Body measurement action detection method and system based on behavior feature recognition

Abstract

The application discloses a physical measurement action detection method and system based on behavior feature recognition, and relates to the technical field of action detection; the method comprises the following steps: obtaining standard data of each physical measurement item, judging core skeleton points of each physical measurement item, and constructing a recognition model based on the core skeleton points; obtaining physical measurement data of each physical measurement item of a user, pre-processing the physical measurement data, collecting historical physical measurement analysis data of the user, and setting an actual recognition interval of the recognition model according to the historical physical measurement analysis data of the user; the technical key points are that a standardized physical measurement system is constructed, the recognition model is constructed by combining standard data and core skeleton points, full-automatic recognition detection from data collection, preprocessing to action evaluation is realized, a unified skeleton point detection standard, scoring rule and data processing mode are adopted, the fairness of physical measurement results is ensured, a foundation for physical measurement intelligentization is laid, standardized physical measurement is realized, and the physical measurement efficiency and reliability are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of action detection, in particular to a body measurement action detection method and system based on behavior feature recognition. BACKGROUND

[0002] Body measurement, i.e. physical fitness test, is a comprehensive evaluation method for systematically evaluating the physical function, movement ability and health status of individuals through a series of standardized physiological and movement indicators. The core purpose is to quantitatively reflect the level of human body in core dimensions such as cardiorespiratory endurance, muscle strength, flexibility, coordination and explosive power. It is widely used in student physical fitness monitoring, athlete training evaluation and public health management, and common projects include standing long jump, pull-up, sit and reach, long run, sit-up, etc.

[0003] Campus body measurement, as the core means of national monitoring of adolescent physical health, covers the whole school group from primary school to university. The evaluation results are not only the physical examination report of individual physical fitness of students, but also an important basis for education departments to formulate sports policies and optimize teaching programs. This feature puts forward rigid requirements for campus body measurement - a unified and objective evaluation standard must be established to ensure the comparability of results between different schools, different batches and different evaluators, and truly reflect the real level of student physical fitness.

[0004] However, in the traditional campus body measurement, the judgment of action specification highly depends on manual observation: whether the chin is below the bar in pull-up, whether the elbow touches the knee in sit-up, whether the feet are simultaneously kicked in standing long jump, etc. Key standards often deviate due to differences in experience, visual errors and even subjective relaxation of evaluators. Different classes in the same school and the same grade in different schools may implement different implicit standards, leading to distorted data, which cannot accurately measure student physical fitness and weakens the application value of body measurement data, and even may cause students to question the fairness of the evaluation;

[0005] Therefore, the existing body measurement method has certain disadvantages. In order to improve the fairness of body measurement, a body measurement action detection method and system based on behavior feature recognition are provided. SUMMARY

[0006] To achieve the above purpose, the present application is realized by the following technical scheme:

[0007] The body measurement action detection method based on behavior feature recognition comprises the following steps:

[0008] S1, obtaining standard data of each body measurement project, judging core skeletal points of each body measurement project, and constructing a recognition model based on the core skeletal points;

[0009] S2, obtain body measurement data of each individual measurement item of a user, and pre-process the body measurement data, collect historical body measurement analysis data of the user, and set an actual recognition interval of a recognition model according to the historical body measurement analysis data of the user;

[0010] S3, deliver the pre-processed body measurement data to the corresponding recognition model with the set actual recognition interval, perform single-view skeleton point detection, and execute a repair strategy when there is occlusion, to obtain a core skeleton point sequence;

[0011] S4, extract action features from the core skeleton point sequence, implement action division, and perform eligibility scoring to generate an evaluation report;

[0012] S5, comprehensively evaluate the student's ability based on the evaluation report, and generate an adjustment strategy according to the eligibility score and the comprehensive evaluation.

[0013] Preferably, the standard data includes standard videos of each individual measurement item and body measurement eligibility indicators, and the steps of judging the core skeleton points of each individual measurement item are as follows:

[0014] An BlazePose model is used to extract L complete skeleton points related to body measurement, to generate a skeleton point coordinate data set for each frame;

[0015] The video is segmented according to the action stage, the key frame of each stage is marked, and a complete action cycle is divided;

[0016] The coordinate change ratio and angle change rate of the L skeleton points in the complete action cycle are calculated, the change amount is calculated by weighting, and the change amount is compared with the set standard change amount. If the change amount is lower than the standard change amount, no processing is performed, otherwise, it is marked as a high dynamic point;

[0017] The importance of the skeleton points of each measurement item is scored, the influence of the skeleton points on the action standard is evaluated, and the skeleton points with a high importance score higher than a preset marking score are marked as important scores;

[0018] The union of the high dynamic points and the important score skeleton points is determined to determine the core skeleton points of each measurement item.

[0019] Preferably, the body measurement data includes body measurement videos of each view during body measurement, body measurement equipment worn on the body measurement personnel during body measurement, and measurement data collected by the body measurement instrument, and the pre-processing of the body measurement data includes:

[0020] Frame rate processing: the body measurement videos of each view are extracted by FFmpeg technology according to fixed timestamps, different frame rate videos are processed by interpolation or frame extraction, and the frame rate is unified;

[0021] De-noising enhancement processing: adopt adaptive Gaussian filter to remove the salt and pepper noise of the video after uniform frame rate, then adopt non-local mean de-noising algorithm to enhance the edge definition of motion blur frame, balance the brightness through gamma correction, and finally adopt limited contrast adaptive histogram equalization algorithm to enhance the local contrast;

[0022] Crop processing: adopt YOLOv5s model to detect the human body region in the video frame of the body measurement video after enhancement processing, output the bounding box coordinates, crop the video frame according to the bounding box, retain the human body region and surrounding 20% redundant pixels, and output the video frame with uniform size;

[0023] Normalization processing: normalize the measurement data collected by the body measurement device and the body measurement instrument.

[0024] Preferably, the user historical body measurement data analysis result includes motion cycle characteristics and stability characteristics, the motion cycle characteristics include single round motion average cycle time length, and the motion rhythm is determined by the inter-frame skeleton point change rate, and the stability characteristic is the standard deviation of the inter-frame displacement of the skeleton point in the continuous 10 rounds of motion; The steps of setting the actual recognition interval of the recognition model according to the user historical body measurement analysis data are as follows:

[0025] Set the initial recognition interval according to the motion cycle characteristics;

[0026] Compare the standard deviation of the inter-frame displacement of the skeleton point with the two preset threshold values BZ1 and BZ2, BZ1 < BZ2;

[0027] If the standard deviation of the inter-frame displacement of the skeleton point is less than BZ1, it is determined that the stability is high, the actual recognition interval = the basic recognition interval × AT1, AT1 is the adjustment coefficient corresponding to high stability, if the actual recognition interval at this time is greater than the maximum interval threshold, the maximum interval threshold is used as the actual recognition interval, otherwise, no processing is performed;

[0028] If BZ1 ≤ the standard deviation of the inter-frame displacement of the skeleton point ≤ BZ2, it is determined that the stability is medium, the actual recognition interval = the basic recognition interval × AT2, AT2 is the adjustment coefficient corresponding to medium stability;

[0029] If the standard deviation of the inter-frame displacement of the skeleton point is greater than BZ2, it is determined that the stability is low, the actual recognition interval = the basic recognition interval × AT3, AT3 is the adjustment coefficient corresponding to low stability;

[0030] During the body measurement process, the standard deviation of the inter-frame displacement of the skeleton point is recalculated every K rounds of motion, 5 < K < 10, if the deviation of the standard deviation of the inter-frame displacement of the skeleton point and the stability characteristic is greater than 3 pixels, the adjustment coefficient is dynamically updated.

[0031] Preferably, the recognition model comprises an input layer, a feature extraction layer, an attention layer, a prediction layer and an output layer. The input layer inputs the pre-processed body measurement data after standardization processing, and outputs a tensor with uniform dimensions. The feature extraction layer inputs the tensor with uniform dimensions, extracts spatial and temporal features, filters redundant information, and outputs a high-dimensional feature vector. The attention layer inputs the high-dimensional feature vector, calculates the weight of each feature to focus on key information, and outputs a weighted fusion feature. The prediction layer inputs the weighted fusion feature, performs full connection layer or classifier operation, and outputs a prediction result. The output layer inputs the prediction result, performs format processing, and outputs a list of N core skeletal points and the confidence of the core skeletal points.

[0032] Preferably, when there is an occlusion, the steps of executing the repair strategy are as follows:

[0033] Based on the current frame timestamp, query the video frames within the timestamp ± 20 ms of the side view angle and the auxiliary view angle;

[0034] Verify the effectiveness of the side view angle and the auxiliary view angle in priority order. When the verification meets the requirements, use a quadratic polynomial interpolation formula to calculate the frame coordinates of each frame;

[0035] Use the human physiological constraint relationship combined with symmetric joint repair to repair the missing core skeletal points through mathematical interpolation and physiological constraints;

[0036] Fuse all core skeletal points at the same timestamp, convert and output the core skeletal point sequence.

[0037] Preferably, the core skeletal point sequence is subjected to action feature extraction to realize action division, which includes calculating spatial features based on the core skeletal point sequence. The spatial features include a multi-dimensional spatial feature vector and a multi-dimensional angle feature vector.

[0038] Based on the core skeletal point sequence, calculate the time features, which include a multi-dimensional velocity vector, a multi-dimensional acceleration vector and a multi-dimensional time sequence feature vector.

[0039] Standardize the spatial features and the time features through min-max standardization;

[0040] Fuse the standardized time features and spatial features to obtain a multi-dimensional action feature vector.

[0041] Analyze the core skeletal point sequence to obtain continuous data of joint angle changes over time, i.e., an angle sequence.

[0042] Based on the fusion features and the angle sequence, realize automatic segmentation of the action cycle.

[0043] Preferably, the steps of eligibility scoring are as follows:

[0044] The core evaluation index score and the auxiliary evaluation index score are obtained by calculating the cosine similarity of the current action period feature and the corresponding part of the template feature multiplied by 100;

[0045] The eligibility score is obtained by weighting the core evaluation index score and the auxiliary evaluation index score;

[0046] The eligibility score is compared with the preset score standard threshold value, if the eligibility score is greater than or equal to the score standard threshold value, it is included in the qualified category, otherwise it is included in the unqualified category.

[0047] Preferably, when the student's ability is comprehensively evaluated based on the evaluation report, the individual scores of each physical test item are first calculated, and then the individual scores of all physical test items are weighted and summed to obtain a comprehensive score;

[0048] According to the eligibility score and the comprehensive evaluation, the evaluation reports of all physical test items of the user in recent times are obtained, the unqualified number and unqualified reasons in each physical test item are extracted, the unqualified reasons are analyzed to obtain a defect reason report, the defect level is judged based on the defect reason report, if the defect level is higher than the set level, no processing is done, otherwise, it is marked as a weak item, and the training plan corresponding to the weak item is searched from the database.

[0049] Preferably, the physical test action detection system based on behavior feature recognition includes a first acquisition module, a model development module, a second acquisition module, a first analysis module, a second analysis module, a physical test evaluation module and a training adjustment module.

[0050] The first acquisition module obtains the standard data of each physical test item, and judges the core skeletal points of each physical test item;

[0051] The model development module constructs a recognition model based on the core skeletal points;

[0052] The second acquisition module obtains the physical test data of the user for each physical test item, and pre-processes the physical test data to collect historical physical test analysis data of the user;

[0053] The first analysis module sets the actual recognition interval of the recognition model according to the historical physical test analysis data of the user;

[0054] The second analysis module sends the pre-processed physical test data to the recognition model with the set actual recognition interval, performs single-view skeletal point detection, and executes a repair strategy when there is an occlusion to obtain a core skeletal point sequence;

[0055] The physical test evaluation module extracts the action features of the core skeletal point sequence, realizes action division, and performs eligibility scoring to generate an evaluation report;

[0056] The training adjustment module: based on the evaluation report, the student's ability is comprehensively evaluated, and the adjustment strategy is generated according to the eligibility score and comprehensive evaluation.

[0057] The present application has the following advantages:

[0058] 1. The present application provides a body measurement action detection method and system based on behavior feature recognition, which constructs a standardized body measurement system, changes the disadvantages of traditional manual observation relying on experience judgment, and realizes full-automatic recognition detection from data acquisition, preprocessing to action evaluation by constructing a recognition model through standard data combined with core skeleton points, adopts unified skeleton point detection standard, scoring rules and data processing mode, eliminates the difference of subjective judgment of different judges, ensures the fairness of body measurement results, lays a foundation for body measurement intelligentization, realizes standardized body measurement, and greatly improves the body measurement efficiency and reliability.

[0059] 2. The present application provides a body measurement action detection method and system based on behavior feature recognition, which marks high dynamic points by calculating the coordinate and angle change rate of L skeleton points, combines the importance score of sports experts, accurately positions the skeleton points which play a decisive role in action evaluation, solves the problems of resource waste and insufficient precision caused by traditional skeleton point non-differential detection, greatly reduces the data processing amount during subsequent recognition analysis, has good use effect, and has good use prospect.

[0060] 3. The present application provides a body measurement action detection method and system based on behavior feature recognition, which constructs actual recognition interval technology through the action cycle and stability characteristics of user historical body measurement data, solves the problems of key action missing detection and data redundancy caused by traditional scheme using fixed interval detection, can guarantee that the core action stage is not missed, and maximally reduces invalid recognition under the premise, greatly reduces the data processing amount during recognition analysis, has good use effect, and has good use prospect.

[0061] 4. The present application provides a body measurement action detection method and system based on behavior feature recognition, which adopts space-time feature coding and automatic scoring technology, solves the defects of traditional manual action segmentation ambiguity and subjective scoring, further sets the scoring mode of action, realizes the judgment of action during body measurement, can make the scoring objective and quantitative, has good use effect, and has good use prospect.

[0062] 5. This invention provides a method and system for detecting physical fitness test movements based on behavioral feature recognition. It solves the problem of delayed feedback in traditional physical fitness tests. During real-time training, it identifies the type of deviation by analyzing the angle of skeletal points in real time, judges the training intensity by combining heart rate data, and triggers corresponding voice correction or rest. It also extracts the data from the last three physical fitness tests and divides strength, endurance, and standard of performance into three defect levels. It matches low / medium / high intensity training plans and provides targeted adjustments for items that are improving slowly. It also allows managers to customize plans to improve weak links, reduce the risk of sports injuries, and has good results and promising application prospects. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the overall method of the present invention;

[0064] Figure 2 This is a flowchart of the process for constructing the recognition model in this invention;

[0065] Figure 3 This is a flowchart illustrating the actual recognition interval of the recognition model in this invention;

[0066] Figure 4 This is a flowchart illustrating the analysis of the core skeletal point sequence in this invention.

[0067] Figure 5 This is a flowchart of the physical fitness assessment in this invention;

[0068] Figure 6 This is a flowchart of the process for generating the adjustment strategy in this invention;

[0069] Figure 7 This is a system framework diagram of the present invention. Detailed Implementation

[0070] 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.

[0071] This invention addresses scenarios where physical education teaching is not highly intelligent, student physical test scores are difficult to record, view, and analyze over a long period, and teachers' inconsistent standards for student physical test movements make it impossible to accurately analyze and judge the results.

[0072] like Figures 1-6 As shown, the body movement detection method based on behavioral feature recognition of the present invention includes the following steps:

[0073] like Figure 2As shown, S1, obtain the standard data of each physical test item, and judge the core skeletal points of each physical test item, and construct an identification model based on the core skeletal points;

[0074] S11, standard data collection;

[0075] The standard data of each physical test item is the standard video of each physical test item and the physical test qualified index, which is derived from the physical test system. The standard video is the action video of the standard of each physical test item completed by the professional sports teachers in the physical test system.

[0076] The physical test qualified index of each physical test item is the qualified / good / excellent frequency standard or time standard of the corresponding age group of each physical test item.

[0077] S12, core skeleton extraction;

[0078] The BlazePose model is used to extract L complete skeletal points related to physical test, and generate a skeletal point coordinate data set for each frame;

[0079] The video is segmented according to the action stage, the key frame of each stage is marked, and the complete action cycle is divided;

[0080] The coordinate change ratio and angle change rate of the L skeletal points in the complete action cycle are calculated, the change amount is calculated by weighting, and the change amount is compared with the set standard change amount. If the change amount is lower than the standard change amount, no processing is done, otherwise, it is marked as a high dynamic point;

[0081] Invite sports experts to score the importance of the skeletal points of each physical test item, 1-10 points, evaluate the influence of the skeletal points on the action standard, and mark the skeletal points with a score higher than 7 as important scores;

[0082] The expert scoring needs to combine the action principle of the physical test item, such as the influence of the skeletal point movement amplitude on the action effectiveness, the necessity of the skeletal point position stability on the action evaluation, the strong correlation of the skeletal point movement trajectory and the action specification, etc. Multiple dimensions are comprehensively scored.

[0083] Take the union of the high dynamic points and the important scoring skeletal points to determine the core skeletal points of each physical test item;

[0084] For example, the jumping jacks project, the hip-knee angle change amount is 45%, the shoulder-elbow movement amplitude is 38%, after screening, and the expert scores of hip, knee, ankle, shoulder and elbow are all >8.2 points, so the core skeletal points corresponding to the jumping jacks project are hip, knee, ankle, shoulder and elbow.

[0085] For example, the squat project, the hip-knee angle determines the change of 58%, the shoulder trajectory reflects the verticality of the trunk, and after screening, the expert scores of hip, knee, ankle and shoulder are all >8.1 points, so the core skeletal points of the squat project are hip, knee, ankle and shoulder;

[0086] For example, the chin-up project, the shoulder-elbow angle changes by 42%, the head position determines whether to pass the bar, and the expert score weight accounts for 65%, after screening, the expert scores of shoulder, elbow, wrist, head and hip are all >8.3 points, so the core skeletal points of the chin-up project are shoulder, elbow, wrist, head and hip;

[0087] By combining objective data with subjective experience, the skeletal points that play a decisive role in action evaluation are accurately positioned, providing a basis for model focusing on key features, and avoiding resource waste caused by indiscriminate detection.

[0088] S13, initial model construction;

[0089] Obtain the core skeletal points of each physical test project;

[0090] Cut the local area of the core skeletal points from the historical physical test frequency frame, and label the coordinate change proportion and angle change rate;

[0091] Enhance the core point features: simulate feature changes under different lighting conditions through Gaussian blur and contrast adjustment to increase the number of samples;

[0092] Establish a core point and non-core point feature distinction mechanism, mark the activation area of the core point feature in the feature map, provide high-quality core point feature samples for the model, strengthen the feature difference between the core point and the non-core point, and help the model learn the feature mode of the key part faster.

[0093] Construct an initial model, including an input layer, a feature extraction layer, an attention layer, a prediction layer and an output layer;

[0094] Input the RGB image obtained after uniform scaling in the input layer;

[0095] The input layer can embed a dynamic cropping layer, which automatically focuses on the human body area according to the physical test project, and the cropping area is dynamically adjusted based on the human body detection result in the preprocessing stage;

[0096] The input layer performs batch normalization on the input features through the built-in batch normalization layer, reduces the influence of lighting changes on skeletal point detection, and speeds up model convergence;

[0097] The feature extraction layer adopts 8 layers of lightweight grouped spatial convolution, which includes four stages, two layers in each stage, the first stage is used to capture low-order edge features, the second stage captures middle-layer texture features, strengthens the key point contour feature extraction, the third stage generates high-order semantic features, and preliminarily fuses the spatial correlation information of the skeleton points, and the fourth stage optimizes the feature channel dimension to provide simplified feature input for the attention layer, ensuring the real-time performance of the model on edge devices.

[0098] After each layer of lightweight grouped spatial convolution, a batch normalization layer and a RelU activation function are connected, the batch normalization layer normalizes the input data of each layer, and the RelU activation function maps the input to a nonlinear output, inhibits overfitting and enhances the nonlinear expression ability.

[0099] The attention layer is two compression and excitation attention modules inserted between the third and fourth stages of the feature extraction layer, where compression is global average pooling of the input feature map to generate a 256-dimensional channel description vector to capture channel-level feature importance, and excitation is through a 2-layer fully connected network and Sigmoid activation to output a 256-dimensional attention weight.

[0100] Based on the importance analysis result of the skeleton points, higher initial weights are given to the feature channels corresponding to the core skeleton points, and the weights are dynamically optimized through training; the attention weight is multiplied with the input feature map channel by channel to enhance the core skeleton point features and suppress irrelevant background noise.

[0101] The prediction layer includes 3 groups of deconvolution layers and 2 groups of convolution layers, the 3 groups of deconvolution layers cooperate to realize the recovery of medium-scale feature details, the enhancement of skeleton point position resolution, and the output of feature maps matching the input size, and the 2 groups of convolution layers cooperate to realize the detection confidence of each skeleton point.

[0102] The output layer outputs the coordinates and confidence of the core skeleton points, providing accurate and simplified skeleton point data for subsequent action counting and standard degree analysis, and the confidence is the reliability score of the skeleton point detection result.

[0103] S14, training and verification;

[0104] The body measurement action data set is adopted, which includes at least 50,000 labeled skeleton points of body measurement video frames, the training data is enhanced, the enhancement includes random rotation, scaling and brightness adjustment, the core skeleton point area is cropped, the key part feature learning is strengthened, the number of enhanced pictures is 3-5 times the original number, and the initial model is trained using the enhanced pictures.

[0105] The loss function adopts a weighted loss function, specifically, coordinate loss and confidence loss are weighted, mean square error loss is adopted for core skeleton points and non-core skeleton points, the weight of the core skeleton point is higher than that of the non-core skeleton point, and the confidence loss is cross-entropy loss, the loss weighting guides the model to preferentially learn the core skeleton point features, and solves the problem of insufficient detection accuracy of key parts.

[0106] The optimizer adopts an Adam optimizer, the initial learning rate is 1e-4, the weight decay is 1e-5, and e is the symbol of scientific notation;

[0107] If the core point error of the verification set does not decrease continuously for 10 rounds, the training is stopped, and the optimal model, i.e., the recognition model, is saved.

[0108] The verification set contains at least 5000 body measurement action video frames, and the real coordinates of the labeled core skeleton points are verified.

[0109] The evaluation index during verification is that the average joint point error is less than or equal to 8 pixels, the non-core point error is less than or equal to 10 pixels, the core skeleton point detection rate is that the proportion of core skeleton points with a confidence degree greater than or equal to 0.6 is greater than or equal to 95%.

[0110] The application provides a body measurement action detection method and system based on behavior feature recognition, high dynamic points are marked by calculating L skeleton point coordinates and angle change rates, and important scores of sports experts are combined to accurately locate skeleton points that play a decisive role in action evaluation, so that the problems of resource waste and insufficient precision caused by traditional skeleton point non-discriminatory detection are solved, the data processing amount during subsequent recognition and analysis is greatly reduced, the use effect is good, and the application has a good use prospect.

[0111] As shown in Figure 3 S2, body measurement data of each body measurement item of a user is obtained, and the body measurement data is preprocessed, historical body measurement analysis data analysis results are collected, and actual recognition intervals of a recognition model are set accordingly;

[0112] S21, body measurement data collection;

[0113] The body measurement data is collected through body measurement software, and the body measurement data includes body measurement videos of each view angle during body measurement, body measurement devices worn on the body of a body measurement person during body measurement, and measurement data collected by a body measurement instrument;

[0114] For example, the body measurement videos of each view angle during body measurement are usually videos of a main view angle, a side view angle and an auxiliary view angle, full-angle coverage is achieved, the body measurement instrument worn on the body of the body measurement person during body measurement includes a motion sensor, a physiological sensor and an environmental sensor, and the body measurement data collected by the body measurement device includes data collected by a body fat scale and an intelligent skipping rope, a pressure sensor in a skipping rope area and the like.

[0115] S22, body measurement data resampling;

[0116] Trigger data re-sampling when there are video data incompleteness, measurement data incompleteness and synchronization anomalies collected by body measurement equipment and body measurement instruments;

[0117] Video data incompleteness includes frame rate fluctuation > ± 5FPS, key frame loss > 3 frames, picture occlusion > 30% area;

[0118] Measurement data incompleteness collected by body measurement equipment and body measurement instruments includes sampling rate loss > 10%, key parameters exceeding the set range;

[0119] Synchronization anomalies include timestamp deviation > 50ms between video and measurement data collected by body measurement equipment and body measurement instruments.

[0120] Timestamp deviation is detected in real time. When in use, the body measurement equipment is connected to a unified NTP time server, which calibrates the local clock every 100ms. Video frame timestamps and measurement data timestamps are compared through a timestamp mapping table. If the deviation of three consecutive frames is > 50ms, data re-sampling is triggered.

[0121] During re-sampling, integrity check is performed every 5s on the collected body measurement data, and abnormal data segments are marked.

[0122] During the re-sampling process, if a single frame is lost, it is repaired through an interpolation algorithm, and data re-sampling is not triggered.

[0123] S23, data preprocessing;

[0124] Existing body measurement equipment and body measurement instruments are filtered when collecting and uploading measurement data to body measurement software, so separate preprocessing is not required.

[0125] Therefore, preprocessing is mainly for video data, and the specific processing method for video data is as follows:

[0126] Frame rate processing: Extract frame data from body measurement videos of each view through FFmpeg technology according to fixed timestamps, interpolate or decimate different frame rate videos, unify frame rate, and use bilinear interpolation algorithm for decimation.

[0127] Denoising and enhancement processing: Use adaptive Gaussian filtering to remove salt and pepper noise in videos after unifying frame rate, then use non-local mean denoising algorithm to enhance edge definition of motion blur frames, then balance brightness through gamma correction, and finally use limited contrast adaptive histogram equalization algorithm to enhance local contrast, ensuring that bone point features can be recognized under complex light.

[0128] Crop processing: use YOLOv5s model to detect human regions in video frames and output bounding box coordinates;

[0129] The video frame is cropped according to the bounding box, the human body region and surrounding 20% redundant pixels are reserved, and the output size of the video frame is uniform.

[0130] The data processing eliminates environmental interference in the video data, provides high-quality input for skeleton point extraction, and improves the subsequent detection accuracy.

[0131] After the video data is processed, the processed video data, the measurement data collected by the body measurement device and the body measurement instrument are added with high-precision timestamps, at least to the millisecond level, and then a timestamp mapping table is constructed to linearly interpolate the measurement data collected by the body measurement device and the body measurement instrument, so that each video frame corresponds to a unique data point.

[0132] The measurement data collected by the body measurement device and the body measurement instrument are normalized.

[0133] During the frame rate processing, the motion needs to be processed by frame segmentation, which is divided according to the motion period, and the key frames of each stage are marked, for example, the deep squat project includes three stages of initial, lower end and rising.

[0134] S24, actual recognition interval setting:

[0135] The user's historical body measurement data analysis result includes motion period characteristics and stability characteristics, the motion period characteristics include single round motion average period length, and the motion rhythm is determined by the inter-frame skeleton point change rate;

[0136] The stability characteristic is the standard deviation of the inter-frame displacement of the skeleton point in the continuous 10-round motion;

[0137] The initial recognition interval is set according to the motion period characteristics, specifically, the basic recognition interval = motion period length x basic frame rate x A1; A1 is a safety factor, the basic frame rate is the uniform video frame rate after preprocessing, which avoids missing key motion stages due to too long interval, and 0.5

[0138] Setting the basic recognition interval ensures that the core motion stage is not missed, while avoiding meaningless high-frequency recognition;

[0139] The basic recognition interval is further adjusted according to the standard deviation of the inter-frame displacement of the skeleton point to obtain the actual recognition interval;

[0140] The specific steps are to compare the standard deviation of the inter-frame displacement of the skeleton point with two groups of preset threshold values; the two groups of threshold values are BZ1 and BZ2, and BZ1

[0141] If the standard deviation of the inter-frame displacement of the skeleton point is less than BZ1, it is determined that the stability is high, and the actual recognition interval = basic recognition interval x AT1, AT1 is the adjustment coefficient corresponding to high stability, and 1.5

[0142] If BZ1≤ the standard deviation of the inter-frame displacement of the skeletal point ≤ BZ2, it is determined that the stability is medium, the actual recognition interval = the basic recognition interval × AT2, AT2 is the adjustment coefficient corresponding to medium stability, 0.8 < AT2 < 1.2;

[0143] If the standard deviation of the inter-frame displacement of the skeletal point > BZ2, it is determined that the stability is low, the actual recognition interval = the basic recognition interval × AT3, AT3 is the adjustment coefficient corresponding to low stability, 0.6 < AT3 < 0.8;

[0144] Real-time correction mechanism: during the body measurement process, the standard deviation of the inter-frame displacement of the skeletal point is recalculated every K rounds of action, 5 < K < 10, if the deviation of the standard deviation of the inter-frame displacement of the skeletal point from the stability feature > 3 pixels, the adjustment coefficient is dynamically updated;

[0145] For example, the user's history is determined to have high stability, but the real-time analysis shows that the standard deviation of the inter-frame displacement of the skeletal point is in the medium stability range, so the adjustment coefficient for medium stability is used to adjust the set basic recognition interval.

[0146] The real-time correction maximizes the reduction of the recognition times under the premise of ensuring the recognition accuracy, and the data processing amount in a high-stability scene can be reduced.

[0147] In actual use, to avoid missing key actions due to too long intervals, a maximum interval threshold for the body measurement project is set, and when the calculated actual recognition interval > the maximum interval threshold, the maximum interval threshold is used as the actual recognition interval.

[0148] As shown in FIG. 3, Figure 4 S3, the preprocessed body measurement data is delivered to the recognition model corresponding to the set actual recognition interval, single-view skeletal point detection is performed, and a repair strategy is executed when there is occlusion, to obtain a core skeletal point sequence;

[0149] S31, skeletal point detection;

[0150] The body measurement data are all classified data, so the data corresponding to the body measurement project are input into the corresponding recognition model, the body measurement data are recognized according to the set actual recognition interval, to obtain a list of N core skeletal points and the confidence of the core skeletal points;

[0151] That is, the model loads the core skeletal points and the actual recognition interval of the body measurement project before use, selectively samples the preprocessed video frames according to the actual recognition interval, and only the relevant data of the core skeletal points are sampled, which greatly reduces the data processing amount.

[0152] S32, occlusion judgment;

[0153] Calculate the average confidence of the core skeleton points in the list of single core skeleton points in real time. If the proportion of core points with confidence <0.6 is ≥30%, mark it as instantaneous occlusion.

[0154] For example, the confidence of 2 core skeleton points in 6 core skeleton points is low.

[0155] If the confidence of the same core point decreases by >0.3 for 2 consecutive times, trigger an occlusion warning.

[0156] Track the trajectory of the core skeleton point identified for 5 consecutive times, calculate the standard deviation of inter-frame displacement, and if the displacement of a certain frame is >3 inter-frame displacement standard deviations and the confidence of the core skeleton point is <0.5, determine it as continuous occlusion.

[0157] S33, occlusion repair;

[0158] After determining occlusion, query the video frames within ±20 ms of the current frame timestamp based on the current frame timestamp, query the side view angle, and auxiliary view angle within the same timestamp;

[0159] The view priority is main view angle > side view angle > auxiliary view angle. If there is occlusion in the high-priority view angle, then the low-priority view angle is called in turn.

[0160] When using a low-priority view angle, effectiveness verification is required, i.e., verifying the completeness rate of the core skeleton points and the average confidence of the core skeleton points to determine whether they meet the standards. For example, when there are 6 core skeleton points, at least 5 are detected, and the average confidence of the core skeleton points is >0.7.

[0161] When it is determined that the called view angle meets the requirements, the frame coordinates of each frame are calculated using a quadratic polynomial interpolation formula. The quadratic polynomial interpolation formula is an existing formula, so it will not be described in detail:

[0162] Using human physiological constraints combined with symmetric joint repair, precise repair of missing skeleton points is achieved through mathematical interpolation and physiological constraints, ensuring the continuity of the skeleton point sequence and providing complete data for subsequent action analysis.

[0163] S34, skeleton point fusion;

[0164] Coordinate unification: Zhang Zhengyou calibration method is used to calibrate the internal parameters of the body measurement site using a checkerboard calibration board, and the internal parameter matrix and distortion coefficient of each camera are output.

[0165] Taking the main view angle as the origin of the world coordinate system, the rotation matrix and translation vector of the side view angle and auxiliary view angle relative to the main view angle are calculated through feature point matching to realize the unification of multi-view coordinate systems.

[0166] Apply distortion coefficients to correct the original images of all view angles to eliminate the skeleton point offset caused by lens distortion.

[0167] Weight distribution: calculate the quality index of each view, the quality index includes the definition score, the illumination score and the occlusion rate score, and the definition score, the illumination score and the occlusion rate score are weighted and summed according to weights to obtain the comprehensive quality score of each view at the same time;

[0168] Skeleton point coordinate weighted fusion: the pixel coordinates of the skeleton points detected by each view are converted to the world coordinate system through the inverse matrix of the internal parameter and the external parameter;

[0169] The standard deviation of the multi-view world coordinates of the same core skeleton point is analyzed, if the deviation of the coordinates of a view from the mean value is greater than 3 times the standard deviation, the view data is marked as an abnormal value, and the effective view data is obtained by temporarily excluding the view data;

[0170] The average position is calculated by weighted sum of the world coordinates of the effective view according to the weight, and the core skeleton point coordinate closest to the true value is obtained;

[0171] S35, skeleton point output;

[0172] The fused core skeleton point coordinates are sorted according to the time stamp to generate a continuous time sequence, and the core skeleton point sequence in JSON format is converted and output; the core skeleton point sequence includes basic information: body measurement item, user name, time stamp range, skeleton point data and occlusion repair ratio; the skeleton point data is the core skeleton point, world coordinate and confidence degree corresponding to each time stamp.

[0173] The application provides a body measurement action detection method and system based on behavior feature recognition, which constructs an actual recognition interval technology through the action cycle and stability features of the user's historical body measurement data, solves the problem that the traditional scheme adopts fixed interval detection and is easy to miss key actions and has data redundancy, can maximize reduce invalid recognition under the premise of ensuring that the core action stage is not missed, greatly reduces the data processing amount during recognition analysis, has good use effect, and has good use prospect.

[0174] As shown in Figure 5 S4, action feature extraction is performed on the core skeleton point sequence, action division is realized, and eligibility scoring is performed to generate an evaluation report;

[0175] S41, space-time feature coding;

[0176] Space feature extraction: based on the core skeleton point sequence, the Euclidean distance and relative coordinate offset between the core skeleton points are calculated to generate a multi-dimensional space feature vector; the angle calculation rule is called, for example, the trunk inclination angle of deep squat, the angle is calculated frame by frame through the vector dot product formula, and the multi-dimensional angle feature vector is expanded;

[0177] Time feature extraction: calculate the displacement difference and time difference of the core skeleton points of consecutive frames to obtain the velocity feature and generate a multi-dimensional velocity vector; perform first-order difference on the velocity feature to obtain the acceleration feature, which reflects the action force intensity, and generate a multi-dimensional acceleration vector; encode the time-space feature sequence of consecutive M1 frames using a bidirectional LSTM network to output a multi-dimensional time sequence feature vector;

[0178] Compress the multi-dimensional space feature vector, multi-dimensional angle feature vector, multi-dimensional velocity vector, multi-dimensional acceleration vector, and multi-dimensional time sequence feature vector to the [0, 1] interval through min-max standardization;

[0179] Feature fusion: fuse the standardized multi-dimensional space feature vector, multi-dimensional angle feature vector, multi-dimensional velocity vector, multi-dimensional acceleration vector, and multi-dimensional time sequence feature vector through a feature concatenation layer to output a multi-dimensional action feature vector.

[0180] S42, action division;

[0181] Based on the multi-dimensional action feature vector and the angle sequence, the action cycle is automatically segmented. The angle sequence is continuous data of joint angles changing over time. It is obtained by analyzing the core skeleton point coordinate sequence, for example, the knee and elbow joint angles are calculated by vector dot product, and are arranged in frame time order to form an angle sequence. When the key angle first falls from the static threshold and the velocity feature is >0.5 pixels / ms, it is marked as the starting frame of the cycle. When the key angle rises to the static threshold and the velocity feature is <0.3 pixels / ms, and at the same time, it remains stable for 3 consecutive frames, it is marked as the end frame of the cycle. At this time, the time interval from the starting frame to the end frame of the cycle is calculated, which is the time length of one round of action. The number of end frames is counted to realize action counting. When in use, a count anti-shake mechanism is set: only when the end frame state is stable for 3 consecutive frames, it is counted as valid count to avoid misjudgment when the action is not completed.

[0182] S43, eligibility score;

[0183] Based on the action cycle feature and the preset standard, the eligibility is automatically judged, and the evaluation index is set for each project;

[0184] The evaluation index includes core evaluation index and auxiliary evaluation index, which is specifically set according to the body measurement project;

[0185] For example, in the deep squat project, the core evaluation index is the minimum angle of the knee joint, and the auxiliary evaluation index is the trunk inclination angle and the time interval from the starting frame to the end frame of the cycle;

[0186] For example, in the pull-up project, the core evaluation index is the jaw over the bar, and the auxiliary evaluation index is the elbow joint extension angle and the time interval from the starting frame to the end frame of the cycle;

[0187] The score of the core evaluation index and the auxiliary evaluation index is obtained by calculating the cosine similarity between the current action period feature and the corresponding part of the template feature multiplied by 100.

[0188] The core evaluation index score is calculated based on the matching degree of the action key feature and the standard template. First, the core feature is extracted from the current action period, such as the knee joint angle sequence of deep squat, the elbow joint angle change of pull-up, etc. These time series data are converted into feature vectors, and the dimension difference is eliminated by Z-scor standardization. The standard feature vector of the standard action data is called, which contains the standard range and time trend of joint angle. The matching degree of the current feature vector and the template vector is calculated by the cosine similarity formula, and the similarity value in the range of [-1, 1] is obtained. Finally, the similarity is multiplied by 100 to convert it into a core score of 0-100 points.

[0189] The auxiliary evaluation index score focuses on the quality of action details, and needs to extract symmetry, rhythm and other derived features. The auxiliary features are obtained from the current action period, such as the left and right knee joint angle difference sequence, the deviation rate of action period length and standard length, etc. A multi-dimensional auxiliary feature vector is constructed and standardized. The standard auxiliary vector of the standard action data is called, which contains the left and right symmetry threshold, rhythm stability parameter and other reference data. The similarity between the current auxiliary vector and the template vector is calculated by the cosine similarity formula, and the similarity is multiplied by 100 to obtain an auxiliary score of 0-100 points.

[0190] In use, the core skeleton point parameter can also be used to calculate the eligibility score, for example, the Euclidean distance of the skeleton points between the detected action and the standard template is calculated frame by frame, the part similarity is calculated by weighting, and the overall eligibility score is obtained by weighted summation. However, due to the different development of students, the length of arms and legs is different, and the similarity naturally exists certain difference, therefore, this method cannot meet the requirements of all people, and needs to be modified by experts.

[0191] The core evaluation index score and the auxiliary evaluation index score are weighted to obtain the eligibility score.

[0192] In use, the weight of the core evaluation index score is 0.7-0.8, and the weight of the auxiliary evaluation index score is 0.2-0.3.

[0193] S44, physical test evaluation;

[0194] The eligibility score is compared with the preset score standard threshold. If the eligibility score is greater than or equal to the score standard threshold, it is included in the qualified category; otherwise, it is included in the unqualified category. The number of eligibility scores in the qualified category is counted, and the number of eligibility scores is compared with the physical test eligibility index. If the number of eligibility scores is greater than or equal to the physical test eligibility index, the user is qualified in the physical test project; otherwise, the user is unqualified in the physical test project.

[0195] For the plank support project, the qualified time is compared with the body measurement standard time, and the principle is the same, so it is not described too much.

[0196] When included in the unqualified category, the unqualified reason is marked, including insufficient amplitude, low stability, cycle anomaly and posture error, for example, the core angle does not reach the threshold, the inter-frame displacement standard deviation exceeds the set threshold, and the cycle anomaly includes the time interval from the cycle start frame to the cycle end frame exceeds the preset range.

[0197] After the evaluation is completed, the evaluation report is output, which includes the body measurement project type, total count, qualified number, unqualified number, unqualified reason and other analysis-related data.

[0198] The application provides a body measurement action detection method and system based on behavior feature recognition, which uses space-time feature coding and automatic scoring technology to solve the defects of traditional manual action segmentation ambiguity and subjective scoring, further sets the scoring method of the action, realizes the judgment of the action during the body measurement, can make the scoring objective and quantitative, has good use effect, and has good use prospect.

[0199] As shown in Figure 6 S5, based on the evaluation report, the student's ability is comprehensively evaluated, and an adjustment strategy is generated;

[0200] S51, body measurement score integration;

[0201] Get the evaluation report of all body measurement projects, calculate the single item score, single item score = qualified number / full score number x 100;

[0202] Sum the single item scores of all body measurement projects to get the comprehensive score;

[0203] S52, real-time adjustment strategy;

[0204] Defect analysis; during daily training, the training video is input into the corresponding recognition model in real time, the body measurement data is recognized according to the set actual recognition interval, the list of real-time core skeleton points and the confidence of the core skeleton points are obtained, the core angle is calculated by the vector dot product formula, and the key angle is calculated, for example, the knee angle of deep squat and the elbow angle of pull-up;

[0205] The mean and standard deviation of the angle data of the continuous 5 frames are calculated, the action trend is analyzed, the deviation type is identified, whether there are multiple same problems is judged, the deviation type includes insufficient amplitude, posture error and rhythm anomaly, and the specific skeleton point marker of the deviation is marked;

[0206] Judge the intensity combined with heart rate data, such as judging whether it is in a high-intensity state or a fatigue state by heart rate, heart rate > maximum heart rate x 85% is a high-intensity state, and the maximum heart rate is calculated by the international general formula, and the acceleration standard deviation > 2g and the continuous rise of heart rate is a fatigue state, g is the acceleration of gravity;

[0207] Strategy generation: start action correction strategy when 2 consecutive frames do not meet the standard and single deviation occurs 3 times; start rest prompt strategy when in high-intensity state and potential fatigue;

[0208] The action correction strategy issues a voice instruction and broadcasts the correction method. If there is no correction for 3 consecutive times, the real-time qualified score is lower than 5% compared with the score before correction, the training video and standard video are played synchronously, the specific bone points of the deviation are identified in the training video, and the training data of this time is stored.

[0209] For example, for the problem of insufficient knee joint angle in deep squat, the voice instruction is to squat to a knee joint angle ≥ 90°; for the problem of insufficient elbow extension in pull-up, the voice instruction is to fully extend the elbow joint when falling back.

[0210] S53, comprehensive adjustment strategy;

[0211] Defect analysis; obtain the evaluation report of all body measurement items of the user, including the evaluation report of the last 3 times of body measurement, extract the number of unqualified items and the reason for unqualification in each body measurement item;

[0212] Analyze the unqualified reasons to get the defect reason report, and the defect reasons in the defect reason report include strength, endurance and standard degree;

[0213] Strength includes upper limb strength and lower limb strength, upper limb strength is evaluated by pull-up single item score, and lower limb strength is evaluated by deep squat single item score; endurance is evaluated by skipping rope single item score, and standard degree is evaluated by comprehensive score;

[0214] The specific evaluation is to compare the corresponding score with the corresponding grade interval to judge the defect level;

[0215] If the defect level is lower than the set level, it is marked as a weak item; for the weak item, compare the defect reason reports at different times to identify whether the improvement proportion of the corresponding single item score of the weak item is lower than 3%, if it is lower, it is marked as an urgent adjustment item;

[0216] For the weak item, the adjustment strategy is formulated according to the defect level, which is to find the training plan corresponding to the defect level of the corresponding weak item from the database, for example, upper limb strength is weak, then develop pull-up training, and lower limb strength is weak, then develop deep squat variable training;

[0217] The defect level is divided into three levels, which is screened by two standard level thresholds DJ1 and DJ2, if the defect level is greater than DJ2, the training plan corresponding to the defect level is high-intensity training, in this case, there is a slight weakness, which needs higher intensity to improve, for example, chin-up training, using weighted chin-up;

[0218] The standard level thresholds DJ1 and DJ2 are preset thresholds, which are determined according to the body measurement standard data statistics of the same age group.

[0219] If DJ1≤defect level≤DJ2, the training plan corresponding to the defect level is medium-intensity training, in this case, it is relatively weak, which needs medium-intensity training to improve, if high-intensity training is used, it is easy to be injured, for example, chin-up training, using self-weight chin-up;

[0220] If the defect level is less than DJ1, the training plan corresponding to the defect level is low-intensity training, in this case, it is very weak, which needs low-intensity training to improve, if high-intensity training and medium-intensity training are used, it is easy to be injured, for example, chin-up training, using elastic band assisted chin-up.

[0221] For the items that need to be adjusted urgently, relevant data is sent to the management personnel, and the management personnel redevelops the training plan according to the training plan of the weak item and the single item score improvement ratio.

[0222] The relevant data includes the single item score and comprehensive score evaluation trend graph of the weak item in the last three times, the current training plan execution record, and the bone point deviation heat map, the management personnel can view real-time data through the system background and adjust the training parameters online.

[0223] The application provides a body measurement action detection method and system based on behavior feature recognition, which solves the problem of traditional body measurement feedback lag, identifies the deviation type through real-time bone point angle analysis in the process of real-time training, judges the training intensity combined with heart rate data, triggers voice correction or rest, extracts the last three body measurement data, divides the strength, endurance and standard degree into three defect levels, matches low / medium / high intensity training plan, adjusts the items that need to be adjusted urgently, customizes the plan with the management personnel, improves the weak links, reduces the risk of sports injury, has good use effect, and has good use prospect.

[0224] As shown in Figure 7 The body measurement action detection system based on behavior feature recognition includes a first acquisition module, a model development module, a second acquisition module, a first analysis module, a second analysis module, a body measurement evaluation module and a training adjustment module.

[0225] The first acquisition module obtains the standard data of each body measurement item, and judges the core bone points of each body measurement item.

[0226] Modeling module: construct a recognition model based on the core skeletal points;

[0227] Second acquisition module: obtain physical measurement data of each physical measurement item of the user, and pre-process the physical measurement data to acquire historical physical measurement analysis data of the user;

[0228] First analysis module: set an actual recognition interval of the recognition model according to the historical physical measurement analysis data of the user;

[0229] Second analysis module: deliver the pre-processed physical measurement data to the recognition model corresponding to the set actual recognition interval, perform single-view skeletal point detection, and execute a repair strategy when there is an occlusion to obtain a core skeletal point sequence;

[0230] Physical measurement evaluation module: perform action feature extraction on the core skeletal point sequence, realize action division, and perform eligibility scoring to generate an evaluation report;

[0231] Training adjustment module: comprehensively evaluate the student's ability based on the evaluation report, and generate an adjustment strategy according to the eligibility score and the comprehensive evaluation.

[0232] The application provides a physical measurement action detection method and system based on behavior feature recognition, constructs a standardized physical measurement system, changes the disadvantages of traditional manual observation relying on experience judgment, and realizes full-automatic recognition detection from data acquisition, preprocessing to action evaluation by constructing a recognition model through standard data combined with core skeletal points, adopts a unified skeletal point detection standard, scoring rules and data processing mode, eliminates differences in subjective judgment of different judges, ensures fairness of physical measurement results, lays a foundation for physical measurement intelligence, realizes standardized physical measurement, and greatly improves physical measurement efficiency and reliability.

[0233] The above embodiments can be realized by software, hardware, firmware or any combination thereof, in whole or in part. When realized by software, the above embodiments can be realized in the form of a computer program product in whole or in part. Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present text can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solutions.

[0234] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, which can be located in one place or distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0235] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A body motion detection method based on behavior feature recognition, characterized by, The method comprises the following steps: S1, obtaining standard data of each physical test item, judging core skeletal points of each physical test item, and constructing an identification model based on the core skeletal points; S2, obtaining physical test data of each physical test item of a user, preprocessing the physical test data, collecting historical physical test analysis data of the user, and setting an actual identification interval of the identification model according to the historical physical test analysis data of the user; S3, delivering the preprocessed physical test data to the identification model corresponding to the set actual identification interval, performing single-view skeletal point detection, and executing a repair strategy when there is an occlusion to obtain a core skeletal point sequence; S4, performing action feature extraction on the core skeletal point sequence, realizing action division, and performing a qualified score to generate an evaluation report; S5, comprehensively evaluating the student's ability based on the evaluation report, and generating an adjustment strategy according to the qualified score and the comprehensive evaluation; The user historical physical test data analysis result includes an action cycle feature and a stability feature, the action cycle feature includes a single-round action average cycle time length, and the action rhythm is determined by an inter-frame skeletal point change rate, and the stability feature is a skeletal point inter-frame displacement standard deviation in 10 consecutive rounds of actions; the step of setting the actual identification interval of the identification model according to the user historical physical test analysis data is as follows: An initial identification interval is set according to the action cycle feature; The skeletal point inter-frame displacement standard deviation is compared with two groups of preset threshold values BZ1 and BZ2, and BZ1 < BZ2; If the skeletal point inter-frame displacement standard deviation < BZ1, it is determined that the stability is high, the actual identification interval = the basic identification interval × AT1, AT1 is an adjustment coefficient corresponding to high stability, if the actual identification interval > the maximum interval threshold value at this time, the maximum interval threshold value is used as the actual identification interval, otherwise, no processing is performed; If BZ1 ≤ the skeletal point inter-frame displacement standard deviation ≤ BZ2, it is determined that the stability is medium, the actual identification interval = the basic identification interval × AT2, AT2 is an adjustment coefficient corresponding to medium stability; If the skeletal point inter-frame displacement standard deviation > BZ2, it is determined that the stability is low, the actual identification interval = the basic identification interval × AT3, AT3 is an adjustment coefficient corresponding to low stability; During the physical test, the skeletal point inter-frame displacement standard deviation is recalculated every K rounds of actions, 5 < K < 10, if the deviation of the skeletal point inter-frame displacement standard deviation and the stability feature > 3 pixels, the adjustment coefficient is dynamically updated.

2. The body motion detection method based on behavior feature recognition according to claim 1, wherein, The standard data includes standard videos and physical test qualified indicators of each physical test item, and the steps of judging the core skeletal points of each physical test item are as follows: An BlazePose model is used to extract L complete skeletal points related to physical test, and a skeletal point coordinate data set of each frame is generated; The video is segmented according to the action stage, the key frame of each stage is marked, and a complete action cycle is divided; The coordinate change ratio and the angle change rate of the L skeletal points in the complete action cycle are calculated, the change amount is calculated by weighting, and the change amount is compared with the set standard change amount, if the change amount is lower than the standard change amount, no processing is performed, otherwise, it is marked as a high dynamic point; The importance of each body measurement project bone point is scored, the influence of the bone point on the action standard is evaluated, and the bone point with a score higher than the preset marking score is marked as important; The high dynamic point and the important scored bone point are taken as a set to determine the core bone points of each body measurement project.

3. The body motion detection method based on behavior feature recognition according to claim 1, wherein, The body measurement data includes body measurement videos of each view angle during body measurement, body measurement equipment worn on the body measurement personnel during body measurement, and measurement data collected by body measurement instruments, and the preprocessing of the body measurement data includes: Frame rate processing: the body measurement videos of each view angle are extracted by FFmpeg technology according to fixed timestamps, interpolation or frame extraction processing is performed on different frame rate videos, and the frame rate is unified; Denoising enhancement processing: the salt and pepper noise of the video after unifying the frame rate is removed by using adaptive Gaussian filtering, then the edge definition of the motion blur frame is enhanced by using a non-local mean denoising algorithm, then the brightness is balanced through gamma correction, and finally the local contrast is enhanced by using a limited contrast adaptive histogram equalization algorithm; Crop processing: the human body region in the video frame of the body measurement video after enhancement processing is detected by using a YOLOv5s model, the bounding box coordinates are output, the video frame is cropped according to the bounding box, the human body region and the surrounding 20% redundant pixels are retained, and the output size of the video frame is unified; Normalization processing: the measurement data collected by the body measurement equipment and the body measurement instruments is normalized.

4. The body motion detection method according to claim 3, wherein The recognition model includes an input layer, a feature extraction layer, an attention layer, a prediction layer and an output layer, the input layer inputs the preprocessed body measurement data after standardization processing, outputs a tensor with uniform dimensions, the feature extraction layer inputs the tensor with uniform dimensions, extracts spatial and temporal features, filters redundant information, and outputs a high-dimensional feature vector, the attention layer inputs the high-dimensional feature vector, calculates the feature weight focusing on key information, outputs the weighted fusion feature, the prediction layer inputs the weighted fusion feature, performs full connection layer or classifier operation, and outputs the prediction result, the output layer inputs the prediction result, performs format processing, and outputs the list of N core bone points and the confidence of the core bone points.

5. The body motion detection method based on behavior feature recognition according to claim 1, wherein, When there is an occlusion, the steps of executing the repair strategy are as follows: Based on the current frame timestamp, query the video frames of the side view angle and the auxiliary view angle within ±20 ms of the same timestamp; According to the priority order, the effectiveness of the side view angle and the auxiliary view angle is verified, when the verification meets the requirements, the frame coordinates of each frame are calculated by using a quadratic polynomial interpolation formula; The missing core bone points are repaired by mathematical interpolation and physiological constraints combined with symmetric joint repair; All core bone points with the same timestamp are fused, converted and output as a core bone point sequence. 6.The body motion detection method based on behavior feature recognition according to claim 1, wherein, Action feature extraction is performed on the core bone point sequence to realize action division, including calculating spatial features based on the core bone point sequence, the spatial features including a multi-dimensional spatial feature vector and a multi-dimensional angle feature vector; Time features are calculated based on the core bone point sequence, the time features including a multi-dimensional velocity vector, a multi-dimensional acceleration vector and a multi-dimensional time sequence feature vector; The spatial features and the time features are standardized by min-max standardization; The standardized time features and spatial features are fused to obtain a multi-dimensional action feature vector; The core skeleton point sequence is analyzed to obtain continuous data of joint angle change over time, i.e., an angle sequence; The motion cycle is automatically segmented based on the fusion features and the angle sequence. 7.The body motion detection method based on behavior feature recognition according to claim 1, wherein, The eligibility score is obtained by the following steps: The core evaluation index score and the auxiliary evaluation index score are obtained by multiplying the cosine similarity of the current motion cycle features and the corresponding part of the template features by 100; The eligibility score is obtained by weighting the core evaluation index score and the auxiliary evaluation index score; The eligibility score is compared with the preset score standard threshold value, if the eligibility score is greater than or equal to the score standard threshold value, it is included in the qualified category, otherwise it is included in the unqualified category. 8.The body motion detection method based on behavior feature recognition according to claim 7, wherein, When the student's ability is comprehensively evaluated based on the evaluation report, the individual scores of each individual test item are first calculated, and then the individual scores of all individual test items are weighted and summed to obtain a comprehensive score; When the adjustment strategy is generated based on the eligibility score and the comprehensive evaluation, the evaluation reports of all individual test items of the user in recent times are obtained, the unqualified number and unqualified reasons in each individual test item are extracted, the unqualified reasons are analyzed to obtain a defect reason report, the defect level is determined based on the defect reason report, if the defect level is higher than the set level, no processing is performed, otherwise, the weak item is marked and the training plan corresponding to the weak item is searched from the database.

9. A body motion detection system based on behavior feature recognition using any one of the methods of claims 1 to 8, characterized in that, It includes: The first acquisition module obtains the standard data of each individual test item and judges the core skeleton points of each individual test item; The model development module constructs an identification model based on the core skeleton points; The second acquisition module obtains the physical test data of the user for each individual test item and pre-processes the physical test data to collect historical physical test analysis data of the user; The first analysis module sets the actual identification interval of the identification model based on the historical physical test analysis data of the user; The second analysis module inputs the pre-processed physical test data into the identification model with the set actual identification interval, performs single-view skeleton point detection, and executes a repair strategy when there is an occlusion to obtain a core skeleton point sequence; The physical test evaluation module extracts motion features from the core skeleton point sequence, divides the motion, performs eligibility scoring, and generates an evaluation report; The training adjustment module comprehensively evaluates the student's ability based on the evaluation report and generates an adjustment strategy based on the eligibility score and the comprehensive evaluation.

Images