Multi-person deep squat real-time counting method based on human body key point measurement

Abstract

The invention discloses a multi-person deep squat real-time counting method based on human body key point measurement, and the method can improve the detection precision and robustness, supports the simultaneous detection of multiple persons, and is suitable for multiple scenes. The method comprises the following steps: based on a multi-target real-time detection and tracking technology of a CSTrack algorithm, identifying and tracking a plurality of detected persons at the same time, and recording and analyzing the motion trail and position of each person; based on a BlazePose algorithm, carrying out attitude estimation on a plurality of detected persons, detecting and tracking the human body attitude data of each detected person in real time, and measuring the distance and angle change between joint points to judge the start and end of the action; kNN similarity discrimination and CNN depth feature learning are combined to analyze and optimize the posture of the human body in real time, whether a detected person performs a deep squat action or not is judged, the deep squat category is determined, and then deep squat counting and action standard scoring are performed.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of image data processing technology, and in particular to a real-time counting method for multiple squats based on human body key point measurement. Background Art

[0002] Squat is a classic fitness movement, and real-time counting of squats is very important for physical exercise. When performing squat detection and counting, manual counting is prone to misjudgment and omission, and it is difficult to cope with the demand for real-time counting of multiple squats. With the advancement of artificial intelligence technology, the intelligence level of schools, sports facilities, fitness venues and various sports facilities has been continuously improved. However, most of the existing technologies focus on single-person squat counting. When there are many people being measured, there are problems such as time-consuming counting, low efficiency, and non-objective scoring. In addition, there is a problem of mutual occlusion in the scene of multi-person squats, which may cause the movements of some of the measured persons to be unable to be accurately identified, affecting the accuracy of counting. In view of the fact that various variants of squats (such as wide squats, narrow squats, freehand squats, kneeling squats, wall squats, kettlebell squats, unbalanced ground squats, split squats, squat jumps, etc.) may cause the recognition system to be difficult to accurately count, this application proposes a real-time counting method for multi-person squats based on human key point measurement. The method designed in this patent helps to promote the intelligent process of sports facilities and can effectively save manpower and material investment in the physical measurement process. Summary of the invention

[0003] In order to overcome the above technical defects, the embodiment of the present application provides a real-time counting method for multiple squats based on human key point measurement, and its specific scheme is as follows:

[0004] S1. Subject detection based on MTCNN algorithm:

[0005] After acquiring the video, the MTCNN algorithm is used to perform face recognition on the subject to determine whether the subject's identity information has changed. If so, the subject's squat count is reset to zero, the squat count of the subject is ended and the data is stored in the background; if there is no change, the real-time squat count continues;

[0006] S2. Real-time tracking of subjects based on CSTrack algorithm:

[0007] After completing the face recognition verification and confirming that the subject's identity information has not changed, the CSTrack algorithm is used to identify and track the subject's movement trajectory and position, and assign a unique identity tag, or ID, to the subject;

[0008] S3. Human posture estimation based on BlazePose algorithm:

[0009] Based on the BlazePose algorithm, each human body detection frame in the environment is detected from top to bottom, and then the posture of each human body is detected independently. The human body posture is described by the three-dimensional coordinates of 33 joint points in each frame to obtain the human body posture data;

[0010] S4. Determination of squat start and end time based on human body key point measurement:

[0011] The bending angle of the knee of the subject is calculated from the human body posture data. When the bending angle of the knee is between 35° and 65°, the action is determined to start and the squat action detection is started. When the bending angle of the knee is between 95° and 105°, the action is determined to end and the squat action detection is ended.

[0012] S5. Squat category discrimination based on KNN algorithm model:

[0013] The KNN algorithm is used to identify the squat category to which the subject's current posture belongs by calculating the maximum distance screening and average distance screening of the left and right hips (23, 24) and the left and right knees (25, 26) and the left and right hips (23, 24) and the left and right ankles (27, 28);

[0014] S6. Count and score the squat movement standardization using the CNN algorithm model:

[0015] Through the CNN algorithm model, the current average distance value is compared with the reference average distance to evaluate the similarity between the squat action and the standard squat action, and the action standardization score is obtained.

[0016] To be more specific, the 33 joints are: nose, inside the left eye, inside the right eye, left eye, right eye, outside the left eye, outside the eye, left ear, right ear, left mouth, right mouth, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left little finger, right little finger, left index finger, right index finger, left thumb, right thumb, left hip, right hip, left knee, right knee, left ankle, right ankle, left heel, right heel, left toe and right toe.

[0017] To be more specific, the video is obtained by simultaneously capturing the subject from different perspectives using multiple cameras, and the captured video images are preprocessed before face recognition, and the preprocessing includes denoising, lighting correction and deblurring.

[0018] The embodiments of the present application adopt the above technical solution to achieve the following technical effects:

[0019] High precision and robustness: KNN and CNN algorithms are used for screening, combined with the distance changes of human joints, to accurately judge the action category, reduce misjudgment, improve counting stability and accuracy, and have strong anti-interference ability.

[0020] Supports multi-person squat counting: It solves the problem of time-consuming and low-precision manual counting. It is suitable for group training environments and can monitor and count multiple types of squat movements of multiple subjects in real time. The output squat movement standardization score is highly feasible.

[0021] Strong adaptability: It integrates multi-view data, maintains high recognition rate in complex environments, is applicable to a variety of scenarios, and facilitates squat counting and movement standardization scoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings exemplarily illustrate the embodiments and constitute a part of the specification, and together with the text description of the specification, are used to explain the exemplary implementation of the embodiments. The embodiments shown are for illustrative purposes only and do not limit the scope of the claims. In all drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0023] Figure 1 A flowchart of a method for real-time counting of multiple squats based on key point measurement of the human body in this application;

[0024] Figure 2 Schematic diagram of the coordinate positions of the 33 joint points used in this application;

[0025] Figure 3 This is a schematic diagram of the angle between the knee and the vertical direction in this application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0027] It should be noted that the technical solutions between the various embodiments of the present application can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0028] In the description of the present application, it should be understood that the numerical labels before the steps do not indicate the order in which the steps are executed, but are only used to facilitate the description of the present application and to distinguish each step, and therefore should not be understood as a limitation on the present application.

[0029] First, the following terms are explained:

[0030] MTCNN algorithm: The full name of MTCNN is Multi-Task Cascaded Convolutional Networks. It is a deep learning algorithm for face detection and face alignment. It is widely used in face detection, facial expression recognition, facial expression change analysis and other fields. The core idea is to perform face detection and key point positioning (key points such as eyes, nose, mouth, etc.) tasks simultaneously, and gradually refine the detection results through cascading, so as to achieve efficient multi-task learning.

[0031] CSTrack algorithm: is a single-stage multi-target tracking algorithm based on ReID (Re-Identification). The main goal of this algorithm is to track multiple targets in real time in applications such as video surveillance and autonomous driving, and to re-identify the same target between different video frames through ReID technology, thereby solving problems such as target occlusion, discreteness, and cross-camera in traditional multi-target tracking methods.

[0032] BlazePose algorithm: is a real-time human posture tracking algorithm proposed by Google, which aims to accurately and quickly identify key points of the human body, especially for real-time human posture estimation on mobile devices. It provides an efficient solution to estimate the position of key points of the human body (such as head, arms, legs, etc.) and provide 2D or 3D position information of each part of the body.

[0033] KNN algorithm: The full name is K-Nearest Neighbors. It is a supervised machine learning algorithm mainly used for classification and regression tasks. The core idea is to find the K samples that are most similar (that is, the closest) to the sample in the training data set given a sample to be predicted, and then predict the category (classification problem) or value (regression problem) of the sample to be predicted through the categories (or values) of these K neighbors.

[0034] CNN algorithm: It is a deep learning model suitable for image processing. It extracts image features through convolution operations, reduces the amount of data through pooling operations, and finally performs classification or recognition through a fully connected layer.

[0035] Secondly, in order to facilitate those skilled in the art to understand the technical solutions provided in the embodiments of the present application, the relevant technologies are described below:

[0036] See also Figure 1 In the embodiment of the present application, the real-time counting method of multiple squats based on human key point measurement includes three parts: recognition tracking, posture estimation and counting scoring, which are as follows:

[0037] 1. Identification and tracking

[0038] S1. Subject detection based on MTCNN algorithm

[0039] After acquiring the video, the MTCNN algorithm is used to perform face recognition on the subject to determine whether the subject's identity information has changed. If so, the subject's squat count is cleared, the squat count of the subject is ended and the data is stored in the background; if there is no change, the real-time squat count continues.

[0040] Furthermore, the video is obtained by simultaneously capturing the subject from different perspectives using multiple cameras, and the captured video images may be pre-processed before face recognition. The pre-processing may specifically include denoising, lighting correction, deblurring, etc.

[0041] S2. Real-time tracking of subjects based on CSTrack algorithm

[0042] After completing the face recognition verification and confirming that the subject's identity information has not changed, the CSTrack algorithm is used to identify and track the subject's movement trajectory and position, and assign a unique identity tag, or ID, to the subject. The uniqueness of the identity tag can achieve continuous tracking of the subject throughout the squat process. Subsequent squat count data is stored in the background based on the ID.

[0043] 2. Pose Estimation

[0044] S3. Human pose estimation based on BlazePose algorithm

[0045] Based on the BlazePose algorithm, each human body detection frame in the environment is detected from top to bottom, and then the posture of each human body is detected independently. The human body posture is described by the three-dimensional coordinates of 33 joint points in each frame to obtain the human body posture data.

[0046] Specifically, the 33 joints are: nose, left eye inside, right eye inside, left eye, right eye, left eye outside, eye outside, left ear, right ear, left mouth, right mouth, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left little finger, right little finger, left index finger, right index finger, left thumb, right thumb, left hip, right hip, left knee, right knee, left ankle, right ankle, left heel, right heel, left toe and right toe. Their specific locations are as follows: Figure 2 The positions indicated by numbers 0-32.

[0047] Squat counting requires extremely high real-time performance. Traditional human feature extraction based on RGB features is greatly affected by the environment and computing equipment, and has poor real-time performance, which is a challenge for squat recognition and counting systems. In contrast, the human posture feature extraction system based on the BlazePose algorithm can effectively capture the 33 key joint positions of multiple human bodies in space during squats, and independently detect the posture of each human body and record relevant data through the three-dimensional coordinates of the 33 key joint positions in each frame.

[0048] S4. Determination of the start and end time of squat based on the measurement of key points of the human body

[0049] The bending angle of the subject's knee is calculated from the human posture data. When the bending angle of the knee is between 35° and 65°, the action is determined to have started and the squat action detection begins. When the bending angle of the knee is between 95° and 105°, the action is determined to have ended and the squat action detection ends.

[0050] Between 35° and 65° means greater than or equal to 35° and less than or equal to 65°. Similarly, between 95° and 105° means greater than or equal to 95° and less than or equal to 105°.

[0051] More specifically, Figure 3 As shown, the angle between the knee-hip and the knee-ankle is selected as the bending angle of the knee with the knee as the center point. The angle value is calculated as follows:

[0052] S4.1. Calculate the vectors from knee to hip and knee to ankle using formula (1)

[0053] Among them, (x k ,y k ,z k )、(x h ,y h ,z h ) are the coordinates of the knee and hip respectively.

[0054] S4.2. Calculate the vector using formula (2) The vector perpendicular to the The angle between them, cos(θ), gives the bending angle of the knee:

[0055]

[0056] in, It should be understood that since the vertical vector is in the z-axis direction, the vertical vector

[0057] 3. Counting and Scoring

[0058] Before combining the KNN algorithm and CNN algorithm model to analyze and optimize the human posture in real time and realize squat counting and scoring, a multi-perspective fusion squat action video training set is collected in advance as sample data for model training of the KNN algorithm and CNN algorithm model. The sample data includes the human posture features in 19 different perspectives and different types of squat action videos extracted by the BlazePose algorithm, which are converted into objects containing joint point coordinates and squat posture samples. These posture features include the position and motion trajectory of each joint point of the human body, which can fully capture the details of the squat action. Then, these feature data are used as sample sets to train the KNN and CNN algorithm models, learn and identify standard squat action postures, and provide standards for the recognition, counting and similarity evaluation of actions in subsequent steps.

[0059] S5. Squat category discrimination based on KNN algorithm model

[0060] The KNN algorithm is used to identify the squat category to which the subject's current posture belongs by calculating the maximum distance screening and average distance screening of the left and right hips (23, 24) and the left and right knees (25, 26) and the left and right hips (23, 24) and the left and right ankles (27, 28).

[0061] Specifically, recognition is to determine whether it meets the KNN classification labeling, identify the squat sample that is most similar to the current posture, and perform a similarity evaluation to determine the squat category to which it belongs.

[0062] The above-mentioned maximum distance screening and average distance screening are used to identify the squat category to which the current posture of the subject belongs, specifically including:

[0063] First, for each pair of joints (ankle and hip, hip and knee), the Euclidean distance between them is calculated, and the calculation formula is shown in formula (3):

[0064]

[0065] Among them, k represents the kth person, is the distance between joint points i and j of the kth person, and are the coordinates of joint i and joint j of the kth person respectively.

[0066] The distances between the joint pairs that need to be calculated in this application include: the distances between the left hip (23) and the left knee (25), the right hip (24) and the right knee (26), the left hip (23) and the left ankle (27), and the right hip (24) and the right ankle (28).

[0067] Secondly, the distance between the joint pairs calculated using formula (4) is normalized to obtain the normalized distance:

[0068]

[0069] in, is the normalized distance, and is the minimum and maximum distances between joints measured by the kth person.

[0070] Again, the set of all normalized distances of the subjects is screened for the maximum distance, and samples that are highly similar to the target squat sample but have a small difference in the feature vector are eliminated to obtain the screened maximum distance set.

[0071] Finally, for the filtered maximum distance set, calculate its average distance value.

[0072] Specifically, let D (k) is the set of all normalized distances of the kth person, by setting a threshold T max , filter out the maximum distance set D max (k) , the specific formula is shown in formula (5):

[0073]

[0074] For the maximum distance set after screening, use formula (6) to calculate its average distance value:

[0075]

[0076] Finally, the squat category to which the subject's current posture belongs is determined according to the calculated average distance value.

[0077] S6. Count and score the squat movement standardization through the CNN algorithm model

[0078] Through the CNN algorithm model, the current average distance value is compared with the reference average distance to evaluate the similarity between the squat action and the standard squat action, and a similarity score (i.e., action standardization score) is obtained.

[0079] Specifically, the similarity calculation formula (7) is as follows:

[0080]

[0081] Among them, S (k) Score the standardness of the k-th person’s actions, μ ref is the average distance of the squat movement obtained from the sample training.

[0082] Based on the KNN algorithm, the squat category that is most similar to the current posture is identified. Then, the CNN algorithm model is used to evaluate the similarity of the squat movement features obtained by training the squat sample of this type, so as to obtain the movement standardization score, which helps the subjects to correct non-standard movements in time. In summary, through the algorithm automation and standardization of the measurement process, human errors are reduced, the accuracy and reliability of multi-person squat measurements are significantly improved, and the subjects are helped to optimize their movements in time.

[0083] In the squat counting, by analyzing the distance between the key points of the human body, the start and end of the action can be accurately judged, and the standard of the action can be evaluated. There are many types of squats (such as wide squats, narrow squats, etc.), but the common feature of squats is that they all require the hips and knees to be parallel or lower than the knees when squatting, and the distance between the hip and ankle joints is small at this time. This patent design starts the action signal acquisition part, first checking whether the coordinates of the human body joint points are within the video frame, and then calculating the angle between the human body knee and the vertical direction (such as Figure 3 The angle between the vector formed by the knee and hip and the z-axis) is used to determine the start and end of the action. If the action starts, the angle between the human hip and ankle joints ( Figure 2 Key points 23, 24, 27, 28) and the human hip and knee joints ( Figure 2 The distance between the key points 23, 24, 25, 26) is used for similarity judgment and standardization scoring. In addition, if the angle between the knee and the vertical direction is between 35° and 65°, the signal is "start", and if the angle is between 95° and 105°, it is "end". The above dual recognition greatly improves the accuracy of squat recognition.

[0084] The squat counting module based on edge devices is designed to achieve accurate action recognition and counting by real-time analysis of the user's squat movements. The squat counting system built through the above steps can be directly deployed on various edge devices. The movement process is monitored in real time by the camera. As can be seen from the above, the counting of squats can be achieved by detecting the angle changes at key moments. When the angle between the knee and the vertical direction gradually exceeds 95° from less than 65°, it is counted as a squat and the corresponding score is output in real time. After the exercise is completed, the squat counting result and the action standardization score of each ID are output. In this way, this application can be applied to scenes such as gyms, training grounds, and school physical tests to achieve real-time counting of multiple squats.

[0085] In the technical solution of this application, the following technical effects can be achieved:

[0086] First, improve the accuracy and robustness of squat recognition.

[0087] By applying the maximum distance and average distance screening of KNN and CNN algorithms, combined with the distance changes between human joints during movement, a squat start and end time recognition, counting and action standardization scoring system is constructed, achieving high-precision action start and end judgment and action standardization scoring. Specifically, during the squat process, the method proposed in this patent accurately judges the start and end of each action by analyzing the angle change between the human knee and the vertical direction, and analyzes the distance between the key points of the human hip and ankle, hip and knee to perform similarity evaluation, thereby ensuring the accuracy of counting and action standardization scoring. Maximum distance screening helps to exclude samples that are very similar to the target sample but have a small difference in the feature vector, while average distance screening ensures that the overall posture is consistent with the standard sample. This method effectively reduces the misjudgment caused by prediction jitter and improves the stability of counting. The system combines the human posture characteristics with different perspective data obtained by multiple cameras to reduce occlusion problems and provide more accurate skeletal key point positioning, thereby improving the accuracy of squat action recognition. It has stronger anti-interference ability against interference factors such as light changes and complex backgrounds, thereby maintaining high accuracy and stable performance in various environments. In addition, CSTrack-based multi-target real-time detection and tracking technology can provide comprehensive support for training and competition in sports scenarios, ensuring that the movement trajectory of each subject is accurately recorded, while also providing data basis for the subject's skill improvement. Through this efficient monitoring and analysis, sports teams can gain greater advantages in competition and promote the continuous improvement of sports levels.

[0088] Second, it can simultaneously perform real-time detection and tracking on multiple persons being tested to achieve multi-person support function.

[0089] When conducting physical examinations, it is generally completed by on-site counting by the test teacher. Manual counting is prone to misjudgment and omission. When there are many subjects, it is easy to cause problems such as low counting efficiency and time-consuming. The system proposed in this patent can simultaneously identify the squat movements of multiple subjects and count and score them, which can improve the counting efficiency in group training scenarios. Through advanced deep learning algorithms, the movement posture and number of squats of each subject can be monitored in real time to ensure that each subject can perform correct training. The system's real-time counting and scoring feedback functions can provide each subject with timely posture correction or adjustment of training intensity. In addition, the system can also obtain overall data, analyze the performance and progress of each subject, and design a more efficient training plan. In addition, the multi-person support function can also enhance the interactivity and competitiveness of training.

[0090] Third, it can be applied to a variety of scenarios and has strong adaptability.

[0091] Fusion of data from different perspectives can maintain high recognition rates in various complex environments and adapt to different training scenarios. This design not only improves the naturalness and convenience of the user experience, but also makes the system suitable for more usage scenarios, whether at home, in the gym, or outdoors, you can easily count and evaluate squats.

[0092] Obviously, those skilled in the art should understand that the modules or steps of the above-mentioned embodiments of the present application can be implemented by general-purpose computer devices, they can be concentrated on a single computer device, or distributed on a network composed of multiple computer devices, optionally, they can be implemented by executable program codes of computer devices, so that they can be stored in a storage device and executed by the computer device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0093] It should be noted that the above are only preferred embodiments of the present application, and the patent protection scope of the present application is not limited thereto. Any equivalent structure or equivalent process transformation made using the contents of the specification and drawings of the present application, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A real-time counting method for multi-person squatting based on human key point measurement, characterized in that: include: S1. Subject detection based on MTCNN algorithm: After acquiring the video, the MTCNN algorithm is used to perform face recognition on the subject to determine whether the subject's identity information has changed. If so, the subject's squat count is reset to zero, the squat count of the subject is ended and the data is stored in the background; if there is no change, the real-time squat count continues; S2. Real-time tracking of subjects based on CSTrack algorithm: After completing the face recognition verification and confirming that the subject's identity information has not changed, the CSTrack algorithm is used to identify and track the subject's movement trajectory and position, and assign a unique identity tag, or ID, to the subject; S3. Human posture estimation based on BlazePose algorithm: Based on the BlazePose algorithm, each human body detection frame in the environment is detected from top to bottom, and then the posture of each human body is detected independently. The human body posture is described by the three-dimensional coordinates of 33 joint points in each frame to obtain the human body posture data; S4. Determination of squat start and end time based on human body key point measurement: The bending angle of the knee of the subject is calculated from the human body posture data. When the bending angle of the knee is between 35° and 65°, the action is determined to start and the squat action detection is started. When the bending angle of the knee is between 95° and 105°, the action is determined to end and the squat action detection is ended. S5. Squat category discrimination based on KNN algorithm model: The KNN algorithm is used to identify the squat category to which the subject's current posture belongs by calculating the maximum distance screening and average distance screening of the left and right hips (23, 24) and the left and right knees (25, 26) and the left and right hips (23, 24) and the left and right ankles (27, 28); S6. Count and score the squat movement standardization using the CNN algorithm model: Through the CNN algorithm model, the current average distance value is compared with the reference average distance to evaluate the similarity between the squat action and the standard squat action, and the action standardization score is obtained.

2. The method for real-time counting of multiple squats according to claim 1, characterized in that: In S4, the angle between the knee-hip and the knee-ankle distances is selected as the bending angle of the knee, and the angle value is calculated as follows: S4.

1. Calculate the vectors from knee to hip and knee to ankle using the following formula (1): S4.2, use the following formula (2) to calculate the vector The vector perpendicular to the The angle between them, cos(θ), gives the bending angle of the knee:

3. The method for real-time counting of multiple squats according to claim 1, characterized in that: In S5, the squat category to which the subject's current posture belongs is identified based on the maximum distance screening and the average distance screening, specifically including: First, for each pair of joints, the Euclidean distance between them is calculated. The calculation formula is shown in formula (3): Each pair of joints includes: a left hip (23) and a left knee (25), a right hip (24) and a right knee (26), a left hip (23) and a left ankle (27), and a right hip (24) and a right ankle (28); Secondly, the distance between the joint pairs calculated using formula (4) is normalized to obtain the normalized distance: Thirdly, the maximum distance set is screened for all normalized distances of the subjects, and samples that are highly similar to the target squat sample but have a small difference in feature vector are eliminated to obtain the screened maximum distance set; for the screened maximum distance set, its average distance value is calculated; Assume D (k) is the set of all normalized distances of the kth person, by setting a threshold T max , filter out the maximum distance set D max (k) , the specific formula is shown in formula (5): D max (k) ={d i ' j (k) ∈D (k) |d i ' j (k) ≤T max } (5) For the maximum distance set after screening, use formula (6) to calculate its average distance value: Finally, the squat category to which the subject's current posture belongs is determined based on the calculated average distance value.

4. The method for real-time counting of multiple squats according to claim 3, characterized in that: The calculation formula for the action standardization score in S6 is shown in formula (7):

5. The method for real-time counting of multiple squats according to claim 1, characterized in that: The 33 joints are: nose, inside the left eye, inside the right eye, left eye, right eye, outside the left eye, outside the eye, left ear, right ear, left mouth, right mouth, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left little finger, right little finger, left index finger, right index finger, left thumb, right thumb, left hip, right hip, left knee, right knee, left ankle, right ankle, left heel, right heel, left toe and right toe.

6. The method for real-time counting of multiple squats according to claim 1, characterized in that: The video is obtained by using multiple cameras to capture the subject from different perspectives at the same time, and the captured video images are preprocessed before face recognition. The preprocessing includes denoising, lighting correction and deblurring.

Images