensive data set; defining corresponding second-category feature vectors for data points in the comprehensive data set; performing a second clustering process on the second-category feature vectors corresponding to the data points in the initial sub-region using a second clustering algorithm to obtain a second-category cluster set; and mapping the second-category cluster set to the initial sub-region to obtain a fine sub-region.

[0123] The main purpose of collecting the second video data is to increase the diversity and representativeness of the data, thereby improving the accuracy of the subsequent cluster analysis. Taking the 400-meter standard track as an example, assume that the track is divided into four main sub-areas: the starting area, the straight area, the curve area, and the finish area. In each initial sub-area, students are arranged to run, and their running postures are recorded using high-speed cameras. Assume that the structure of the collected second running posture dataset D' is the same as that of the first running posture dataset D: D'={(P'1,v'1,a'1,t'1,l'1),(P'2,v'2,a'2,t'2,l'2),...,(P'm,v'm,a'm,t'm,l'm)}, where m is the total amount of newly collected data.

[0124] Merge the second running posture dataset D' with the first running posture dataset D to obtain a comprehensive dataset D*: D*=D∪D'={(P1,v1,a1,t1,l1),...,(Pn,vn,an,tn,ln),(P'1,v'1,a'1,t'1,l'1),...,(P'm,v'm,a'm,t'm,l'm)}.

[0125] Next, define a feature vector F (i.e., the second type of feature vector) for each data point in D* to capture the key features of the running posture: F=[P,v,a,S(l),Q(P,v,a)]. Here, P represents the coordinates of the key points of the body, v is the speed, a is the acceleration, S(l) represents the initial sub-region to which the data point belongs (determined by step S130), and Q(P,v,a) is a posture feature function used to capture the overall characteristics of the running posture and provide input for clustering analysis. The Q function is designed based on expert knowledge, Q(P,v,a)=w1*A(P)+w2*B(v)+w3*C(a) where A(P) evaluates the overall standard degree of the posture, B(v) evaluates the rationality of the speed, C(a) evaluates the stability of the acceleration, and w1,w2,w3 are weight coefficients.

[0126] Continue with fine clustering to obtain the second set of clusters. Perform fine clustering on each initial sub-region Sj. Taking the bend area as an example, first extract all data points belonging to the sub-region: D_bend={(P,v,a,t,l,F)∈D*|l∈bend area}. Use the K-means++ algorithm to cluster D_bend. The K-means++ algorithm is an improved version of the K-means algorithm, which improves the clustering effect by optimizing the selection of the initial center point. The algorithm first randomly selects a data point as the first cluster center, then calculates the distance from each data point to the nearest cluster center, and selects a new cluster center with the square of the distance as the weight. This process is repeated until K cluster centers are selected. Then, use these K cluster centers to start the standard K-means algorithm.

[0127] The objective function of clustering is: J=Σ(F∈D_bend)Σ(k=1toK)||F-μk||^2*I(F∈Ck), where K is the number of clusters in the bend area, μk is the center of the kth cluster, Ck is the kth cluster, ||·|| represents the Euclidean distance, and I(·) is the indicator function. The number of clusters K is selected using the Silhouette Coefficient method. For different K values ​​(usually starting from 2 to a reasonable upper limit, such as 10), the corresponding silhouette coefficient is calculated: SC(K)=Σ(F∈D_bend)(b(F)-a(F)) / max(a(F),b(F)), where a(F) is the average distance between F and other points in the same cluster, and b(F) is the distance between F and the nearest other cluster center. The K value that maximizes SC(K) is selected as the final number of clusters.

[0128] When executing the K-means++ algorithm, first use the above method to select the initial cluster center, then assign each data point to the nearest cluster center, and recalculate the center of each cluster (that is, the average position of all points in the cluster). Repeat this assignment and recalculation process until the cluster center no longer changes significantly or reaches the preset maximum number of iterations. In practical applications, it may be necessary to run the K-means++ algorithm multiple times (such as 100 times), each time using a different random seed, and then select the result with the smallest objective function J value as the final clustering result. This can reduce the risk of the algorithm falling into a local optimum.

[0129] After executing the clustering algorithm, three fine clusters C1, C2, and C3 (i.e., the second type of clustering set) may be obtained in the curve area, representing the running posture characteristics of the curve entry, curve middle, and curve exit stages, respectively. In order to map these clustering results back to the physical space of the runway, the fine sub-area Sk is defined as follows: This means that if there is a data point belonging to cluster Ck with position coordinate l, then l belongs to the refined subregion Sk.

[0130] By this method, the fine division results S1, S2, S3 of the curve area are obtained. Each fine sub-area Sk is a continuous interval or the union of several continuous intervals on the runway. For example:

[0131] S1=[150m,180m)∪[350m,380m) indicates the cornering area.

[0132] S2=[180m,220m)∪[380m,420m) represents the middle section of the curve.

[0133] S3=[220m,250m)∪[420m,450m) indicates the exit from the curve.

[0134] A similar fine clustering process is performed for other sub-areas (starting area, straight area, and finish area). In the end, the target clustering result obtained includes the fine division of all sub-areas. For example, it may be possible to obtain: the "starting stage" and "acceleration stage" of the starting area; the "acceleration section" and "steady speed section" of the straight area; the "entry zone", "mid-curve section" and "exit zone" of the curve area; the "final sprint section" and "deceleration section" of the finish area.

[0135] Step S150, obtaining typical posture problems and posture data density distribution corresponding to the initial sub-region according to the refined sub-region, and determining the main problem region according to the typical posture problems and posture data density distribution.

[0136] In one example, according to the fine sub-region, the typical posture problems and posture data density distribution corresponding to the initial sub-region are obtained, and the steps of determining the main problem area according to the typical posture problems and posture data density distribution include the steps of: using the posture scoring function to calculate the posture score corresponding to the data points in the fine sub-region; calculating the mean and standard deviation of the posture features in the fine sub-region, and obtaining the abnormal threshold according to the mean and standard deviation; taking the posture features corresponding to the posture score greater than the abnormal threshold as the typical posture problem; using the kernel density estimation method to calculate the posture data density distribution of the fine sub-region; calculating the spatial distribution of the problem severity corresponding to the fine sub-region according to the posture data density distribution and the weight of the typical posture problem; sorting the spatial distribution of the problem severity from large to small, and taking the first part of the sorting result as the main problem area.

[0137] 1. Identify typical posture problems:

[0138] For each data point (P, v, a, t, l, F) in the fine sub-region Sk, the pose score function G(P, v, a) is used to calculate its pose score:

[0139] G(P,v,a)=w1*G1(P)+w2*G2(v)+w3*G3(a)+...+wn*Gn(P,v,a)

[0140] Among them, G1, G2, G3, ..., Gn evaluate different posture features, such as trunk tilt, stride, arm swing, etc. The output of each subfunction is normalized to the interval [0, 1], where 0 represents the worst posture and 1 represents the best posture. The weights w1, w2, w3, ..., wn reflect the importance of each feature in the overall score, and the sum of these weights is equal to 1.

[0141] For each posture feature, calculate its mean and standard deviation in the sub-region:

[0142] ;

[0143] The abnormal threshold is defined as μ_feature±2σ_feature. If the posture score of a posture feature exceeds this range, it is considered that there is a typical problem with the feature.

[0144] Define a set of typical posture problems F={f1,f2,...,fm}, where each fi represents a specific posture problem (such as excessive torso tilt, uneven stride, etc.).

[0145] For each location l, define the indicator function I_feature(l):

[0146] I_feature(l)=1, if there is a specific posture problem fi at position l;

[0147] 0, otherwise.

[0148] 2. The kernel density estimation (KDE) method is used to calculate the density distribution of the posture data. This is a non-parametric method for estimating the probability density function of a random variable. For each position l in the fine sub-region Sk, the density estimation formula is:

[0149] D(l)=(1 / (n*h))*Σ(K((l-li) / h))

[0150] In this formula, n represents the total number of data points in the fine sub-region Sk, h is the bandwidth parameter, K is the kernel function, and li represents the position of the i-th data point. The summation symbol Σ represents the accumulation of all data points in Sk. This formula actually calculates the contribution of each data point to the position l, then accumulates all contributions, and finally divides by a normalization factor (n*h).

[0151] Usually the Gaussian kernel is chosen as the kernel function, and its expression is:

[0152]

[0153] The Gaussian kernel function is a bell-shaped curve that has a maximum value at the center and then decays rapidly to both sides. In density estimation, each data point is regarded as the center of a Gaussian distribution, and the superposition of these distributions forms the final density estimate.

[0154] The choice of bandwidth parameter h has an important impact on the result of density estimation. The h value can be understood as the "width" of each Gaussian distribution. The smaller the h value, the sharper the density estimation is, and it can capture more subtle changes; the larger the h value, the smoother the density estimation is, and it can resist the influence of noise, but it may ignore some local features.

[0155] In order to select the optimal h value, a cross-validation method is used. This method tries a series of different h values ​​and then selects the one with the best performance. Specifically, for a series of candidate h values, the leave-one-out cross-validation score is calculated:

[0156] CV(h)=(1 / n)*Σ(log(D_-i(li)))

[0157] Where D_-i(li) is the density estimate of the position li after removing the i-th point. This formula can be understood as removing one data point at a time, then using the remaining points to estimate the density of the removed point, and finally taking the logarithmic average of all these estimates. The h value that maximizes CV(h) is selected as the final bandwidth parameter.

[0158] This process actually simulates the situation where data points are repeatedly sampled, and the stability and generalization ability of the model are evaluated by calculating the density estimate after each sampling. The larger the CV(h) value, the stronger the predictive ability of the model.

[0159] The final density estimate D(l) is a continuous function that describes the distribution of data in the entire fine sub-region Sk. In high-density areas, the value of D(l) will be larger; in low-density areas, the value of D(l) will be smaller. This density estimate can help identify which areas on the runway have the most concentrated data points, thus providing an important reference for determining the main problem areas.

[0160] 3. Identify major problem areas:

[0161] Combining typical posture problems and posture data density distribution, we can determine the main problem areas. First, we define the problem severity index P(l):

[0162] P(l)=D(l)*Σ(w_feature*I_feature(l))

[0163] in:

[0164] -D(l) is the density distribution of the posture data calculated in step 2;

[0165] -w_feature is the weight of each typical posture problem, reflecting the relative importance of different posture problems;

[0166] - I_feature(l) is the indicator function defined in step S140, indicating whether there is a specific posture problem at position l;

[0167] By calculating the P(l) value at each position, we can get the spatial distribution of the problem severity. We sort the P(l) values ​​and select the top (for example, the top 20%) regions as the main problem regions Pmain. These regions have both high data density and obvious posture problems, and are the areas that should be focused on in subsequent feedback and training.

[0168] Through the above analysis, a detailed feature description of each fine sub-area is obtained. Take the curve area as an example:

[0169] Sub-areas: bend area (S3);

[0170] Location range: [150m, 250m) ∪ [350m, 450m);

[0171] Typical posture problems:

[0172] -f1: excessive trunk tilt (average tilt angle: 12°, normal range: 5°-8°);

[0173] -f2: uneven stride length (coefficient of variation of stride length: 0.15, normal range: <0.1);

[0174] Attitude data density distribution: The density is highest at the cornering (150m-170m and 350m-370m), reaching 1.5 times the average density;

[0175] Main problem areas: The first 10 meters into a turn (150m-160m and 350m-360m).

[0176] Step S160 , determining candidate feedback positions in the fine sub-region according to the posture data density distribution and the problem severity index corresponding to the main problem region.

[0177] In one example, the step of determining candidate feedback positions in a fine sub-region according to the posture data density distribution and the problem severity index corresponding to the main problem region includes the steps of: performing weighted processing on the reaction time function, the posture data density distribution and the problem severity index corresponding to the main problem region to obtain a feedback position scoring function; wherein the reaction time function indicates how much time and space the student has to adjust the posture after providing feedback at a certain position; using the feedback position scoring function to calculate the feedback position score corresponding to the position in the fine sub-region; and obtaining the local maximum value in the feedback position score based on a sliding window algorithm;

[0178] The position corresponding to the local maximum value greater than the preset threshold is taken as the primary candidate feedback position; the preset threshold is set according to the average value and standard deviation of the feedback position score; the primary candidate feedback positions are sorted from large to small, and the distance between any two primary candidate feedback positions in the sorting result is calculated, and when the distance is less than the preset interval, the primary candidate feedback position with the smallest local maximum value among the two primary candidate feedback positions is deleted to obtain the secondary candidate feedback position; according to the length of the fine sub-region and the preset maximum number of feedback positions, the first W secondary candidate feedback positions from large to small are taken as candidate feedback positions.

[0179] Specifically, in order to determine the candidate feedback positions, a feedback position scoring function F(l) is defined to evaluate the suitability of each position as a feedback point:

[0180] F(l)=α*D(l)+β*P(l)+γ*R(l)

[0181] Where α, β, and γ are weight coefficients used to balance the importance of different factors, and α+β+γ=1. D(l) is the density distribution of posture data calculated in step S150, and P(l) is the severity index calculated in step S150. R(l) is a newly introduced reaction time function, which indicates how much time and space the student has to adjust the posture after providing feedback at position l. R(l) is defined as:

[0182] R(l)=min(1,(Ll) / d)

[0183] Here L is the end position of the sub-area, and d is the preset reaction distance (e.g., 10 meters). This function ensures that the feedback position is not too close to the end of the sub-area, leaving enough room for students to adjust.

[0184] The specific steps for determining candidate feedback locations are as follows:

[0185] First, for each position l in the fine sub-region Sk, F(l) is calculated. This step requires traversing all positions in the sub-region, and it may be necessary to discretize continuous positions, for example, calculating the F(l) value every 0.1 meter. This ensures that potential important feedback points are not missed. In actual implementation, the discretization interval can be adjusted according to computing resources and accuracy requirements.

[0186] Discretization is the process of converting a continuous space into a finite set of discrete points. This process is necessary for numerical calculations because computers cannot process truly continuous data. Choosing an appropriate discretization interval requires a trade-off between computational accuracy and computational efficiency. An interval that is too large may miss some important local features, while an interval that is too small will increase the computational burden. An interval of 0.1 meters usually provides sufficient accuracy while keeping the amount of computation within an acceptable range.

[0187] During this discretization process, F(l) will be calculated repeatedly. Each calculation requires the evaluation of three functions: D(l), P(l), and R(l). D(l) reflects the data density at position l. High-density areas may mean that this is where students often have problems, so they may need more feedback. P(l) directly reflects the severity of the problem at position l. Positions with high P(l) values ​​obviously need more feedback. R(l) takes into account the student's reaction time after receiving feedback, ensuring that the student has enough time and space to adjust their posture.

[0188] The second step is to find the local maximum point of F(l). This can be achieved by comparing the F(l) value of each point with its neighboring points. If the F(l) value of a point is greater than the F(l) values ​​of all points within the range of δ / 2 before and after it, then the point is considered to be a local maximum point. Here δ is the preset minimum feedback point interval, which is used to prevent the feedback points from being too dense. When choosing the δ value, it is necessary to consider the student's reaction time and running speed to ensure that the student has enough time to respond to each feedback.

[0189] The process of detecting local maximum points can be imagined as sliding a window of width δ on the F(l) curve. When the F(l) value at the center of the window is greater than the F(l) values ​​of all other points in the window, a local maximum point is considered to have been found. This process requires traversing all discrete points and performing a window comparison on each point. Therefore, if the length of the sub-region is L and the discretization interval is 0.1 meters, then approximately 10L window comparisons are required.

[0190] The third step is to apply a peak detection algorithm, such as threshold-based peak detection. A preset threshold θ is set, and l is considered as a primary candidate feedback position only when F(l)>θ and F(l) is a local maximum. The selection of the preset threshold θ can be based on the mean and standard deviation of F(l), for example, it can be set to the mean plus a standard deviation. This ensures that the selected feedback point has a sufficiently high importance. The selection of the threshold requires a trade-off between the quantity and quality of the feedback points. A too high threshold may cause important feedback points to be ignored, while a too low threshold may introduce too many unnecessary feedback points.

[0191] The calculation of the preset threshold requires the calculation of the mean and standard deviation of F(l). The mean requires the sum of all F(l) values ​​and then divide by the total number of points, while the standard deviation requires the calculation of the variance (the average of the squares of the difference between each F(l) value and the mean) and then taking the square root. This process requires traversing all F(l) values ​​twice, once to calculate the mean and once to calculate the standard deviation.

[0192] After setting the preset threshold, it is necessary to traverse all local maximum points again to check whether they exceed the threshold. This process will further filter the candidate feedback positions and only retain those with sufficiently high F(l) values. This ensures that the selected feedback point is not only the best choice locally, but also has sufficient importance overall.

[0193] The fourth step is to further screen the detected primary candidate feedback positions to ensure that the distance between adjacent primary candidate feedback positions is not less than the minimum interval δ. This can be achieved through the following iterative process: first sort the primary candidate feedback positions from large to small according to the F(l) value, and then start from the primary candidate feedback position with the highest score and delete other candidate feedback positions with a distance less than δ from it. Repeat this process for the remaining candidate feedback positions until the interval between all adjacent candidate feedback positions is not less than δ to obtain the secondary candidate feedback positions. This step ensures that there is enough interval between the feedback points to avoid too frequent feedback and interfere with the student's performance.

[0194] This process first requires sorting the primary candidate feedback positions, which is usually done using an efficient sorting algorithm such as quick sort or merge sort. After sorting, start with the candidate feedback position with the highest score and check the other candidate feedback positions one by one. If a position with a distance less than δ is found, it is deleted from the candidate list. This process needs to be repeated many times until no candidate feedback positions need to be deleted. In the worst case, if all candidate feedback positions are concentrated in a small area, then this process may require deleting most of the candidate feedback positions, leaving only the one with the highest score to obtain the secondary candidate feedback position.

[0195] The last step is to select the top W candidate feedback positions with the highest scores according to the length of the refined sub-region and the preset maximum number of feedback positions M, where . This step ensures that the number of feedback points is not too large while taking into account the length of the sub-region. Through this process, a set of candidate feedback positions Ck={c1,c2,...,cN} for each fine sub-region Sk can be obtained. These candidate feedback positions represent the most suitable locations for providing feedback in the sub-region. Their selection takes into account factors such as data density, problem severity, and response time, so as to maximize the effect of feedback.

[0196] This final selection process only requires comparing two values , and then select the smaller one as the final number of candidate feedback positions. If the candidate feedback positions have been sorted from large to small according to the F(l) value in the previous step, then the first W positions can be selected.

[0197] Step S170 , comprehensively evaluate the candidate feedback positions, and select the candidate feedback position with the highest comprehensive evaluation score and covering the main problem area as the feedback system deployment position.

[0198] In one example, a step of comprehensively evaluating candidate feedback locations and selecting a candidate feedback location with the highest comprehensive evaluation score and covering a major problem area as a feedback system deployment location includes the following steps: calculating the coverage effect score, problem targeting score, feedback timeliness score, and environmental adaptability score of the candidate feedback location; weighting the coverage effect score, problem targeting score, feedback timeliness score, and environmental adaptability score to obtain a comprehensive evaluation score for the candidate feedback location; sorting the candidate feedback locations from large to small according to the comprehensive evaluation score to obtain a sorted list; and obtaining the feedback system deployment location based on the coverage function and the sorted list; wherein the coverage function represents the coverage degree of the selected location set to the major problem area.

[0199] Specifically, for the candidate feedback position set Ck={c1,c2,...,cN} in each fine sub-region Sk, a comprehensive scoring function E(c) is defined to evaluate the suitability of each candidate feedback position c:

[0200] E(c)=w1*Ecov(c)+w2*Eprob(c)+w3*Etime(c)+w4*Eenv(c)

[0201] Among them, w1, w2, w3, and w4 are weight coefficients used to balance the importance of different factors, and w1+w2+w3+w4=1. These weights can be adjusted according to actual needs to reflect the relative importance of different factors in specific application scenarios.

[0202] The coverage effect score Ecov(c) measures the coverage of the candidate feedback position c on the main problem area:

[0203] Ecov(c)=|A(c)∩Pmain| / |Pmain|

[0204] Where A(c) is the influence area of ​​candidate feedback position c, and Pmain is the main problem area determined in step 5. |·| represents the area or length of the set. The definition of the influence area A(c) may depend on the characteristics of the feedback device, such as the field of view of a visual feedback system or the effective sound propagation range of an audio feedback system. This coverage effect metric ensures that the selected feedback positions can target the problem areas that need the most attention.

[0205] The problem-targeting score Eprob(c) evaluates the ability of the candidate feedback position c to target typical posture problems:

[0206] Eprob(c)=Σ(w_feature*I_feature(c)) / Σw_feature

[0207] Where w_feature is the weight of each posture feature, and I_feature(c) is the indicator function, which takes 1 when the position c can address the feature problem, otherwise it takes 0. This indicator takes into account the importance of different posture problems and ensures that the selected feedback position can address the most critical posture problem. For example, if multiple important posture problems can be detected at a certain position at the same time, then the Eprob(c) value of this position will be higher.

[0208] The feedback timeliness score Etime(c) considers the likelihood that the student will adjust their posture after feedback is provided at position c:

[0209] Etime(c)=min(1,(Lc) / d)

[0210] Where L is the end position of the sub-area, and d is the preset reaction distance. This function ensures that the selected feedback position is not too close to the end of the sub-area, giving students enough time and space to adjust their posture. If the feedback position is too close to the end of the sub-area, students may not have enough time to respond to the feedback and adjust their posture.

[0211] The environmental adaptability score Eenv(c) evaluates the impact of environmental conditions at location c on the performance of the feedback system:

[0212] Eenv(c)=1-|H(c)-Hideal| / Hmax

[0213] Where H(c) is the environmental factor score at position c, Hideal is the ideal environmental score, and Hmax is the maximum possible difference in environmental scores. This metric takes into account environmental factors such as lighting conditions and background noise to ensure that the selected feedback position can work effectively in the actual environment. For example, if the lighting conditions at a certain location are particularly poor, which may affect the effectiveness of the visual feedback system, then the Eenv(c) value of this location will be lower.

[0214] After calculating the comprehensive evaluation score E(c) of each candidate feedback location, the final feedback system layout location selection is performed. This process not only takes into account the scores of individual locations, but also ensures that the selected location can cover the main problem area determined in step S150. The following strategy is adopted:

[0215] First, the candidate feedback positions in each fine sub-region Sk are sorted from high to low according to E(c), and a sorted list Lk={lk1,lk2,...,lkN} is obtained. This sorting process ensures that the positions with the highest scores are considered first.

[0216] Then, a coverage function Cov(S) is defined to indicate the coverage of the selected location set S to the main problem area Pmain:

[0217] Cov(S)=|∪(c∈S)A(c)∩Pmain| / |Pmain|

[0218] This coverage function calculates the intersection of the impact area of ​​all selected locations with the main problem area, and then divides it by the total area or length of the main problem area to get a coverage ratio between 0 and 1.

[0219] Next, initialize the selection set S to an empty set, and then select the feedback system layout location according to the following steps:

[0220] 1. For each refined sub-region Sk, select positions c from its sorted list Lk in sequence.

[0221] 2. If Cov(S∪{c})-Cov(S)≥θ, where θ is the preset minimum gain threshold, then add c to S.

[0222] 3. Repeat steps 1 and 2 until one of the following conditions is met:

[0223] a) Cov(S) = 1, which means that the main problem area is completely covered.

[0224] b) |S|=M, where M is the preset maximum number of feedback points.

[0225] c) Unable to find a new position that satisfies the gain threshold condition.

[0226] This process ensures that the selected locations not only have high scores but also effectively cover the main problem areas. The setting of the minimum gain threshold θ can avoid selecting locations that only bring minor improvements, thereby controlling the total number of feedback points. For example, if θ is set to 0.05, then a newly added location will only be selected if it can increase the coverage by at least 5%. The setting of the maximum number of feedback points M prevents the system from becoming too complex, ensuring the manageability and practicality of the system.

[0227] The finally selected feedback system layout positions constitute the set S = {c1, c2, ..., cm}, where ci (i = 1, 2, ..., m) represents the i-th selected feedback system layout position, and m is the total number of feedback points finally selected.

[0228] Step S180, taking the feedback system placement position as the center and combining the posture data density distribution, determine the effective feedback range of the feedback system placement position.

[0229] In one example, the steps of determining the effective feedback range of the feedback system layout position with the feedback system layout position as the center and in combination with the posture data density distribution include the steps of: smoothing the posture data density distribution; setting a corresponding initial search range for the feedback system layout position; within the initial search range, determining the primary effective feedback range using the smoothed posture data density distribution; determining whether there is an overlapping area between adjacent effective feedback ranges, and if there is an overlapping area, adjusting the adjacent primary effective feedback ranges based on the overlapping area to obtain the effective feedback range of the feedback system layout position.

[0230] First, the original posture data density distribution D(l) is smoothed to reduce the impact of local fluctuations and obtain a more stable density distribution. The moving average method is used to calculate the smoothed density distribution D'(l):

[0231] D'(l)=(1 / (2k+1))*Σ(j=-ktok)D(l+j*Δl)

[0232] Where k is the half width of the smoothing window and Δl is the minimum resolution of the position. The size of the smoothing window needs to be adjusted according to the characteristics of the actual data. Usually k=5 and Δl=0.1 meters can be selected. This smoothing process helps to eliminate noise and short-term fluctuations in the data, making the subsequent range determination more stable and reliable.

[0233] Next, for each feedback system placement ci, set an initial search range [ci-d, ci+d], where d is a preset distance, such as 10 meters. The selection of this initial range takes into account the average speed and reaction time of the students, while avoiding unnecessary computational burden. The setting of the initial search range provides a reasonable starting point for subsequent precise calculations, which not only ensures sufficient search space but also limits the computational complexity.

[0234] Within this initial search range, the smoothed posture data density distribution D'(l) is used to determine the actual effective feedback range. Starting from position ci, expand to both sides until the density value drops below 50% of the density of position ci. The selection of this threshold is based on the assumption that when the density drops below half, the data points become relatively sparse and may no longer be suitable as an effective feedback area. Specifically, the process of determining the effective feedback range R(ci) is as follows:

[0235] Search forward:

[0236] R_forward(ci)=max{l|ci≤l≤ci+d,D'(l)≥0.5*D'(ci)}

[0237] Search backward:

[0238] R_backward(ci)=min{l|ci-d≤l≤ci,D'(l)≥0.5*D'(ci)}

[0239] This search method ensures that the effective feedback range is concentrated in the area with relatively high data density, which improves the pertinence and effectiveness of the feedback. The preliminary determined primary effective feedback range R(ci) is [R_backward(ci), R_forward(ci)].

[0240] After determining the primary effective feedback range of each feedback system layout location, it is necessary to consider the possible range overlap between adjacent feedback system layout locations. Range overlap may lead to feedback redundancy or interference, so adjustments are required. For two adjacent feedback system layout locations ci and ci+1, if R_forward(ci)>R_backward(ci+1), make the following adjustments:

[0241] R_forward_adjusted(ci)=(R_forward(ci)+R_backward(ci+1)) / 2

[0242] R_backward_adjusted(ci+1)=R_forward_adjusted(ci)

[0243] This adjustment method evenly distributes the overlapping area to two adjacent feedback system deployment locations, which not only ensures the continuity of the feedback range, but also avoids the range of a single feedback system deployment location being too large. In this way, a balanced feedback coverage network can be formed on the entire runway.

[0244] For the adjacent feedback system locations near the start and end of the runway, special processing is required to ensure complete coverage. For the adjacent feedback system location c1 closest to the start, its backward effective range is extended to the start of the runway: R_backward_adjusted(c1)=0. For the feedback system location cm closest to the end, its forward effective range is extended to the end of the runway: R_forward_adjusted(cm)=L, where L is the total length of the runway. This processing ensures that the entire runway from the start to the end is covered by the effective feedback range, and there will be no blind spots.

[0245] After all adjustments are completed, the final effective feedback range R(ci) of each feedback system placement position ci is determined as [R_backward_adjusted(ci), R_forward_adjusted(ci)]. This range represents the area around the feedback system placement position where the attitude data density is relatively high and does not overlap with the ranges of other feedback system placement positions.

[0246] Finally, the output of step S180 is a series of optimized effective feedback ranges, each of which is associated with a specific feedback system deployment position. These outputs will be directly used for the specific implementation of the feedback system in the next step, providing an accurate range of action for real-time running posture feedback.

[0247] Step S190, deploy the feedback system at the feedback system deployment location, set a corresponding trigger area for the feedback system deployment location according to the effective feedback range, and when the student enters the trigger area, trigger the feedback system to provide running posture feedback to the student based on typical posture problems.

[0248] First, a feedback system is deployed at each feedback system deployment location ci. This feedback system usually includes a high-speed camera, a data processing unit, and a feedback output device. The installation position and angle of the high-speed camera need to be carefully designed to ensure that the student's full-body movements within the effective feedback range R(ci) can be fully captured. The resolution and frame rate of the camera should be high enough, for example, a device with a resolution of not less than 1920x1080 pixels and a frame rate of at least 120fps can be selected to ensure that the details of the student's rapidly changing movements are captured. The data processing unit needs to be equipped with sufficient computing power to process the image data captured by the high-speed camera in real time. This unit usually includes a high-performance CPU and GPU to run specially designed action recognition algorithms. These algorithms can be based on deep learning models, such as convolutional neural networks (CNN) or long short-term memory networks (LSTM), which are pre-trained to quickly and accurately identify and locate key points of the human body, such as the head, shoulders, elbows, wrists, hips, knees, and ankles, from the video stream. The feedback output device can be selected in different forms according to actual needs. For example, an LED display can be used to provide visual feedback, or a speaker system can be used to provide audio feedback. In some cases, you can also consider using wearable devices, such as smart bracelets or headphones, to provide students with more personalized feedback.

[0249] In one example, a feedback system is deployed at a feedback system deployment location, a corresponding trigger area is set for the feedback system deployment location according to an effective feedback range, and when a student enters the trigger area, the feedback system is triggered to provide the student with running posture feedback based on typical posture problems, including the steps of: scoring the student's current posture based on a real-time posture evaluation function to obtain a current posture score; when the current posture score exceeds a trigger threshold corresponding to the trigger area, providing the student with running posture feedback based on a feedback strategy corresponding to the typical posture problem.

[0250] According to the effective feedback range R(ci) determined in step S180, a trigger area is set for each feedback system layout position. The trigger area T(ci) is defined as:

[0251] T(ci)=[R_backward(ci),R_backward(ci)+min(5m,0.3*|R(ci)|)]

[0252] Where |R(ci)| represents the length of the effective feedback range. This setting ensures that the trigger area is located at the front of the effective feedback range, so that the system has enough time to recognize the student's posture and generate feedback. The starting point of the trigger area R_backward(ci) is the starting point of the effective feedback range, and the end point is 5 meters backward from the starting point or 30% of the length of the effective feedback range (the smaller of the two). This setting takes into account the system response time and avoids the problem of untimely feedback caused by a too long trigger area.

[0253] Next, based on the typical posture problems identified in step S150, a targeted feedback strategy is designed for each feedback system placement location ci. These strategies should consider the most common posture problems and their severity in that area. For example, if the most common problem in a certain area is excessive torso tilt, then the feedback system should focus on the student's torso angle.

[0254] Design a real-time posture evaluation function E(P,v,a), where P is the student's current posture (including the position of each key point), v is the velocity, and a is the acceleration. This function should be able to quickly evaluate the student's current posture and identify existing problems. Design E(P,v,a) as a weighted sum of multiple sub-functions:

[0255] E(P,v,a)=w1*E1(P)+w2*E2(v)+w3*E3(a)+...+wn*En(P,v,a)

[0256] Each sub-function Ei evaluates a specific pose feature and wi is the corresponding weight. These weights can be set according to the importance of typical problems identified in step 5.

[0257] Specifically, E1(P) can evaluate the torso inclination, which is quantified by calculating the angle between the torso and the vertical line. E2(v) can evaluate the cadence, which is measured by calculating the number of steps per unit time. E3(a) can evaluate the acceleration change, which is used to determine the smoothness of running by calculating the degree of acceleration fluctuation. Other sub-functions can evaluate aspects such as arm swing, knee lift height, and landing posture.

[0258] The design of each sub-function should take into account the normal range and abnormal threshold of the feature, which may need to be adjusted according to the age and physical fitness level of the students. For example, for the trunk tilt E1(P), the normal range can be set to 0-10 degrees, 10-20 degrees is mildly abnormal, and more than 20 degrees is severely abnormal. The output of the sub-function can be a normalized value between 0 and 1, where 0 means completely normal and 1 means severely abnormal.

[0259] The setting of weights wi should reflect the importance of each posture feature in the overall evaluation. This can be based on the experience of physical education teachers or determined by analyzing a large amount of student running data. For example, if research shows that trunk posture is most important for student running efficiency and injury prevention, then w1 can be set to a larger value, such as 0.3 or 0.4.

[0260] When the student enters the trigger area T(ci), the system begins to continuously evaluate the student's posture. If the value of E(P, v, a) exceeds the preset threshold θ, the feedback mechanism is triggered. The setting of the threshold θ requires a trade-off between sensitivity and specificity, and the optimal value can be determined by analyzing the student's running data. For example, θ can be set to 0.6, which means that feedback is triggered when the overall posture score exceeds 60% of the abnormality.

[0261] The specific content of the feedback should be determined by the specific sub-function that causes E(P,v,a) to exceed the threshold. For example, if the main reason is that the value of E1(P) (assessing the torso inclination) is too high, then the feedback content should be guidance on torso posture. The system can maintain a feedback content library, preset multiple feedback methods for each posture problem, and select appropriate feedback content according to the severity of the problem.

[0262] The form and content of feedback should be concise and clear, so that students can quickly understand and apply it during running. For visual feedback, simple graphic symbols can be used, such as an upward arrow for "head up" and an inward arrow for "belly in". These symbols should be large and eye-catching enough to ensure that students can see them clearly even at high speeds. For audio feedback, short voice instructions such as "straighten your body" and "lift your feet" can be used. These instructions should be concise, clear, and use an appropriate volume to ensure that they can be clearly heard in a sports environment.

[0263] The frequency of feedback needs to be properly controlled to avoid too frequent feedback interfering with students' normal performance. A minimum feedback interval can be set, such as 3-5 seconds, to ensure that students have enough time to understand and apply feedback. At the same time, the system should also be able to identify whether the student has responded to the feedback. If the system detects that the student's posture has improved after giving feedback, the interval between the next feedback can be extended to avoid unnecessary repeated reminders.

[0264] In this way, the motion recognition system deployed at each selected feedback system layout location can provide real-time and targeted running posture feedback for students entering the trigger area based on the predetermined effective feedback range and trigger area and the typical posture problems in the area.

[0265] Example

[0266] A primary school installed this running posture feedback system on its 400-meter standard track. The following is the process of system installation and operation:

[0267] Step 1-2: Data collection and preliminary analysis.

[0268] Ten high-speed cameras were installed around the 400-meter standard track, covering the entire track. In the following week, the system collected video data of 200 students running at different time periods. Through analysis, the system found that students were most likely to have posture problems at the first bend (150 meters to 200 meters), and this area was identified as the high-incidence area of ​​abnormal posture R_high.

[0269] Step 3-4: Runway area division and fine clustering.

[0270] Based on R_high, the system initially divides the track into several main areas: the starting area (0-50 meters), the first straight (50-150 meters), the first curve (150-250 meters), the second straight (250-350 meters) and the last straight (350-400 meters). The system then performs a fine cluster analysis on these areas. In particular, for the first curve area, the system further subdivides it into the entry area (150-180 meters), the middle of the curve (180-220 meters) and the exit area (220-250 meters).

[0271] Step 5: Identify typical problems and density distribution.

[0272] The system uses the posture scoring function G(P,v,a) to analyze the data of each sub-area. In the cornering area (150-180 meters), the system found that the most common problem of students was excessive torso tilt, with an average tilt angle of 12 degrees (the normal range should be 5-8 degrees). At the same time, the posture data density D(l) in this area is the highest, which is 1.5 times the average density. The system identified this area as part of the main problem area Pmain.

[0273] Step 6: Determine candidate feedback system deployment locations.

[0274] The system uses the feedback system placement scoring function F(l) to evaluate each position in the cornering area. Through calculation and comparison, the system preliminarily determines several candidate feedback system placement positions, among which the F(l) value at 155 meters is the highest.

[0275] Step 7: Comprehensive evaluation and final selection.

[0276] The system uses a comprehensive scoring function to evaluate candidate locations. The location at 155 meters not only performs well in terms of coverage effect Ecov(c) and problem targeting Eprob(c), but also performs well in terms of feedback timeliness Etime(c) and environmental adaptability Eenv(c). After comprehensive consideration, the system finally decided to place the feedback point at 155 meters.

[0277] Step 8: Determine the effective feedback range.

[0278] Based on the feedback point position at 155 meters, the system calculates that the effective feedback range R(c) of this feedback point is 150 meters to 165 meters. This range takes into account the data density distribution D(l) and ensures monitoring in areas with higher density.

[0279] Step 9: Deploy the feedback system.

[0280] Based on the analysis results of the previous steps, the school installed an LED display and a small speaker as feedback devices at 155 meters. The system sets the starting point of the trigger area T(c) at 150 meters. When the students run to 150 meters, the system starts to analyze their posture in real time. If it is detected that the student's torso tilts more than 10 degrees, the LED screen will display a text prompt of "Stay upright" and the speaker will issue a short voice reminder of "Stand up straight".

[0281] The present application comprises the following steps: collecting first video data of a student running in a target track area, extracting coordinates of key body points of the student from the first video data, and constructing a first running posture data set; analyzing the first running posture data set to determine high-incidence areas of posture abnormalities; performing a first clustering process on the first running posture data set based on the high-incidence areas of posture abnormalities, and dividing the target track area into M initial sub-areas according to the first clustering process result; collecting second video data of a student running in the initial sub-area, extracting coordinates of key body points of the student from the second video data, and constructing a second running posture data set; performing a second clustering process on the first running posture data set and the second running posture data set together, and dividing the initial sub-area into N refined sub-areas according to the second clustering process result; and obtaining, according to the refined sub-areas, Typical posture problems and posture data density distribution corresponding to the initial sub-area, and the main problem area is determined based on the typical posture problems and posture data density distribution; candidate feedback positions are determined in the fine sub-area based on the posture data density distribution and the problem severity index corresponding to the main problem area; the candidate feedback positions are comprehensively evaluated, and the candidate feedback positions with the highest comprehensive evaluation score and covering the main problem area are selected as the feedback system layout position; the feedback system layout position is taken as the center, and the effective feedback range of the feedback system layout position is determined in combination with the posture data density distribution; the feedback system is deployed at the feedback system layout position, and the corresponding trigger area is set for the feedback system layout position according to the effective feedback range. When the student enters the trigger area, the feedback system is triggered to provide running posture feedback to the student based on the typical posture problem. This application analyzes the dynamic three-dimensional movements of students running based on video data. In the process of video analysis, a specific runway area is analyzed in combination with the running process, so that the running posture analysis is more targeted. In addition, in the process of analyzing the running posture, the differences of the students themselves are considered, so that the analysis results are closer to the actual situation of each student. Finally, the student's running posture can be reminded and adjusted in real time.

[0282] 1. Refined track area division and cluster analysis: This solution innovatively proposes a method of track area division and refined cluster analysis in step S130 and step S140. This method takes into account the impact of the characteristics of different areas of the track (such as the starting area, the straight area, and the curve area) on the running posture. By performing independent cluster analysis on each sub-area, the system can capture the posture characteristics and problems unique to different areas. For example, in the curve area, more attention may be paid to the changes in body inclination and stride frequency, while in the straight area, more attention is paid to the consistency of stride and speed. This refined analysis greatly improves the pertinence and accuracy of posture assessment.

[0283] 2 Dynamically adaptive feedback position selection: The feedback position selection method proposed in step S160 and step S170 is highly innovative. This method not only takes into account the density distribution of posture data and the severity of the problem, but also introduces the reaction time function R(l) to ensure that students have enough time to respond to feedback. More importantly, by defining the comprehensive scoring function E(c), the system can dynamically balance multiple factors such as coverage effect, problem targeting, feedback timeliness and environmental adaptability. This method enables the layout of the feedback system to be optimized according to actual conditions to maximize the effectiveness of feedback. For example, in some areas where problems occur frequently, the system may choose a denser layout of feedback points; while in relatively stable areas, it may choose fewer feedback points. This dynamic adaptive method greatly improves the flexibility and practicality of the system.

[0284] 3 Real-time personalized trigger feedback mechanism: The trigger feedback mechanism designed in step S190 shows another innovation of this solution. The system realizes personalized real-time monitoring of each student by setting the trigger area T(ci) and the real-time posture evaluation function E(P,v,a). This mechanism not only considers static posture, but also incorporates dynamic factors such as speed and acceleration, making the evaluation more comprehensive. When the student enters the trigger area and the posture score exceeds the threshold, the system triggers feedback to avoid unnecessary interference.

[0285] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0286] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 2 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as video data and running posture data sets. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a real-time feedback method for students' running posture based on motion recognition is implemented.

[0287] Those skilled in the art will understand that Figure 2 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0288] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0289] Collecting first video data of a student running in a target track area, and constructing a first running posture data set according to the first video data;

[0290] Analyze the first running posture data set to determine high incidence areas of abnormal posture;

[0291] Based on the high-incidence areas of abnormal posture, a first clustering process is performed on the first running posture data set, and the target runway area is divided into M initial sub-areas according to the first clustering process result;

[0292] Collecting second video data of running in the initial sub-region, extracting coordinates of key points of the student's body from the second video data, constructing a second running posture data set, performing a second clustering process on the first running posture data set and the second running posture data, and dividing the initial sub-region into N fine sub-regions according to the second clustering process result;

[0293] According to the refined sub-region, typical posture problems and posture data density distribution corresponding to the initial sub-region are obtained, and the main problem area is determined according to the typical posture problems and posture data density distribution;

[0294] Determine candidate feedback locations in the fine sub-region based on the posture data density distribution and the problem severity index corresponding to the main problem area;

[0295] Conduct a comprehensive evaluation of candidate feedback locations, and select the candidate feedback location with the highest comprehensive evaluation score and covering the main problem area as the location for the feedback system deployment;

[0296] Taking the feedback system layout location as the center and combining the attitude data density distribution, determine the effective feedback range of the feedback system layout location;

[0297] The feedback system is deployed at the feedback system layout location, and a corresponding trigger area is set for the feedback system layout location according to the effective feedback range. When the student enters the trigger area, the feedback system is triggered to provide the student with running posture feedback based on typical posture problems.

[0298] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0299] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0300] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A method for real-time feedback of students' running posture based on motion recognition, characterized in that: The following steps are involved: Collecting first video data of a student running in a target track area, extracting coordinates of key points of the student's body from the first video data, and constructing a first running posture data set; Analyze the first running posture data set to determine high-incidence areas of abnormal posture; Based on the high-incidence area of ​​the abnormal posture, performing a first clustering process on the first running posture data set, and dividing the target runway area into M initial sub-areas according to the first clustering process result; Collecting second video data of running in the initial sub-region, extracting coordinates of key points of the student's body from the second video data, constructing a second running posture data set, performing a second clustering process on the first running posture data set and the second running posture data, and dividing the initial sub-region into N fine sub-regions according to the second clustering process result; According to the refined sub-region, obtaining typical posture problems and posture data density distribution corresponding to the initial sub-region, and determining the main problem area according to the typical posture problems and the posture data density distribution; Determining a candidate feedback position in the fine sub-region according to the posture data density distribution and the problem severity index corresponding to the main problem region; Performing a comprehensive evaluation on the candidate feedback locations, and selecting the candidate feedback location with the highest comprehensive evaluation score and covering the main problem area as the feedback system deployment location; Taking the feedback system arrangement position as the center and combining the posture data density distribution, determining the effective feedback range of the feedback system arrangement position; Deploy a feedback system at the feedback system deployment position, set a corresponding trigger area for the feedback system deployment position according to the effective feedback range, and trigger the feedback system to provide running posture feedback to the student based on the typical posture problem when the student enters the trigger area; Among them, the step of obtaining typical posture problems and posture data density distribution corresponding to the initial sub-region according to the refined sub-region, and determining the main problem area according to the typical posture problems and the posture data density distribution, includes the steps of: Calculate the posture score corresponding to the data point in the fine sub-region by using the posture score function; Calculating the average value and standard deviation of the posture features in the fine sub-region, and obtaining an abnormal threshold value according to the average value and the standard deviation; The posture feature corresponding to the posture score being greater than the abnormal threshold is regarded as the typical posture problem; Using a kernel density estimation method, the density distribution of the posture data of the fine sub-region is calculated; Calculating the spatial distribution of the severity of the problem corresponding to the fine sub-region according to the posture data density distribution and the weight of the typical posture problem; Sorting the spatial distribution of the severity of the problem from large to small, and taking the first part of the sorting result as the main problem area; The step of determining a candidate feedback position in the fine sub-region according to the posture data density distribution and the problem severity index corresponding to the main problem region comprises the steps of: The reaction time function, the posture data density distribution and the problem severity index corresponding to the main problem area are weighted to obtain a feedback position scoring function; wherein the reaction time function indicates how much time and space the student has to adjust the posture after providing feedback at a certain position; Using the feedback position scoring function, calculating the feedback position score corresponding to the position in the fine sub-region; Based on a sliding window algorithm, obtaining a local maximum value in the feedback position score; The position corresponding to the local maximum value being greater than a preset threshold is taken as a primary candidate feedback position; the preset threshold is set according to the average value and standard deviation of the feedback position score; sorting the primary candidate feedback positions from large to small, and calculating the distance between any two of the primary candidate feedback positions in the sorting result, and when the distance is less than a preset interval, deleting the primary candidate feedback position with the smallest local maximum value among the two primary candidate feedback positions, to obtain a secondary candidate feedback position; According to the length of the fine sub-region and the preset maximum number of feedback positions, the first W of the secondary candidate feedback positions from largest to smallest are used as the candidate feedback positions.

2. The method for real-time feedback of students' running posture based on motion recognition according to claim 1 is characterized in that: The step of analyzing the first running posture data set to determine the high-incidence area of ​​abnormal posture includes the following steps: Dividing the target runway area into N intervals on average; Calculating a posture abnormality score corresponding to the interval; The sum of the average value and the standard deviation of the posture abnormality score is calculated, and the interval where the posture abnormality score is greater than the sum is defined as a high-incidence area of ​​the posture abnormality.

3. The method for real-time feedback of students' running posture based on motion recognition according to claim 1 is characterized in that: The step of performing a first clustering process on the first running posture data set based on the high-incidence area of ​​the abnormal posture, and dividing the target runway area into M initial sub-areas according to the first clustering process result includes the steps of: defining a first type of feature vector corresponding to a data point in the first running posture data set; Performing a first clustering process on the first type of feature vectors using a first clustering algorithm to obtain a first type of cluster set; The first cluster set is mapped to the target runway area to obtain the initial sub-area.

4. The method for real-time feedback of students' running posture based on motion recognition according to claim 1 is characterized in that: The steps of collecting second video data of running in the initial sub-region, extracting coordinates of key points of the student's body from the second video data, constructing a second running posture data set, performing a second clustering process on the first running posture data set and the second running posture data set, and dividing the initial sub-region into N fine sub-regions according to the second clustering process result include the steps of: Merging the first running posture data set and the second running posture data set to obtain a comprehensive data set; defining corresponding second-type feature vectors for data points in the comprehensive data set; Performing a second clustering process on the second type of feature vectors corresponding to the data points in the initial sub-region using a second clustering algorithm to obtain a second type of cluster set; The second cluster set is mapped to the initial sub-region to obtain the refined sub-region.

5. The method for real-time feedback of students' running posture based on motion recognition according to claim 1 is characterized in that: The step of comprehensively evaluating the candidate feedback positions and selecting the candidate feedback position having the highest comprehensive evaluation score and covering the main problem area as the feedback system deployment position comprises the following steps: Calculating the coverage effect score, problem targeting score, feedback timeliness score and environmental adaptability score of the candidate feedback location; Performing weighted processing on the coverage effect score, the problem targeting score, the feedback timeliness score, and the environmental adaptability score to obtain a comprehensive evaluation score of the candidate feedback location; Sorting the candidate feedback positions according to the comprehensive evaluation scores from large to small to obtain a sorted list; The feedback system deployment position is obtained according to the coverage function and the sorting list; wherein the coverage function represents the coverage degree of the selected position set to the main problem area.

6. The method for real-time feedback of students' running posture based on motion recognition according to claim 1 is characterized in that: The step of determining the effective feedback range of the feedback system layout position with the feedback system layout position as the center and in combination with the posture data density distribution includes the steps of: Smoothing the density distribution of the posture data; Setting a corresponding initial search range for the feedback system deployment location; Within the initial search range, determining an effective feedback range using the smoothed density distribution of the posture data; It is determined whether there is an overlapping area between adjacent effective feedback ranges. If there is an overlapping area, the adjacent effective feedback ranges are adjusted based on the overlapping area to obtain the effective feedback range of the feedback system deployment position.

7. The method for real-time feedback of student running posture based on motion recognition according to claim 1 is characterized in that: The step of deploying a feedback system at the feedback system deployment position, setting a corresponding trigger area for the feedback system deployment position according to the effective feedback range, and triggering the feedback system to provide running posture feedback to the student based on the typical posture problem when the student enters the trigger area comprises the following steps: Score the student's current posture based on the real-time posture evaluation function to obtain the current posture score; When the current posture score exceeds the trigger threshold corresponding to the trigger area, running posture feedback is provided to the student based on the feedback strategy corresponding to the typical posture problem.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.