A physical education classroom teaching evaluation method and system based on big data

Abstract

The application relates to the technical field of physical education, and particularly discloses a physical classroom teaching evaluation method and system based on big data, which collects continuous multi-frame multi-target skeleton point sequence data; maps the skeleton point sequence data into a space-time graph structure containing space edges and time edges; performs graph convolution processing on the space-time graph data, extracts motion trajectory features and posture features, and calculates kinematic consistency indexes based on skeleton length constancy constraints and motion trajectory smoothness constraints; detects value discontinuous points caused by identity exchange according to the indexes, corrects identity labels through global optimization backtracking, and outputs continuous identity label skeleton point sequences; finally, standard action templates are compared frame by frame, action completion degrees and motion load parameters are calculated, and individualized teaching evaluation reports are generated; the application solves the problem of space-time fracture of skeleton point sequences caused by identity exchange in a multi-person dense scene, and provides stable and reliable quantitative evaluation data for physical education.

Description

Technical Field

[0001] This invention relates to the field of physical education teaching technology, specifically to a method and system for evaluating physical education classroom teaching based on big data. Background Technology

[0002] With the development of information technology, intelligent teaching and assessment methods are gradually being applied to physical education classes. Currently, some physical education classes are beginning to use cameras to collect videos of students' movements, extract key points of the human skeleton through computer vision technology, and then analyze the students' movement standardization and exercise intensity.

[0003] In complex physical education teaching scenarios with many people densely packed and mutually obstructing each other, the visual algorithm frequently swaps identity identifiers, causing the output skeletal point sequence to exhibit data characteristics that violate the physical laws of human movement on the time axis. The bone length of the same target changes abruptly between adjacent frames, and the trajectory of the centroid moves instantaneously. This non-physiological spatiotemporal data breakage problem is difficult to fundamentally solve through single-frame image optimization or hardware upgrades. However, this invention constructs constraints on the constancy of bone length and the smoothness of movement trajectory, starting from the spatiotemporal continuity of the movement data itself, automatically detects and backtracks to correct identity identifiers, so that the output skeletal point sequence conforms to the physical laws of real human movement in the time dimension. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for evaluating physical education classroom teaching based on big data, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A big data-based method for evaluating physical education classroom teaching includes the following steps:

[0007] S1. Collect multi-target skeleton point sequence data of multiple consecutive frames in a physical education class;

[0008] S2. Map the multi-target skeleton point sequence data into a spatiotemporal graph structure, where the skeleton points in each frame are naturally connected according to human joints to form spatial edges, and the candidate targets between adjacent frames are formed temporal edges according to their positional proximity. The output is spatiotemporal graph data containing spatial topology and temporal correlation.

[0009] S3. Perform graph convolution processing on the spatiotemporal graph data to extract the motion trajectory features and pose features of each target. Calculate the kinematic consistency index of targets between adjacent frames based on the constant bone length constraint and motion trajectory smoothness constraint, and output a feature spatiotemporal graph with additional consistency index.

[0010] S4. Based on the kinematic consistency index, detect the numerical discontinuity points caused by identity swapping, locate the conflict start frame, and backtrack and correct the identity identifier through a global optimization algorithm to ensure that the bone point sequence of the same target satisfies constant bone length and smooth trajectory, and output the corrected continuous identity identifier bone point sequence.

[0011] S5. Based on the corrected continuous identity identifier skeleton point sequence, compare the preset standard action template frame by frame, calculate the action completion degree and exercise load parameters, and generate a personalized teaching evaluation report.

[0012] As a further aspect of the present invention: S2 specifically includes:

[0013] Extract the 3D coordinates of the skeletal key points of each candidate target in each frame, and generate spatial edges based on the human joint connection relationship;

[0014] The system calculates the positional similarity, pose similarity, and motion velocity continuity indices between the current frame candidate target and the candidate targets in adjacent frames. It then weights and fuses these three indices to generate a cross-frame association cost. The system solves for the global optimal matching relationship through dynamic programming and connects the successfully matched candidate targets with time edges, outputting spatiotemporal graph data with spatiotemporal consistency.

[0015] As a further aspect of the present invention: the generation of spatial edges based on the connection relationships of human joints specifically includes:

[0016] Feature point matching is performed on multi-view images acquired at the same time to obtain the initial three-dimensional coordinates of each joint.

[0017] Based on the connection relationship of human joints, a skeletal topological constraint is constructed, and the initial three-dimensional coordinates are projected onto a preset skeletal length ratio field for optimization, so that the spatial distance between adjacent joints conforms to the anatomical proportion of the human body.

[0018] The optimized 3D coordinates are weighted and smoothed on the time axis according to the motion continuity to eliminate outlier coordinates caused by mismatches in a single frame, and the 3D coordinates of the skeletal key points of each candidate target and the corresponding spatial edge connection relationship are output.

[0019] As a further aspect of the present invention: S3 specifically includes:

[0020] The spatiotemporal graph data is input into the graph convolutional layer. The coordinate differences of adjacent joints are aggregated along the spatial edge to extract spatial pose features, and the displacement vectors of the same joint between adjacent frames are aggregated along the temporal edge to extract motion trajectory features.

[0021] Based on the constraint of constant bone length, the fluctuation amplitude of the bone segment length of the same target between consecutive frames is calculated. Based on the constraint of smoothness of motion trajectory, the number of acceleration abrupt changes of the centroid trajectory on the time axis is calculated.

[0022] The fluctuation amplitude and the number of acceleration mutations are weighted and fused to generate a kinematic consistency index that characterizes identity stability.

[0023] The kinematic consistency index is attached to the features of each node in the spatiotemporal graph data, and the spatiotemporal graph with the added consistency index is output.

[0024] As a further aspect of the present invention: the feature spatiotemporal graph of the output additional consistency index specifically includes:

[0025] The kinematic consistency index is converted into confidence weights for each node, and a confidence weight vector is constructed.

[0026] The adjacency matrix of the spatiotemporal graph data is weighted and modulated based on the confidence weight vector to generate an adaptive adjacency relationship that reflects the stability of identity.

[0027] The original node features are smoothly diffused using graph Laplacian by adaptive adjacency relationship, so that the features of high confidence nodes are propagated to adjacent frames along the time edge, and the features of low confidence nodes are replaced by the features of neighboring high confidence nodes.

[0028] The node features after smooth diffusion are recombined with the original spatiotemporal graph structure to output a feature spatiotemporal graph with an additional consistency index.

[0029] As a further aspect of the present invention: S4 specifically includes:

[0030] Scan the feature spatiotemporal graph with additional consistency index, extract the bone length sequence and centroid trajectory sequence of each candidate target on the time axis, and mark the frames whose bone length fluctuation exceeds the preset tolerance threshold and are accompanied by a sudden change in centroid acceleration as candidate conflict frames.

[0031] Starting from each candidate conflict frame, backtrack to the previous identity-stable frame, and extract the time window in between to construct a local spatiotemporal subgraph;

[0032] In the local spatiotemporal subgraph, all possible identity assignment combinations are enumerated, and the cumulative amount of bone length fluctuation and trajectory acceleration mutation of each target under each combination is calculated. The combination with the smallest weighted sum of the two is selected as the optimal identity assignment.

[0033] Based on the optimal identity identifier allocation result, the identity identifiers of each node in the local spatiotemporal subgraph are rewritten so that the skeletal point sequence of the same target satisfies constant bone length and smooth trajectory after splicing, and the corrected continuous identity identifier skeletal point sequence is output.

[0034] As a further aspect of the present invention: the construction of the local spatiotemporal subgraph specifically includes:

[0035] Mark the position of each candidate conflict frame on the time axis as the backtracking starting point, and retrieve the bone length fluctuation amplitude and the centroid acceleration change rate frame by frame in reverse along the time axis;

[0036] Frames in which the bone length fluctuation amplitude is below a preset threshold and the centroid acceleration change rate does not change abruptly are identified as identity-stable frames, and identity-stable frames are used as the backtracking termination point.

[0037] Extract the spatiotemporal graph data of all frames between the backtracking start point and the backtracking end point, and retain the connection relationship of the skeletal key points in each frame and the associated edges of the candidate targets between frames;

[0038] The extracted spatiotemporal graph data is reorganized into an independent local spatiotemporal subgraph in chronological order. The local spatiotemporal subgraph contains the complete skeletal topology and cross-frame association information, and the output is a local spatiotemporal subgraph used for identity correction.

[0039] As a further aspect of the present invention: S5 specifically includes:

[0040] The corrected continuous identity skeletal point sequence is compared frame by frame with the preset standard motion template to calculate the deviation of each joint angle from the standard angle and the spatial distance of the joint trajectory from the standard trajectory.

[0041] Based on deviation values ​​and spatial distance, combined with exercise physiological parameter tables, exercise load is quantified and graded to generate a performance score and exercise load intensity level.

[0042] Input the action completion score and exercise load intensity level into the report generation unit. The report generation unit automatically matches the comment text and improvement strategy according to the correspondence between the score and the level.

[0043] The evaluation text, improvement suggestions, and corresponding skeletal point sequence visualization images are integrated to output a personalized teaching evaluation report that includes text descriptions and image markers.

[0044] As a further aspect of the present invention: the generation of the action completion score and the exercise load intensity level specifically includes:

[0045] The deviation values ​​of each joint are assigned weight coefficients according to the joint type. The weight coefficients are preset based on the contribution of the joints to the completion of the movement. The weighted sum is then used to obtain the initial score of the movement completion.

[0046] The spatial distance is projected onto the corresponding exercise load interval in the exercise physiology parameter table, and the instantaneous heart rate range and energy consumption coefficient matched with the spatial distance are extracted to generate the exercise load intensity level.

[0047] Based on the correlation between the initial score of movement completion and the level of exercise load intensity, the initial score of movement completion is corrected so that movements completed under high exercise load intensity can gain a performance gain.

[0048] Output the corrected performance score and the corresponding exercise load intensity level.

[0049] A big data-based physical education classroom teaching evaluation system includes:

[0050] The skeleton point data acquisition module is used to collect multi-target skeleton point sequence data in multiple consecutive frames during physical education classes;

[0051] The spatiotemporal graph construction module maps multi-target skeletal point sequence data into a spatiotemporal graph structure. In this structure, skeletal points within each frame are naturally connected according to human joints to form spatial edges, and candidate targets between adjacent frames are connected according to their positional proximity to form temporal edges. The output is spatiotemporal graph data containing spatial topology and temporal correlation.

[0052] The graph convolution processing module performs graph convolution processing on the spatiotemporal graph data, extracts the motion trajectory features and pose features of each target, and calculates the kinematic consistency index of targets between adjacent frames based on the constant skeleton length constraint and the smoothness constraint of motion trajectory, and outputs a feature spatiotemporal graph with additional consistency index.

[0053] The identity correction module detects the numerical discontinuities caused by identity swapping based on the kinematic consistency index, locates the conflict start frame, and corrects the identity identifiers by backtracking through a global optimization algorithm, so that the skeletal point sequence of the same target satisfies the constant bone length and smooth trajectory, and outputs the corrected continuous identity identifier skeletal point sequence.

[0054] The evaluation report generation module compares the corrected continuous identity identifier skeleton point sequence frame by frame with the preset standard action template, calculates the action completion degree and exercise load parameters, and generates a personalized teaching evaluation report.

[0055] The beneficial effects of this invention are:

[0056] (1) This invention introduces constraints on the constancy of bone length and the smoothness of motion trajectory to construct a kinematic consistency index and backtrack to correct identity identifiers, effectively solving the problem of identity swapping and key point loss caused by occlusion in multi-person dense motion scenes. Compared with traditional visual evaluation methods, this invention can output spatiotemporally continuous and identity-stable bone point sequences, improving the reliability and accuracy of motion capture data and providing a high-quality data foundation for subsequent teaching evaluation.

[0057] (2) Based on the modified continuous identity marker skeleton point sequence, this invention compares the standard action template frame by frame and combines it with exercise physiology parameters for quantitative grading, generating a personalized evaluation report containing comment text, improvement strategies and visualization images. This scheme can not only objectively evaluate students' action completion and exercise load intensity, but also provide teachers with accurate teaching feedback and analysis of individual student differences, realizing scientific classroom management that teaches according to aptitude. Attached Figure Description

[0058] The invention will now be further described with reference to the accompanying drawings.

[0059] Figure 1 This is a flowchart of the method of the present invention;

[0060] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1 As shown, this invention is a method for evaluating physical education classroom teaching based on big data, comprising the following steps:

[0063] S1. Collect multi-target skeleton point sequence data of multiple consecutive frames in a physical education class;

[0064] S2. Map the multi-target skeleton point sequence data into a spatiotemporal graph structure, where the skeleton points in each frame are naturally connected according to human joints to form spatial edges, and the candidate targets between adjacent frames are formed temporal edges according to their positional proximity. The output is spatiotemporal graph data containing spatial topology and temporal correlation.

[0065] S3. Perform graph convolution processing on the spatiotemporal graph data to extract the motion trajectory features and pose features of each target. Calculate the kinematic consistency index of targets between adjacent frames based on the constant bone length constraint and motion trajectory smoothness constraint, and output a feature spatiotemporal graph with additional consistency index.

[0066] S4. Based on the kinematic consistency index, detect the numerical discontinuity points caused by identity swapping, locate the conflict start frame, and backtrack and correct the identity identifier through a global optimization algorithm to ensure that the bone point sequence of the same target satisfies constant bone length and smooth trajectory, and output the corrected continuous identity identifier bone point sequence.

[0067] S5. Based on the corrected continuous identity identifier skeleton point sequence, compare the preset standard action template frame by frame, calculate the action completion degree and exercise load parameters, and generate a personalized teaching evaluation report.

[0068] In S1, multi-frame multi-target skeleton point sequence data from a physical education class is collected, specifically including:

[0069] First, multiple depth cameras are installed around the perimeter of the physical education classroom, covering the entire teaching area from different perspectives. Each depth camera simultaneously captures color and depth images of the classroom scene and transmits the acquired image data to the data processing unit in real time.

[0070] Secondly, the data processing unit jointly processes the received multi-view color images and depth images, and extracts the three-dimensional coordinates of the skeletal key points of all students in each frame image through a feature point matching algorithm. The skeletal key points include major joints such as the head, shoulders, elbows, wrists, hips, knees, and ankles.

[0071] Then, the 3D coordinates of the skeletal key points extracted from different perspectives at the same time are fused to eliminate duplicate detections in overlapping areas of perspectives, and a temporary identity is assigned to each detected student to form a single frame of skeletal point data for that moment.

[0072] Finally, the single-frame data of multiple human skeleton points are arranged sequentially according to time, with the time interval between adjacent frames being the frame acquisition cycle of the depth camera, thus forming a sequence of multiple target skeleton points in consecutive frames. In the sequence of multiple target skeleton points, each skeleton point is associated with three-dimensional coordinate information, the timestamp of the frame to which it belongs, and a temporary identity identifier of the target to which it belongs.

[0073] In S2, multi-target skeletal point sequence data is mapped to a spatiotemporal graph structure. Skeletal points within each frame are naturally connected according to human joints to form spatial edges, while candidate targets between adjacent frames are connected based on positional proximity to form temporal edges. The output is spatiotemporal graph data containing both spatial topology and temporal correlation, specifically including:

[0074] First, for each frame in the multi-target skeletal point sequence data, the 3D coordinates of the skeletal keypoints of each candidate target within that frame are extracted. Here, candidate targets refer to each individual student detected in the current frame. After extracting the 3D coordinates, based on the natural connections between joints in the human anatomical structure, skeletal keypoints belonging to the same candidate target are connected pairwise to form spatial edges. For example, shoulder joints are connected to elbow joints, elbow joints to wrist joints, and so on, thereby constructing a spatial topology reflecting the human posture within a single frame. Spatial edges do not contain length attributes; they only indicate a direct skeletal connection between two joints.

[0075] Secondly, based on the construction of spatial edges within each frame, the association between candidate targets in adjacent frames is established. For each candidate target in the current frame, the positional similarity, pose similarity, and motion velocity continuity index are calculated between it and each candidate target in adjacent frames. The positional similarity is measured by calculating the centroid spatial distance between the two candidate targets in the same world coordinate system; the smaller the distance, the higher the similarity. The pose similarity is measured by calculating the cumulative sum of the differences in the three-dimensional coordinates of the corresponding joints of the two candidate targets; the smaller the difference, the closer the poses. The motion velocity continuity index is obtained by dividing the difference between the centroid coordinates of the candidate target in the current frame and the centroid coordinates of the candidate targets in the adjacent frames by the frame interval time to obtain the motion velocity vector, and then comparing the change amplitude of this velocity vector with the velocity vector at the previous moment; the smaller the change amplitude, the better the continuity. The above three indices are weighted and summed according to a preset weight ratio to obtain the cross-frame association cost. The weight ratio is set according to the intensity of motion in actual operation; for example, the positional similarity weight is 40%, the pose similarity weight is 40%, and the motion velocity continuity index weight is 20%.

[0076] Then, with the goal of minimizing the global association cost, a dynamic programming algorithm is used to solve for the globally optimal matching relationship of candidate targets across multiple consecutive frames. The specific implementation of dynamic programming is as follows: a multi-stage decision graph is constructed, where each frame is a stage, and the candidate targets within each frame are the state nodes of that stage; between adjacent stages, the transition cost between nodes is constructed based on the aforementioned cross-frame association cost; by recursively calculating the path with the minimum cumulative cost stage by stage, the optimal matching sequence spanning multiple consecutive frames is obtained. Successfully matched candidate targets are connected across different frames using time edges, with each time edge pointing to the continuation of the identity identifier of the same student individual at different times.

[0077] Finally, the constructed spatial and temporal edges are integrated to output spatiotemporal graph data that simultaneously contains the spatial topology within a single frame and the temporal correlations between frames. In the spatiotemporal graph data, nodes are the skeletal keypoints in each frame, spatial edges are used to describe the connection relationships between different keypoints at the same time, and temporal edges are used to describe the correspondence relationships between the same keypoint at different times.

[0078] The process of extracting the 3D coordinates of the skeletal key points of each candidate target in each frame and generating spatial edges based on the human joint connection relationship is specifically implemented in the following way:

[0079] First, feature point matching is performed on multiple images captured simultaneously by depth cameras from multiple viewpoints. Specifically, in each viewpoint image, the pixel coordinates of each joint are located using a pre-defined joint detection algorithm. Then, using camera calibration parameters across multiple viewpoints, the initial 3D coordinates of each joint in the world coordinate system are calculated using triangulation principles. Since different viewpoints may detect the same joint, the average of multiple 3D coordinates for the same joint is taken as the initial 3D coordinates of that joint.

[0080] Secondly, a skeletal topological constraint is constructed based on the joint connections of the human body, and the initial 3D coordinates are optimized by projection. Specifically, a pre-defined skeletal length ratio field is used, storing the length ratios of various skeletal segments in students of different ages and genders, such as the ratio of upper arm to forearm length, and thigh to calf length. Using the initial 3D coordinates of all joints of the current candidate target as input, the spatial distance between adjacent joints is calculated to obtain the measured length of each skeletal segment. The measured lengths are compared with the corresponding standard ratios in the skeletal length ratio field to calculate the ratio deviation of each skeletal segment. Through an iterative nearest-point algorithm, while maintaining the joint connection topology, the 3D coordinate positions of each joint are gradually adjusted to bring the length ratios of each skeletal segment closer to the standard ratio, while minimizing the overall displacement. The iteration terminates when the length ratio deviations of all skeletal segments are less than a preset threshold, such as less than 5%.

[0081] Then, the optimized 3D coordinates are weighted and smoothed along the time axis according to motion continuity. Specifically, for each joint, the 3D coordinates of the two frames before, the current frame, and the two frames after are taken (a total of five frames). A Gaussian weighted average is used to calculate the final 3D coordinates of the joint in the current frame, with frames closer to the current frame assigned higher weights. For example, the current frame has a weight of 40%, the adjacent frames before and after each have 25%, and the second and third frames before and after each have 5%. This weighted smoothing process eliminates outlier coordinate values ​​caused by mismatches in single-frame images, resulting in a smooth motion trajectory for the skeletal point sequence over time.

[0082] Finally, the three-dimensional coordinates of the skeletal key points of each candidate target are output, along with the spatial edge connection relationship constructed from these coordinate points based on the human joint connection relationship. The spatial edge connection relationship records which joint points have direct skeletal connections, which are used for the construction of subsequent spatiotemporal graph data.

[0083] In S3, graph convolution processing is performed on the spatiotemporal graph data to extract the motion trajectory and pose features of each target. Based on the constant bone length constraint and motion trajectory smoothness constraint, the kinematic consistency index of targets between adjacent frames is calculated, outputting a feature spatiotemporal graph with an additional consistency index, specifically including:

[0084] First, the spatiotemporal graph data output from step S2 is used as input for graph convolution processing. Graph convolution processing is divided into two directions: spatial and temporal. First, the coordinate differences between adjacent joints are aggregated along spatial edges to extract spatial pose features. For each node, its own features are weighted and summed with the features of all adjacent nodes connected by spatial edges. The weights are determined by preset values ​​for the edges, for example, each spatial edge has a weight of 1, thus obtaining the spatial aggregated features of that node. Simultaneously, the displacement vectors of the same joint between adjacent frames are aggregated along temporal edges to extract motion trajectory features. For each node, its own features are differentially calculated with the features of nodes connected to the same joint in the previous and next frames via temporal edges to obtain a displacement vector. The displacement vectors of multiple consecutive frames are accumulated to form motion trajectory features. After graph convolution processing, each node has two types of features: spatial pose feature vector and motion trajectory feature vector.

[0085] Secondly, based on the constraint of constant bone length, the fluctuation range of bone segment lengths for the same target across consecutive frames is calculated. A bone segment refers to a limb part formed by two adjacent joints connected by spatial edges, such as the upper arm bone segment from the shoulder joint to the elbow joint. For each target, the bone point coordinates of its N consecutive frames (N is an integer greater than or equal to 3, for example, N equals 5) are selected. For each bone segment, the spatial length of the bone segment in each frame is calculated to obtain the length sequence. Let the length values ​​of the bone segment in the N consecutive frames be respectively... Its arithmetic mean is The length fluctuation range P of the bone segment is obtained by calculating the standard deviation of the sequence. The specific calculation formula is as follows: ;in, For frame index, This is the average value for N frames. For the same target, the sum of the fluctuation amplitudes of all its bone segments is used to obtain the overall bone length fluctuation amplitude for that target. .

[0086] Then, based on the motion trajectory smoothness constraint, the number of acceleration abrupt changes on the time axis of the centroid trajectory is calculated. The centroid is defined as the average of the three-dimensional coordinates of all skeletal keypoints of the target. First, the centroid coordinates of each frame are calculated to obtain the centroid trajectory sequence. Then, the instantaneous velocity vector is obtained by dividing the difference between the centroid coordinates of adjacent frames by the frame interval time, and the instantaneous acceleration vector is obtained by dividing the difference between the velocity vectors of adjacent frames by the frame interval time. The number of times the magnitude of the acceleration vector exceeds a preset threshold (e.g., 0.5 m / s²) during the entire motion process is counted and recorded as the acceleration abrupt change number Q. This index reflects the degree of non-smoothness of the motion trajectory; the larger the Q value, the more severe the trajectory jitter.

[0087] Next, the overall bone length fluctuation range obtained above will be... The kinematic consistency index C, representing identity stability, is generated by weighted fusion with the number of acceleration mutations Q. The specific method of weighted fusion is as follows: ;in, and The preset weighting coefficients, and and The sum is 1, for example Take 0.6, Take 0.4. This indicator. The smaller the value, the more stable the target's identity is across consecutive frames; The larger the value, the more likely there may be identity swapping or tracking errors.

[0088] Subsequently, the calculated kinematic consistency index This is appended to the features of each node in the spatiotemporal graph data. This is achieved through the following steps:

[0089] The first step is to set the kinematic consistency index for each node. This is converted into a confidence weight for that node, where all nodes have the same weight for the same target. Value, so for the first Kinematic consistency index value of each node The conversion method uses a monotonically decreasing function, making... Nodes with smaller values ​​receive higher confidence levels. Specifically, for the ... Each node, its confidence weight ,in This represents the kinematic consistency index value for that node. A confidence weight vector with a dimension equal to the total number of nodes is then constructed based on this. , Each element in the table corresponds to the confidence weight of a node.

[0090] The second step involves weighting and modulating the adjacency matrix of the spatiotemporal graph data based on the confidence weight vector to generate adaptive adjacency relationships that reflect identity stability. (Original adjacency matrix) It is a symmetric matrix whose elements Represents a node With nodes The confidence level is set to 1 if an edge (spatial or temporal) exists between the edges, and 0 otherwise. To reflect the importance of the confidence level to the edge, a diagonal matrix is ​​constructed. Its diagonal elements are the confidence weights of the corresponding nodes, i.e. Then the new adjacency matrix Through formula The calculation yields results where the multiplication is matrix multiplication. Each element in In other words, the weight of an edge is modulated by the product of the confidence weights of the two endpoints. Thus, the edge connecting two high-confidence nodes has a higher weight, while the edge involving low-confidence nodes has a lower weight.

[0091] The third step involves using adaptive adjacency relationships to perform graph Laplacian smooth diffusion on the original node features, allowing the features of high-confidence nodes to propagate along the temporal edges to adjacent frames, while the features of low-confidence nodes are replaced by neighboring high-confidence features. First, based on the adaptive adjacency matrix... Calculate the degree matrix , For a diagonal matrix, its diagonal elements equal The Middle Sum of all elements in the row. Then construct the Laplace matrix. Next, the node feature matrix is ​​updated iteratively. The iterative formula is ,in For smoothing step size, the value is a decimal between 0 and 1, for example... =0.1. This iterative process moves the features of each node toward the features of its neighboring nodes, and the magnitude of the movement is controlled by the Laplacian matrix. After a preset number of iterations (e.g., 10), the features of low-confidence nodes are smoothed by the features of surrounding high-confidence nodes, thereby correcting the abnormal feature values ​​caused by identity swapping.

[0092] The fourth step involves recombining the smoothed and diffused node features with the original spatiotemporal graph structure (i.e., the original node connection relationships). The original spatial and temporal edges remain unchanged, and only the node feature vectors are updated, thus outputting a feature spatiotemporal graph with an added consistency index. This feature spatiotemporal graph contains node features corrected for kinematic consistency, providing a more reliable data foundation for subsequent identity correction steps.

[0093] In S4, based on the kinematic consistency index, numerical discontinuities caused by identity swapping are detected, the conflict initiation frame is located, and a global optimization algorithm is used to backtrack and correct the identity identifiers, ensuring that the skeletal point sequence of the same target satisfies constant bone length and smooth trajectory. The corrected continuous identity identifier skeletal point sequence is output, specifically including:

[0094] First, a comprehensive scan of the feature spatiotemporal map with the additional consistency index output in step S3 is performed. For each candidate target in the feature spatiotemporal map, the bone length sequence and centroid trajectory sequence of the target on the continuous time axis are extracted. The bone length sequence refers to multiple time series composed of the length values ​​of each bone segment of the target in each frame, with each bone segment corresponding to one length sequence; the centroid trajectory sequence refers to the point sequence composed of the average three-dimensional coordinates of all bone keypoints of the target in each frame. After extracting the above sequences, for each frame, the change amplitude of the length of each bone segment in the frame relative to the length of the corresponding bone segment in the previous frame is calculated, and the change amplitudes of all bone segments are summed to obtain the bone length fluctuation amplitude of the frame. At the same time, the velocity vector of the centroid position in the frame relative to the centroid position in the previous frame is calculated, and then the acceleration vector of the velocity vector in the frame relative to the velocity vector in the previous frame is calculated. If the bone length fluctuation of a frame exceeds a preset fluctuation threshold (e.g., fluctuation greater than 0.05 meters), and at the same time, the magnitude of the centroid acceleration vector of that frame exceeds a preset acceleration mutation threshold (e.g., acceleration greater than 0.8 meters per second squared), then that frame is marked as a candidate conflict frame. The preset fluctuation threshold and acceleration mutation threshold are pre-set based on the normal physiological range of human movement in the teaching scenario and stored in the configuration parameters.

[0095] Secondly, starting from each marked candidate conflict frame, a local spatiotemporal subgraph for identity correction is constructed by backtracking. The specific backtracking construction process is as follows: the position of the candidate conflict frame on the time axis is determined as the backtracking starting point; starting from this starting point, the frame is searched frame by frame in the opposite direction of the time axis. For each frame, the bone length fluctuation amplitude and the centroid acceleration change rate are calculated; when the bone length fluctuation amplitude of three consecutive frames is lower than the preset stability threshold (e.g., fluctuation amplitude less than 0.02 meters), and the centroid acceleration change rate of each of these three frames does not exceed the preset abrupt change threshold (i.e., acceleration change rate less than 0.5 meters per second squared), the last frame of these three consecutive frames (i.e., the frame farthest from the backtracking starting point) is determined as the corrected frame. The identity-stable frame is used as the backtracking endpoint. Then, starting from the next frame after the backtracking endpoint, the spatiotemporal graph data of all frames within this time window is extracted until the backtracking start point. During the extraction process, the spatial edge connections between key points of each skeleton in each frame are preserved, as well as the temporal edge association information between candidate targets in adjacent frames. Finally, the spatiotemporal graph data of these extracted frames are reorganized in chronological order to form an independent local spatiotemporal subgraph. This local spatiotemporal subgraph contains the complete skeleton topology and cross-frame association information, and is output as a local spatiotemporal subgraph for subsequent identity correction.

[0096] Then, for each of the constructed local spatiotemporal subgraphs, a global optimization method is used to correct the identity identifiers. Within the time window covered by the local spatiotemporal subgraph, the identity identifiers of each candidate target may be chaotic due to possible identity swapping. Specifically, all possible identity identifier allocation combinations within the time window are enumerated. Assuming the time window contains M candidate targets and T frames of images, the identity identifier allocation combination refers to reassigning a globally unique identity number to each candidate target detected in each of the T frames, ensuring that the same real student individual has the same identity number in the T frames. For each allocation combination, the cumulative bone length fluctuation and cumulative trajectory acceleration mutation of each target under that combination are calculated. The cumulative bone length fluctuation refers to the sum of the fluctuation amplitudes of all bone segment lengths of each target in the T frames under that combination, and then the sums of these values ​​for all targets are added together to obtain the total cumulative bone length fluctuation. The cumulative trajectory acceleration mutation amount refers to the sum of the number of acceleration mutations in the centroid trajectory of each target within T frames under this combination, and then summing these sums for all targets to obtain the total cumulative trajectory acceleration mutation amount. The total cumulative bone length fluctuation amount is multiplied by a first weighting coefficient, and the total cumulative trajectory acceleration mutation amount is multiplied by a second weighting coefficient; then the two are added together to obtain the total cost of this allocation combination. The first and second weighting coefficients are preset, for example, the first weighting coefficient is 0.6 and the second weighting coefficient is 0.4. Among all possible allocation combinations, the combination with the minimum total cost is selected as the optimal identity assignment.

[0097] Finally, based on the optimal identity assignment results determined above, the identity of each node (i.e., each skeletal keypoint in each frame) in the local spatiotemporal subgraph is rewritten, so that the skeletal points of the same real student individual in different frames are assigned the same identity number. After rewriting, the skeletal point sequence of the same target within the local spatiotemporal subgraph, after being spliced, can meet the requirements that the bone length remains constant on the time axis and the centroid trajectory is smooth and continuous. The corrected local spatiotemporal subgraph is reintegrated into the complete feature spatiotemporal graph, outputting a corrected continuous identity-identified skeletal point sequence. Each skeletal point in this sequence is accompanied by a stable and continuous identity, providing an accurate data foundation for subsequent motion evaluation.

[0098] In S5, based on the corrected continuous identity marker skeleton point sequence, the system compares the preset standard action template frame by frame to calculate the action completion degree and motion load parameters, generating a personalized teaching evaluation report, which specifically includes:

[0099] First, the corrected continuous identity marker skeleton sequence output from step S4 is used as input data and compared frame-by-frame with a preset standard action template. The standard action template is stored in the teaching resource library. This template defines the standard angle values ​​of each joint at each moment and the standard motion trajectory of each joint for each sports action (e.g., standing long jump, basketball shooting). For each student's skeleton sequence, the difference between the actual angle and the standard angle of each joint is calculated frame-by-frame to obtain the joint angle deviation value. Simultaneously, the Euclidean distance between the actual three-dimensional coordinates of each joint in each frame and the corresponding point on the standard trajectory is calculated to obtain the joint trajectory spatial distance. The deviation values ​​and spatial distances of all frames are accumulated and averaged to obtain the student's average joint angle deviation and average joint trajectory spatial distance for that action.

[0100] Secondly, based on the calculated average joint angle deviation and average joint trajectory spatial distance, combined with the exercise physiology parameter table, a movement completion score and exercise load intensity level are generated. The specific implementation process is as follows: First, assign different weight coefficients to the average angle deviation values ​​of each joint according to the joint type. The weight coefficients are set according to the contribution of the joint in completing a specific movement. For example, in a basketball shooting motion, the weight coefficients for the wrist and elbow joints are set to 0.4 and 0.3, while the weight coefficients for the knee and hip joints are set to 0.2 and 0.1, and the sum of the weight coefficients for all joints is 1. Multiply the average angle deviation value of each joint by its corresponding weight coefficient, and then add the products of all joints to obtain the initial movement completion score. This initial score is a value between 0 and 100, with a higher score for a smaller deviation. Second, use the average joint trajectory spatial distance as input to query the exercise physiology parameter table. This table is pre-constructed based on exercise physiology experimental data and records the instantaneous heart rate range (e.g., heart rate per minute) and energy consumption coefficient (e.g., kcal per kilogram of body weight per minute) corresponding to different spatial distance intervals. Based on the calculated average joint trajectory spatial distance within the calculated range, the corresponding instantaneous heart rate range and energy consumption coefficient are extracted. The exercise load intensity is then categorized into low, medium, and high intensity levels according to the heart rate range. Thirdly, the initial exercise completion score is corrected based on the correlation between the initial score and the exercise load intensity level. The correction rule is as follows: if the exercise load intensity level is high, a correction value is added to the initial exercise completion score, for example, 5 points; if it is medium, 2 points are added; if it is low, no points are added or points are subtracted. The purpose of the correction is to reflect the extra effort students put into completing the movement under a higher exercise load. Finally, the corrected exercise completion score and the corresponding exercise load intensity level are output.

[0101] Then, the revised action completion score and exercise load intensity level are input into the report generation unit. The report generation unit pre-stores multiple comment text templates and improvement strategy text templates, each corresponding to a specific score range and load level combination. For example, when the score is above 90 and the load level is high intensity, the matched comment is "The movement is standard and the exercise engagement is high," and the improvement strategy is "Maintain the current training intensity and pay attention to post-exercise cool-down." When the score is below 60 and the load level is low intensity, the matched comment is "Movement standardization needs improvement, exercise intensity is insufficient," and the improvement strategy is "Strengthen basic movement practice and appropriately increase the exercise load." Based on the actual obtained score and load level, the report generation unit automatically finds and selects the corresponding comment text and improvement strategy text.

[0102] Finally, the selected comments and improvement strategies were integrated with the corresponding skeletal point sequence visualization. The visualization overlays the student's skeletal key points, joint angle markers, and trajectory lines onto the original video frames, highlighting areas with significant deviations in different colors. The integrated data was packaged into a complete personalized teaching evaluation report file, containing both textual descriptions (comments and improvement strategies) and image annotations (skeletal point sequence visualization), and output to a display terminal or storage medium for teachers and students to review.

[0103] Please see Figure 2 As shown, a big data-based physical education classroom teaching evaluation system includes:

[0104] The skeleton point data acquisition module is used to collect multi-target skeleton point sequence data in multiple consecutive frames during physical education classes;

[0105] The spatiotemporal graph construction module maps multi-target skeletal point sequence data into a spatiotemporal graph structure. In this structure, skeletal points within each frame are naturally connected according to human joints to form spatial edges, and candidate targets between adjacent frames are connected according to their positional proximity to form temporal edges. The output is spatiotemporal graph data containing spatial topology and temporal correlation.

[0106] The graph convolution processing module performs graph convolution processing on the spatiotemporal graph data, extracts the motion trajectory features and pose features of each target, and calculates the kinematic consistency index of targets between adjacent frames based on the constant skeleton length constraint and the smoothness constraint of motion trajectory, and outputs a feature spatiotemporal graph with additional consistency index.

[0107] The identity correction module detects the numerical discontinuities caused by identity swapping based on the kinematic consistency index, locates the conflict start frame, and corrects the identity identifiers by backtracking through a global optimization algorithm, so that the skeletal point sequence of the same target satisfies the constant bone length and smooth trajectory, and outputs the corrected continuous identity identifier skeletal point sequence.

[0108] The evaluation report generation module compares the corrected continuous identity identifier skeleton point sequence frame by frame with the preset standard action template, calculates the action completion degree and exercise load parameters, and generates a personalized teaching evaluation report.

[0109] The working principle of this invention is as follows: Multi-target skeletal point sequence data from multiple consecutive frames in a classroom is collected and mapped into a spatiotemporal graph structure containing spatial and temporal edges. Then, graph convolution processing is performed on the spatiotemporal graph data to extract the motion trajectory and posture features of each target. Based on the constraints of constant bone length and smooth motion trajectory, a kinematic consistency index for targets between adjacent frames is calculated, and a feature spatiotemporal graph with this index is output. Furthermore, based on the kinematic consistency index, points of numerical discontinuity caused by identity swapping are detected, the conflict initiation frame is located, and a global optimization algorithm is used to backtrack and correct the identity identifier, ensuring that the skeletal point sequence of the same target satisfies constant bone length and smooth trajectory, outputting a corrected continuous identity identifier skeletal point sequence. Finally, based on the corrected sequence, frame-by-frame comparison with a preset standard action template is performed to calculate the action completion degree and motion load parameters, generating a personalized teaching evaluation report containing comment text, improvement strategies, and a visualized image of the skeletal point sequence.

[0110] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for evaluating physical education classroom teaching based on big data, characterized in that, Includes the following steps: S1. Collect multi-target skeleton point sequence data of multiple consecutive frames in a physical education class; S2. Map the multi-target skeleton point sequence data into a spatiotemporal graph structure, where the skeleton points in each frame are naturally connected according to human joints to form spatial edges, and the candidate targets between adjacent frames are formed temporal edges according to their positional proximity. The output is spatiotemporal graph data containing spatial topology and temporal correlation. S3. Perform graph convolution processing on the spatiotemporal graph data to extract the motion trajectory features and pose features of each target. Calculate the kinematic consistency index of targets between adjacent frames based on the constant bone length constraint and motion trajectory smoothness constraint, and output a feature spatiotemporal graph with additional consistency index. S4. Based on the kinematic consistency index, detect the numerical discontinuity points caused by identity swapping, locate the conflict start frame, and backtrack and correct the identity identifier through a global optimization algorithm to ensure that the bone point sequence of the same target satisfies constant bone length and smooth trajectory, and output the corrected continuous identity identifier bone point sequence. S5. Based on the corrected continuous identity skeletal point sequence, compare the preset standard action template frame by frame, calculate the action completion degree and exercise load parameters, and generate a personalized teaching evaluation report. S3 specifically includes: The spatiotemporal graph data is input into the graph convolutional layer. The coordinate differences of adjacent joints are aggregated along the spatial edge to extract spatial pose features, and the displacement vectors of the same joint between adjacent frames are aggregated along the temporal edge to extract motion trajectory features. Based on the constraint of constant bone length, the fluctuation amplitude of the bone segment length of the same target between consecutive frames is calculated. Based on the constraint of smoothness of motion trajectory, the number of acceleration abrupt changes of the centroid trajectory on the time axis is calculated. The fluctuation amplitude and the number of acceleration mutations are weighted and fused to generate a kinematic consistency index that characterizes identity stability. The kinematic consistency index is attached to the features of each node in the spatiotemporal graph data, and the spatiotemporal graph with the added consistency index is output. S4 specifically includes: Scan the feature spatiotemporal graph with additional consistency index, extract the bone length sequence and centroid trajectory sequence of each candidate target on the time axis, and mark the frames whose bone length fluctuation exceeds the preset tolerance threshold and are accompanied by a sudden change in centroid acceleration as candidate conflict frames. Starting from each candidate conflict frame, backtrack to the previous identity-stable frame, and extract the time window in between to construct a local spatiotemporal subgraph; In the local spatiotemporal subgraph, all possible identity assignment combinations are enumerated, and the cumulative amount of bone length fluctuation and trajectory acceleration mutation of each target under each combination is calculated. The combination with the smallest weighted sum of the two is selected as the optimal identity assignment. Based on the optimal identity identifier allocation result, the identity identifiers of each node in the local spatiotemporal subgraph are rewritten so that the skeletal point sequence of the same target satisfies constant bone length and smooth trajectory after splicing, and the corrected continuous identity identifier skeletal point sequence is output.

2. The method for evaluating physical education classroom teaching based on big data according to claim 1, characterized in that, S2 specifically includes: Extract the 3D coordinates of the skeletal key points of each candidate target in each frame, and generate spatial edges based on the human joint connection relationship; The system calculates the positional similarity, pose similarity, and motion velocity continuity indices between the current frame candidate target and the candidate targets in adjacent frames. It then weights and fuses these three indices to generate a cross-frame association cost. The system solves for the global optimal matching relationship through dynamic programming and connects the successfully matched candidate targets with time edges, outputting spatiotemporal graph data with spatiotemporal consistency.

3. The method for evaluating physical education classroom teaching based on big data according to claim 2, characterized in that, The generation of spatial edges based on the connection relationships of human joints specifically includes: Feature point matching is performed on multi-view images acquired at the same time to obtain the initial three-dimensional coordinates of each joint. Based on the connection relationship of human joints, a skeletal topological constraint is constructed, and the initial three-dimensional coordinates are projected onto a preset skeletal length ratio field for optimization, so that the spatial distance between adjacent joints conforms to the anatomical proportion of the human body. The optimized 3D coordinates are weighted and smoothed on the time axis according to the motion continuity to eliminate outlier coordinates caused by mismatches in a single frame, and the 3D coordinates of the skeletal key points of each candidate target and the corresponding spatial edge connection relationship are output.

4. The method for evaluating physical education classroom teaching based on big data according to claim 3, characterized in that, The characteristic spatiotemporal graph of the output additional consistency index specifically includes: The kinematic consistency index is converted into confidence weights for each node, and a confidence weight vector is constructed. The adjacency matrix of the spatiotemporal graph data is weighted and modulated based on the confidence weight vector to generate an adaptive adjacency relationship that reflects the stability of identity. The original node features are smoothly diffused using graph Laplacian by adaptive adjacency relationship, so that the features of high confidence nodes are propagated to adjacent frames along the time edge, and the features of low confidence nodes are replaced by the features of neighboring high confidence nodes. The node features after smooth diffusion are recombined with the original spatiotemporal graph structure to output a feature spatiotemporal graph with an additional consistency index.

5. The method for evaluating physical education classroom teaching based on big data according to claim 4, characterized in that, The construction of the local spatiotemporal subgraph specifically includes: Mark the position of each candidate conflict frame on the time axis as the backtracking starting point, and retrieve the bone length fluctuation amplitude and the centroid acceleration change rate frame by frame in reverse along the time axis; Frames in which the bone length fluctuation amplitude is below a preset threshold and the centroid acceleration change rate does not change abruptly are identified as identity-stable frames, and identity-stable frames are used as the backtracking termination point. Extract the spatiotemporal graph data of all frames between the backtracking start point and the backtracking end point, and retain the connection relationship of the skeletal key points in each frame and the associated edges of the candidate targets between frames; The extracted spatiotemporal graph data is reorganized into an independent local spatiotemporal subgraph in chronological order. The local spatiotemporal subgraph contains the complete skeletal topology and cross-frame association information, and the output is a local spatiotemporal subgraph used for identity correction.

6. The method for evaluating physical education classroom teaching based on big data according to claim 1, characterized in that, S5 specifically includes: The corrected continuous identity skeletal point sequence is compared frame by frame with the preset standard motion template to calculate the deviation of each joint angle from the standard angle and the spatial distance of the joint trajectory from the standard trajectory. Based on deviation values ​​and spatial distance, combined with exercise physiological parameter tables, exercise load is quantified and graded to generate a performance score and exercise load intensity level. Input the action completion score and exercise load intensity level into the report generation unit. The report generation unit automatically matches the comment text and improvement strategy according to the correspondence between the score and the level. The evaluation text, improvement suggestions, and corresponding skeletal point sequence visualization images are integrated to output a personalized teaching evaluation report that includes text descriptions and image markers.

7. The method for evaluating physical education classroom teaching based on big data according to claim 6, characterized in that, The generation of the action completion score and exercise load intensity level specifically includes: The deviation values ​​of each joint are assigned weight coefficients according to the joint type. The weight coefficients are preset based on the contribution of the joints to the completion of the movement. The weighted sum is then used to obtain the initial score of the movement completion. The spatial distance is projected onto the corresponding exercise load interval in the exercise physiology parameter table, and the instantaneous heart rate range and energy consumption coefficient matched with the spatial distance are extracted to generate the exercise load intensity level. Based on the correlation between the initial score of movement completion and the level of exercise load intensity, the initial score of movement completion is corrected so that movements completed under high exercise load intensity can gain a performance gain. Output the corrected performance score and the corresponding exercise load intensity level.

8. A physical education classroom teaching evaluation system based on big data, characterized in that, A method for evaluating physical education classroom teaching based on big data, as described in any one of claims 1-7, comprises: The skeleton point data acquisition module is used to collect multi-target skeleton point sequence data in multiple consecutive frames during physical education classes; The spatiotemporal graph construction module maps multi-target skeletal point sequence data into a spatiotemporal graph structure. In this structure, skeletal points within each frame are naturally connected according to human joints to form spatial edges, and candidate targets between adjacent frames are connected according to their positional proximity to form temporal edges. The output is spatiotemporal graph data containing spatial topology and temporal correlation. The graph convolution processing module performs graph convolution processing on the spatiotemporal graph data, extracts the motion trajectory features and pose features of each target, and calculates the kinematic consistency index of targets between adjacent frames based on the constant skeleton length constraint and the smoothness constraint of motion trajectory, and outputs a feature spatiotemporal graph with additional consistency index. The identity correction module detects the numerical discontinuities caused by identity swapping based on the kinematic consistency index, locates the conflict start frame, and corrects the identity identifiers by backtracking through a global optimization algorithm, so that the skeletal point sequence of the same target satisfies the constant bone length and smooth trajectory, and outputs the corrected continuous identity identifier skeletal point sequence. The evaluation report generation module compares the corrected continuous identity identifier skeleton point sequence frame by frame with the preset standard action template, calculates the action completion degree and exercise load parameters, and generates a personalized teaching evaluation report.

Images