Squat Posture Analysis Method Based on Multiple Regression

Abstract

A squatted movement standardization evaluation method and device based on multiple regression, comprising the following steps: marking joint points matched with the squatted movement data of a tested person and establishing the coordinates of standardized joint points; setting action evaluation parameter thresholds and evaluating a single squatted movement in the squatted movement data based on the spatial coordinates to generate a determination parameter vector including multiple determination parameters for the single squatted movement; and making a determination on the determination parameter vector to determine whether the corresponding squatted movement meets the standard. The present invention can accurately analyze the movement details of each squat, and obtain a standard counting result according to the measurement standard of the standard squatted movement; moreover, action evaluation parameter thresholds can be set according to action requirements, and the thresholds can be modified according to scene requirements and the conditions of the tested person, with high detection accuracy, strong anti-interference ability, and being more accurate and scientific.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and further relates to a method for recognizing human postures based on multiple regression in the field of computer vision for analyzing the squat postures of testers. The present invention can be used to analyze the standard degree of the squat postures of the testers to be tested and perform standard counting. Background Art

[0002] In the existing technologies, there are the following three methods for judging whether a squat posture is standard: directly judging by a human according to the standardization theory, judging by using sensors, and the method based on computer vision. Directly judging by a human according to the standardization theory often requires the judge to learn the standardization theory by himself / herself, and then conduct a standardization evaluation through naked-eye observation and his / her own experience; the method of judging by using sensors is generally based on an infrared sensor and a pressure sensor. The infrared sensor is used to judge whether the body parts of the exerciser reach the specified positions, and the pressure sensor is used to judge whether the body's center of gravity is evenly distributed. With the continuous development and progress of modern science, the method of judging based on computer vision has come into people's sight. This method first uses a camera and a computer to collect video and image information during the squat process, then designs an algorithm to simplify and analyze the collected image information, and compares it with the standard squat action images, so as to achieve a standardization evaluation.

[0003] Hunan University of Technology and Business disclosed a method for recognizing fitness actions based on human posture estimation in its patent document "Method, device and related equipment for recognizing fitness actions based on human posture estimation" (application number: 202211531164.8, publication number: CN 115880774A). The specific steps of this method are as follows: 1) Obtain a video stream, and extract human key point data based on the video stream; 2) Input the human key point data into a trained fitness action classification network for action recognition to obtain an action classification result, where the trained fitness action classification network is a stacked neural network of a multi-layer neural network MLP and a long short-term memory neural network LSTM; 3) Obtain a preset standard fitness action dictionary, and for each recognized action data, calculate the similarity between the recognized action data and the standard action data of the corresponding category in the standard fitness action dictionary to obtain a similarity value; 4) Based on the similarity value corresponding to each group of recognized action data, conduct a fitness action quality evaluation to obtain a quality evaluation result. The disadvantages of this method are as follows: Since the standard fitness action dictionary is generated by the set demonstrators during the evaluation, and the collected image data comes from the fitness person himself / herself, and the body proportions vary from person to person, this will impose certain limitations on the accuracy and rationality of the fitness action quality evaluation result.

[0004] Tianjin University disclosed a posture analysis method based on periodic length segmentation analysis in its patent document "A Posture Analysis and Evaluation Method Based on Periodic Length Segmentation Analysis" (Application No.: 202310486404.5, Publication No.: CN 116543458 A). The specific steps of this method are as follows: 1) Obtain the standard time series diagram and angle time series diagram of key joints in the motion video; 2) Obtain the teaching joint time series diagram and teaching angle time series diagram of key joints in the teaching video; Use the DTW algorithm to segment and align the standard time series diagram with the teaching joint time series diagram, and the angle time series diagram with the teaching angle time series diagram; 3) Calculate the joint coordinate offset value and joint angle offset value of key joints for each posture respectively, and determine the number of abnormal postures based on the joint angle offset value and joint coordinate offset value; 4) Obtain the joint angle offset value, joint coordinate offset value, and posture abnormal frames of all postures to calculate the overall accuracy of the action, the accuracy of the action coordinates, and the accuracy of the action angles. The deficiencies of this method are as follows: Since a preset teaching video is used for comparison during evaluation, this teaching video has a preset stable image shooting angle, while the collected motion video comes from the person to be analyzed. There may be slight jitter in the position of the shooting device or angle change, resulting in slight jitter or angle change in the image. At this time, comparing the motion video with the teaching video will cause a certain error in the evaluation result of the action quality of the person to be analyzed. Summary of the Invention

[0005] The object of the present invention is to address the deficiencies of the above-mentioned prior art and propose a squat posture analysis method based on multiple regression to solve the problems existing in the prior art that in the method of evaluating the squat posture through computer vision, there is jitter in the image, resulting in errors in the collected bone coordinate nodes and reducing the accuracy of the evaluation result; and the data collected only comes from the person to be analyzed, while the standard action subject during evaluation is a demonstrator set by the algorithm, and the body proportions of people vary, resulting in certain limitations on the accuracy and rationality of the evaluation result.

[0006] To achieve the above object, the technical idea of the present invention is: conduct a multicollinearity test on the coordinate of each bone node of the person to be analyzed. By calculating the variance inflation factor value VIF (Variance Inflation Factor), test the correlation between the coordinate of each bone node and the coordinates of other bone nodes during the squat posture, and eliminate the bone node coordinates that cause multicollinearity with VIF > 10. During this process, by gradually discarding irrelevant bone node coordinates, finally retain the key bone node coordinates. After eliminating the useless joint point information in the evaluation of the squat posture standard degree, the coefficient of determination R 2The present invention performs a backward stepwise regression analysis on the retained skeletal node coordinates to obtain a multiple linear regression model of the key skeletal node coordinates of each analyst, and calculates the coefficients of the multiple linear regression model using the least squares regression algorithm LS (Least Square). On the basis of the multiple linear regression model, the change threshold of the action standard degree L is found to evaluate the current tester's squatting posture. When the tester is changed, the multiple linear regression model can adaptively adjust the coefficient, which solves the problem that the body proportions of different people in the prior art are different, resulting in certain limitations on the accuracy and rationality of the evaluation results.

[0007] The steps implemented by the present invention include the following:

[0008] Step 1, obtaining a motion video of the person to be analyzed, identifying the human posture of the person to be analyzed in each frame of the motion video through the Mediapipe algorithm, generating human skeletal joint coordinates, and normalizing the horizontal and vertical coordinate data in the human skeletal joint coordinates respectively;

[0009] Step 2, perform a multicollinearity test on each bone joint coordinate after normalization, and calculate the variance inflation factor value VIF to test the correlation between each bone joint coordinate and other bone joint coordinates in the squat posture, and eliminate the bone joint coordinates with VIF greater than 10;

[0010] Step 3, perform backward stepwise regression analysis on the eliminated bone joint coordinates to obtain a multiple linear regression model of the tester's posture standard degree L with respect to each key bone joint coordinate, and calculate the final regression equation coefficient using the least squares regression algorithm LS; determine the range of change of the posture standard degree L using an adaptive threshold algorithm;

[0011] Step 4: determine the posture standard level L of each frame in the motion video of the person to be analyzed within a preset range as normal, and determine the rest as abnormal.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] First, since the present invention conducts a multicollinearity test on the coordinate of each bone node of the analyzer to be analyzed, and examines the correlation between the coordinate of each bone node and the coordinates of other bone nodes during the squatting motion by calculating the variance inflation factor value VIF, it can effectively eliminate the bone node coordinates that cause multicollinearity, overcoming the defect of inaccurate evaluation caused by the interference of useless joint point information in the traditional method. By improving the goodness of fit and recognition accuracy of the final regression equation, the present invention has the advantages of higher accuracy and reliability in squatting posture evaluation.

[0014] Second, since the present invention conducts a backward stepwise regression analysis on the remaining bone node coordinates and constructs a multiple linear regression model of the key bone node coordinates of each tester by using the least square regression algorithm LS (Least Square), it overcomes the deficiency that the traditional method cannot adaptively adjust the coefficients when replacing the tester, making the present invention have the advantages of strong adaptability and more accurate and reasonable evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flowchart of the present invention;

[0016] Figure 2 is a schematic diagram showing the change of the action standard degree L of the standard squatting posture of the present invention over time. DETAILED DESCRIPTION OF THE INVENTION

[0017] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings in the preferred embodiments of the present invention. In the drawings, the same or similar reference numerals represent the same or similar components or components with the same or similar functions throughout. The described embodiments are some but not all of the embodiments of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention and should not be construed as limiting the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention. The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Refer to Figure 1 for a further description of the implementation steps of the embodiments of the present invention.

[0020] Step 1: Obtain the motion video of the analyzer to be analyzed, identify the human posture of the analyzer to be analyzed in each frame of the motion video through the Mediapipe algorithm, generate the human bone joint coordinates, and perform normalization processing on the horizontal and vertical coordinate data in the human bone joint coordinates respectively.

[0021] The motion video of the person to be analyzed refers to a motion video obtained by taking the side view of the person to be analyzed indoors or outdoors and capturing the postures and the moving paths of the skeletal joints at each action stage of the squatting posture.

[0022] The steps of the Mediapipe algorithm are as follows:

[0023] In the first step, analyze the positions of each joint of the person to be analyzed in each frame of the motion video. Taking the lower left corner of the video image as the coordinate origin (0, 0), generate a plane coordinate system, and obtain the 33 skeletal joint coordinates (x, y) of the person to be analyzed. Here, x and y respectively represent the horizontal pixel distance and the vertical pixel distance from the skeletal joint coordinate to the lower left corner of the video image.

[0024] In the second step, mark the positions of the 33 skeletal joint coordinates of the person to be analyzed in each frame of the video image.

[0025] The steps for normalizing the human skeletal joint coordinates are as follows:

[0026] In the first step, find the maximum values x max and y max and the minimum values x min and y min of the horizontal and vertical coordinate values of each skeletal joint coordinate in all frames of the motion video;

[0027] In the second step, according to the formulas and normalize the horizontal and vertical coordinate values of each skeletal joint coordinate into the interval [-1, 1]. Here, represents the normalized coordinate value of the j-th skeletal joint coordinate.

[0028] In the embodiments of the present invention, a total of 8 persons to be analyzed are selected. Each person to be analyzed performs 10 groups of squat tests, and each group contains 5 squat actions, so as to obtain a data set containing 400 squat actions. Specifically, first, when the person to be analyzed is ready, the analyst clicks the start button to control the start and end of the collection of squat data. Among them, the collection of squat data is usually realized by recording a complete squat action video. In this video, it can be seen that the person to be analyzed performs multiple actions within a period of time, and these action data will subsequently be segmented to form multiple single-action data. Each single-action data is composed of a squat starting point to a squat ending point.

[0029] After the acquisition is completed, a segmentation method is used to detect the regularity of the body rising or falling with the change of the squat action in the complete squat action video, and the relatively stable characteristics of the key points of the eyes, thighs, calves, and buttocks during the rising and falling process. The segmentation method refers to identifying and extracting each individual squat action in the video, and segmenting it from the entire video for separate analysis. The start and end time of each individual squat action is recorded, and multiple frame segments of a single squat action are extracted to obtain the coordinates of the key points of each frame, and the coordinates of the key points of each frame are normalized.

[0030] Step 2: Perform a multicollinearity test on each bone joint coordinate after normalization. By calculating the variance inflation factor value VIF, test the correlation between each bone joint coordinate and other bone joint coordinates in the squat posture, and eliminate the bone joint coordinates with VIF greater than 10.

[0031] The steps of the multicollinearity test are as follows:

[0032] The first step is to follow Formula, calculate the variance inflation factor VIF of each bone joint coordinate, where, It indicates the goodness of fit of the transverse and ordinate values of the mth bone joint obtained through regression analysis to the transverse and ordinate values of other bone joints;

[0033] In the second step, all the horizontal and vertical coordinate values of the skeletal joint coordinates with VIF>10 are judged as weakly correlated, and the rest are retained as key skeletal joint coordinates.

[0034] In step 2 of the embodiment of the present invention, first the key point coordinates of each frame obtained after the normalization process in step 1 are subjected to a multicollinearity test. This process is achieved by calculating the variance inflation factor VIF (Variance Inflation Factor) of each skeletal joint coordinate, so as to analyze the correlation between each skeletal joint coordinate and other coordinates when in a squatting posture. In order to ensure that each coordinate provides independent and useful information to analyze the squatting posture, the multicollinearity problem caused by the highly correlated coordinates is avoided, so that subsequent regression analysis is more accurate and reliable, the skeletal joint coordinates with VIF (Variance Inflation Factor) greater than 10 will be eliminated, and a more concise and more targeted key skeletal coordinate is obtained, which lays the foundation for the next stepwise regression analysis backward.

[0035] Step 3: Perform backward stepwise regression analysis on the eliminated bone joint coordinates to obtain the multivariate linear regression model of the test subject's posture standard degree L with respect to each key bone joint coordinate, and use the least squares regression algorithm LS to calculate the final regression equation coefficients; determine the range of change of the posture standard degree L through an adaptive threshold algorithm.

[0036] The steps of the backward stepwise regression analysis are as follows:

[0037] The first step is to establish a multivariate linear regression equation for the posture standard degree L with respect to the coordinates of all key bone joints through the least squares method.

[0038] The second step is based on Formula to calculate the residual sum of squares SSE, where x i ,y i Represents the horizontal and vertical coordinate values of the i-th skeletal joint motion video, Represents the value after fitting the multiple linear regression equation.

[0039] In the third step, each bone node coordinate is removed from the multiple linear regression equation in turn to obtain k multiple linear regression equations containing k-1 key bone joint coordinates, where k represents the total number of all key bone joint coordinates.

[0040] Step 4: Calculate the residual sum of squares SSE for the k equations in step 3 i , and according to the formula ΔSSE i =|SSE-SSE i |Calculate the absolute error of the residual sum of squares ΔSSE i , remove the key skeletal joint coordinates corresponding to the multivariate linear regression equation with the smallest absolute error.

[0041] Step 5: Repeat steps 3 and 4 until removing any bone node coordinate from the multivariate linear regression equation does not make the absolute error greater than 0.05. At this time, a final regression equation containing kn bone node coordinates is obtained, where n is the bone node coordinate removed during the iteration process.

[0042] The sixth step is to use the least squares regression algorithm LS to obtain the final regression equation of the posture standardization degree L with respect to the coordinates of the key bone joints.

[0043] The preset range of the posture standard degree L determined by the adaptive threshold algorithm refers to the posture standard degrees of all frames calculated by traversing the final regression equation, and selecting the maximum value L of the posture standard degree in all frames. max With the minimum value L min , [0.9L max ,1.1L max∪ [0.9L min , 1.1L min interval is used as the preset range of the posture standard degree L.

[0044] In step 3 of the embodiment of the present invention, backward stepwise regression analysis is started on the bone joint coordinates after elimination. First, a multiple linear regression model including all key coordinates is constructed. Then, each coordinate is removed one by one, and a new multiple linear regression model is constructed for the coordinate set after each removal. After each coordinate is removed, the sum of squared residuals of the new model is calculated and compared with the sum of squared residuals of the original model to calculate the absolute error. The coordinate with the smallest absolute error is selected for removal. This process continues until an optimal model is found, which can best describe the standard degree of the squat by the smallest sum of squared residuals. Finally, the adaptive threshold algorithm is used to determine the change range of the posture standard degree L, providing a basis for the next squat posture judgment. A suitable threshold range is determined through multiple experiments to judge whether the squat posture of the person to be analyzed is standard.

[0045] The expression of the regression equation is as follows:

[0046]

[0047] where y nose represents the ordinate of the tip of the nose; represents the ordinate of the inner side of the right eye; represents the ordinate of the right eye; represents the ordinate of the outer side of the right eye; represents the ordinate of the inner side of the left eye; represents the ordinate of the left eye; represents the ordinate of the outer side of the left eye; represents the ordinate of the right ear; represents the ordinate of the left ear; represents the ordinate of the right side of the mouth; represents the ordinate of the left side of the mouth; represents the ordinate of the right shoulder; represents the ordinate of the left shoulder; represents the ordinate of the right elbow; represents the ordinate of the left elbow; represents the ordinate of the right wrist; represents the ordinate of the left wrist; represents the ordinate of the right little finger; represents the ordinate of the left little finger; represents the ordinate of the right index finger; represents the ordinate of the left index finger; represents the ordinate of the right thumb; Represents the vertical coordinate of the left thumb; Represents the vertical coordinate of the right hip; Represents the vertical coordinate of the left hip; Represents the vertical coordinate of the right ankle; Represents the vertical coordinate of the left ankle; Represents the vertical coordinate of the right heel; Represents the vertical coordinate of the left heel; Represents the vertical coordinate of the right toe; Represents the vertical coordinate of the left toe. Similarly, for the horizontal coordinates of each bone joint, simply replace y with x.

[0048] Step 4: Determine that the posture standard degree L of each frame in the movement video of the person to be analyzed within the preset range is normal, and the rest are abnormal.

[0049] In the embodiments of the present invention, the obtained multiple linear regression model and the change range of the posture standard degree L are used to judge the squat posture of each frame. Specifically, substitute the coordinate values of each normalized bone joint of each frame into the multiple linear regression model to calculate the posture standard degree L of each frame. Judge whether the posture of each frame is standard according to the preset change range of the posture standard degree L. In this way, it can be clearly known which parts are standard and which parts are abnormal in the squat movement, so as to more targeted guide the person to be analyzed for improvement.

[0050] Refer to Figure 2 , and make a further description of the change process of the posture standard degree L of the person to be analyzed in the embodiments of the present invention over time.

[0051] For the posture standard degree L data of the person to be analyzed in the embodiments of the present invention, draw a curve graph with time as the abscissa and the posture standard degree L as the ordinate. The peaks of the posture standard degree L of the first four standard squats are concentrated at 1.15. Therefore, it can be judged that when the posture standard degree L = 1.5, the person to be analyzed is at the end of the squat, and when the posture standard degree L < 0.2, the person to be analyzed is in the state of returning to the upright position from the squat. Accordingly, the squat video of the person to be analyzed can be segmented into single-squat videos with the posture standard degree L = 0.2 as the key frame for one-by-one inspection. Traverse the posture standard degree of all frames calculated by the final regression equation, and select the maximum value L max and the minimum value L min of the posture standard degree in all frames. Take the interval of [0.9L max , 1.1L max ∪ [0.9L min , 1.1L min as the preset range of the posture standard degree L. If the posture standard degree of the person to be analyzed in the embodiment is within this preset range, the squat posture of the person to be analyzed is standard, otherwise it is abnormal.

[0052] The effects of the present invention will be further described below in conjunction with simulation experiments:

[0053] 1. Simulation experiment conditions:

[0054] The hardware platform for the simulation experiment of the present invention is as follows: the processor is AMD 5800H, the memory is 16G, and the main frequency is 4.20GHz.

[0055] The software platform for the simulation experiment of the present invention is: Windows 10 operating system and python 3.8.

[0056] The test set used in the simulation experiment of the present invention consists of 1000 videos. The videos are the squat videos of 495 different male test subjects and 505 different female test subjects collected by a mobile phone camera. Each test subject corresponds to one video, and there is exactly one squat action in one video. The size of each video is 640*480, and the video file format is.mp4. The human squat actions in the videos are divided into two categories, namely standard and abnormal.

[0057] 2. Simulation content and its result analysis:

[0058] The simulation experiment of the present invention adopts the method of the present invention. After all steps of the present invention for all videos in the input video test set, the standard degree of all squat actions in each sample of the 1000 samples in the test set is estimated. The results are shown in Table 1.

[0059] Table 1: List of recognition results of all squat actions in the test set

[0060] Number of samples Standard squat motion Abnormal squat motion Total Male 231 264 495 Female 259 246 505 Total 490 510 1000

[0061] In order to verify the effect of the simulation experiment of the present invention, the following two evaluation criteria are used to evaluate the correct detection degree of bone nodes and the recognition accuracy of squat actions respectively.

[0062] The first evaluation criterion is the percentage of the number of correctly located bone nodes of the person to be analyzed in the total number of bone nodes (Percentage of Detected Joints, PDJ). Its specific meaning is to measure the Euclidean distance between the predicted bone nodes and the true bone nodes. If it is less than 0.1 times the trunk diameter, then this bone node is regarded as correctly located. After accumulating all correctly located bone nodes and dividing by the total number of all bone nodes, the PDJ value is obtained.

[0063] The second evaluation criterion is that when the person to be analyzed squats down, the thighs are parallel to the ground, the toes are slightly turned outwards by no more than 30 degrees, and the back is kept straight and the pelvis is kept in a neutral position during the deep squat. The standard degree of all squat movements in each video is evaluated according to this evaluation criterion, and the coincidence degree between the estimation result of the present invention and this evaluation result is used as the recognition accuracy rate. Tables 2 and 3 respectively give whether all bone nodes are correctly positioned and the recognition accuracy rate obtained in the simulation experiment of the present invention.

[0064] Table 2: List of whether all bone nodes are correctly positioned

[0065]

[0066] Table 3: List of recognition accuracy rates

[0067]

[0068] As can be seen from Table 2, the human posture estimated by the present invention reaches 97% in terms of the PDJ value. As can be seen from Table 3, the present invention reaches 98% in terms of the standard degree of the squat movement, proving that the present invention can effectively identify and evaluate the squat movement.

Claims

1. A squat posture analysis method based on multiple regression, characterized in that, The cleaned bone joint coordinates are subjected to multicollinearity test and screening, and the removed bone joint coordinates are subjected to backward stepwise regression analysis; the analysis method comprises the following steps: Step 1, obtaining a motion video of the person to be analyzed, identifying the human posture of the person to be analyzed in each frame of the motion video through the Mediapipe algorithm, generating human skeletal joint coordinates, and normalizing the horizontal and vertical coordinate data in the human skeletal joint coordinates respectively; Step 2, perform a multicollinearity test on each bone joint coordinate after normalization, and calculate the variance inflation factor value VIF to test the correlation between each bone joint coordinate and other bone joint coordinates in the squat posture, and eliminate the bone joint coordinates with VIF greater than 10; Step 3, perform backward stepwise regression analysis on the eliminated bone joint coordinates to obtain a multiple linear regression model of the tester's posture standard degree L with respect to each key bone joint coordinate, and calculate the final regression equation coefficient using the least squares regression algorithm LS; determine the range of change of the posture standard degree L using an adaptive threshold algorithm; Step 4: determine the posture standard level L of each frame in the motion video of the person to be analyzed within a preset range as normal, and determine the rest as abnormal.

2. The squat posture analysis method based on multiple regression according to claim 1, wherein The motion video of the person to be analyzed in step 1 refers to a motion video obtained by shooting the posture and movement path of the bone nodes of each action stage of the squat posture of the person to be analyzed from a side perspective indoors or outdoors.

3. The method for analyzing squat postures based on multiple regression according to claim 1, wherein The steps of the Mediapipe algorithm described in step 1 are as follows: The first step is to analyze the position of each joint of the person to be analyzed in each frame of the motion video, and generate a plane coordinate system with the lower left corner of the video image as the coordinate origin (0,0), and obtain the coordinates of 33 skeletal joints of the person to be analyzed (x, y), where x and y represent the horizontal pixel distance and vertical pixel distance from the skeletal joint coordinates to the lower left corner of the video image respectively; The second step is to mark the coordinate positions of the 33 skeletal joints of the person to be analyzed in each frame of the video image.

4. The squat posture analysis method based on multiple regression according to claim 1, wherein The steps for normalizing the human skeleton joint coordinates described in step 1 are as follows: First step, obtain the maximum values x max and y max of the horizontal and vertical coordinates of each skeletal joint coordinate in all frames of the motion video, min as well as the minimum values x min and y; Step 2: According to the formulas and normalize the horizontal and vertical coordinate values of each bone joint coordinate to the interval [-1, 1], where represents the normalized coordinate value of the j-th bone joint coordinate.

5. The squat posture analysis method based on multiple regression according to claim 1, wherein The steps of the multicollinearity test described in step 2 are as follows: First step, according to formula, calculate the variance inflation factor VIF of each skeletal joint coordinate, where represents the goodness of fit of the horizontal and vertical coordinate values of the m-th skeletal joint obtained through regression analysis to the horizontal and vertical coordinate values of other skeletal joints; In the second step, the skeletal joint coordinates with variance inflation factors less than 10 are determined as key skeletal joint coordinates, and the rest are eliminated.

6. The squatted posture analysis method based on multiple regression according to claim 4, characterized in that The steps for the backward stepwise regression analysis described in Step 3 are as follows: The first step is to establish a multivariate linear regression equation of the posture standard degree L with respect to the coordinates of all key bone joints through the least squares method; Step 2: According to the formula, calculate the sum of squared residuals SSE, where x i , y i represent the horizontal and vertical coordinate values in the i-th skeletal joint movement video, represents the value after fitting by the multiple linear regression equation; The third step is to remove each bone node coordinate from the multiple linear regression equation in turn to obtain k multiple linear regression equations containing k-1 key bone joint coordinates, where k represents the total number of all key bone joint coordinates; Step 4: Calculate the sum of squared residuals SSE for the k equations in Step 3 i , and calculate the absolute error ΔSSE of the sum of squared residuals according to the formula ΔSSE i = |SSE - SSE i |; remove the key skeletal joint coordinates corresponding to the multiple linear regression equation with the smallest absolute error i ; Step 5: Repeat steps 3 and 4 until removing any bone node coordinate from the multiple linear regression equation does not make the absolute error greater than 0.

05. At this point, a final regression equation containing kn bone node coordinates is obtained, where n is the bone node coordinate removed during the iteration process. Step 6: Using the least squares regression algorithm LS, obtain the final regression equation of the attitude standard degree L with respect to the coordinates of the key skeletal joints.

7. The method for analyzing squat postures based on multiple regression according to claim 1, wherein The preset range of the attitude standard degree L determined by the adaptive threshold algorithm in step 3 refers to traversing the attitude standard degrees of all frames calculated by the final regression equation, and selecting the maximum value L of the attitude standard degree among all frames max and the minimum value L min , and taking the interval of [0.9L max , 1.1L max ∪ [0.9L min , 1.1L min as the preset range of the attitude standard degree L.

Images