Multi-person taijiquan movement image recognition scoring method and storage medium

Abstract

The present application relates to the technical field of computer vision, in particular to a multi-person Taijiquan movement image recognition and scoring method and a storage medium. The scoring method comprises the following steps: synchronously capturing N global images of multi-person Taijiquan movement in a target area from N different angles, and simultaneously tracking the three-dimensional coordinates of UWB tags worn by each participant at each shooting moment by using UWB positioning technology; according to the three-dimensional coordinates of the UWB tags worn by each participant at each shooting moment, N local images corresponding to each participant are cut out from the N global images; the pre-trained key point detection model is used to detect the human body key point coordinates of the participant in each local image, and then generate N key point coordinate time series of the participant, and input the N key point coordinate time series into N pre-trained scoring models respectively to obtain the final score of the participant by averaging the N scores of the participant. The present application solves the scoring problem in multi-person Taijiquan movement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a multi-person Taijiquan movement image recognition and scoring method and a storage medium. BACKGROUND

[0002] With the acceleration of modern life pace and the increasing of social pressure, people's attention to health problems is increasing day by day. Under this background, traditional fitness exercises, especially Taijiquan, have gradually been favored by the public because of its unique health effects and slow-paced exercise methods. Taijiquan not only can enhance physical fitness and improve body coordination, but also has the function of self-cultivation and health preservation, and has an important role in promoting mental health.

[0003] In the scoring process of Taijiquan movement, the traditional method mainly relies on manual operation. This method is often affected by the subjectivity of the coach or the judge, and lacks unified and objective evaluation criteria. In recent years, with the breakthrough development of computer vision technology, through image recognition and processing, the movements and postures of the athletes can be captured and analyzed, thereby providing objective basis for movement scoring and improving the accuracy and fairness of scoring.

[0004] However, most of the existing technical solutions are designed for single-person movement. How to use computer vision technology to accurately identify and score the movements of each athlete in multi-person Taijiquan movement has become a problem to be solved. SUMMARY

[0005] In order to solve the above problems in the prior art, the present application provides a multi-person Taijiquan movement image recognition and scoring method, an electronic device and a storage medium, which solve the problem of scoring in multi-person Taijiquan movement.

[0006] The first aspect of the present application provides a multi-person Taijiquan movement image recognition and scoring method, which comprises:

[0007] N cameras are used to synchronously capture N global images of multi-person Taijiquan movement in a target area from N different angles, and UWB (Ultra Wide Band) positioning technology is used to track the three-dimensional coordinates of the UWB tags worn by each participant at each shooting moment;

[0008] According to the three-dimensional coordinates of the UWB tags worn by each participant at each shooting moment, N local images corresponding to each participant are cropped from the N global images;

[0009] According to the N local images corresponding to each participant, a pre-trained key point detection model is used to detect the human key point coordinates of the participant in each local image, and then generate N key point coordinate time sequences of the participant.

[0010] inputting the N key point coordinate time series of each participant into N pre-trained scoring models respectively to obtain N scores of the participant;

[0011] averaging the N scores of each participant to obtain a final score of the participant;

[0012] wherein N is a preset value.

[0013] Preferably, the target area is a rectangular sports field, and the outer frame is drawn in a prominent color.

[0014] The N cameras are fixedly arranged above or obliquely above the target area, and the position of each camera is higher than all the participants to avoid portrait overlap, and the shooting range of each camera is sufficient to cover the target area.

[0015] Each participant wears a UWB positioning tag in front of the chest, and each UWB positioning tag has an ID number as the ID of the corresponding participant.

[0016] Four UWB positioning base stations are arranged directly above the four vertices of the target area, and the positions of the four UWB positioning base stations are higher than all the participants.

[0017] The N cameras, all UWB positioning tags, and the four UWB positioning base stations are pre-uniformly calibrated in time.

[0018] Preferably, the step of "cutting out N local images corresponding to each participant from the N global images according to the three-dimensional coordinates of the UWB tag worn by each participant at each shooting time" comprises:

[0019] preprocessing the N global images to obtain N preprocessed images;

[0020] According to the N preprocessed images and the three-dimensional coordinates of the UWB tags of each participant at each shooting time, the two-dimensional projection coordinates of the UWB tags of each participant in each image are calculated.

[0021] According to the N preprocessed images and the two-dimensional projection coordinates of each participant, a pre-trained human detection model is used to detect the rectangular frame position of each participant in each preprocessed image.

[0022] According to the rectangular frame position, the N preprocessed images are cut, and the cut rectangular images are scaled to a preset first size to obtain N local images corresponding to each participant.

[0023] Preferably, the step of "calculating the two-dimensional projection coordinates of the UWB tag of each participant in each image according to the N pre-processed images and the three-dimensional coordinates of the UWB tag of each participant at each shooting time" comprises:

[0024] According to the positions of the four UWB positioning base stations and the three-dimensional coordinates of the UWB tag of each participant at each shooting time, the vertical projection coordinates of the UWB tag of each participant in the target area at each shooting time are calculated.

[0025] The four vertex coordinates of the target area image in each image are detected using OpenCV (Open Source Computer Vision Library, an open-source computer vision and machine learning software library).

[0026] According to the projection coordinates of the UWB tag of each participant in the target area at each shooting time, and the four vertex coordinates of the target area image in each image, the two-dimensional projection coordinates of the UWB tag of each participant in the image are calculated.

[0027] Preferably, the step of "detecting the human key point coordinates of each participant in each local image using a pre-trained key point detection model, and then generating an N key point coordinate time sequence of the participant" comprises:

[0028] From the first local image of a participant, one image is selected each time and input into the pre-trained key point detection model to obtain the two-dimensional human key point coordinates of the participant on each image in the first local image, and then a first key point coordinate time sequence of the participant is generated.

[0029] From the second local image of the participant, one image is selected each time and input into the key point detection model to obtain the two-dimensional human key point coordinates of the participant on each image in the second local image, and then a second key point coordinate time sequence of the participant is generated.

[0030] Repeat until the Nth key point coordinate time sequence of the participant corresponding to the Nth local image is generated.

[0031] Preferably, the training step of the human detection model comprises:

[0032] Collecting M times of multi-person Taijiquan exercise data; wherein, the multi-person Taijiquan exercise data of each time includes: N global sample images synchronously photographed by the N cameras from N different angles on the target area and three-dimensional coordinates of UWB tags worn by P participants in each shooting moment in this time; M and P are preset values, and the participants in different times are the same, partially the same or different;

[0033] Preprocessing each global sample image to obtain a set of preprocessed global sample images;

[0034] According to each image in the set of preprocessed global sample images and the three-dimensional coordinates of the UWB tags of each corresponding participant at the shooting moment of the image, the two-dimensional projection coordinates of the UWB tags of each corresponding participant in each preprocessed global sample image are calculated;

[0035] For each of the preprocessed global sample images, the rectangular frame position of the human body image of each corresponding participant is obtained using OpenCV;

[0036] Selecting a sample image from the set of preprocessed global sample images, and inputting the selected sample image and the two-dimensional projection coordinates of each corresponding participant in the image into the human body detection model to detect the rectangular frame position of the human body image of each corresponding participant in the sample image;

[0037] According to the rectangular frame position obtained by openCV and the detected rectangular frame position, a first loss function is calculated, and the parameters of the human body detection model are adjusted;

[0038] Repeat the steps of selecting a sample image, inputting a model, calculating the first loss function and adjusting parameters until the first loss function no longer decreases, and obtain the trained human body detection model.

[0039] Preferably, the training steps of the key point detection model include:

[0040] According to the manually obtained rectangular frame position, each image in the set of preprocessed global sample images is cropped, and the cropped rectangular image is scaled to the preset first size, thereby obtaining a set of local sample images;

[0041] For each local sample image in the set of local sample images, human body key points are labeled using OpenCV, and two-dimensional key point coordinates are obtained;

[0042] Selecting a sample image from the set of local sample images, and inputting the selected sample image into the key point detection model to detect the two-dimensional key point coordinates in the sample image;

[0043] calculate a second loss function according to the two-dimensional key point coordinates obtained by the OpenCV and the detected two-dimensional key point coordinates, and adjust parameters of the key point detection model;

[0044] The steps of selecting a sample image, inputting a model, calculating the second loss function, and adjusting parameters are repeatedly performed until the second loss function no longer decreases, and a trained key point detection model is obtained.

[0045] Preferably, N score models correspond to N shooting angles one by one.

[0046] The N score models have the same structure, and the training steps include:

[0047] According to the two-dimensional key point coordinates in the N local sample images corresponding to each participant in each session in the local sample image set, an N key point coordinate time sequence of each participant in each session is generated, thereby obtaining a time sequence sample set; the time sequence sample set includes MxP time sequence samples, and each time sequence sample includes N key point coordinate time sequences.

[0048] According to the N local sample images corresponding to each participant in each session in the local sample image set, the standardization of the participant's action is scored by an artificial person, thereby obtaining MxP scores.

[0049] From the time sequence sample set, a time sequence sample is selected, and the N key point coordinate time sequences in the sample are respectively input into N score models, and N predicted scores are output.

[0050] According to the artificial score and the N predicted scores, N third loss functions are respectively calculated, and N score models are respectively adjusted.

[0051] The steps of selecting a time sequence sample, inputting N score models, calculating N third loss functions, and adjusting N score models are repeatedly performed until a preset iteration number is reached, and N trained score models are obtained.

[0052] Preferably, the human key points of each participant include: head, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left middle finger tip, right middle finger tip, left hip, right hip, left knee, right knee, left ankle, right ankle, left foot palm, and right foot palm.

[0053] The human form detection model is a modified design of a YOLOv3 (You Only Look Once version 3, a target detection algorithm) network structure, and the guide model focuses attention on the two-dimensional projection coordinates of each participant.

[0054] The key point detection model is OpenPose (a human pose estimation framework based on deep learning, which can estimate the key points of the human body in real time);

[0055] The scoring model comprises an input layer, an LSTM layer, a full connection layer and an output layer.

[0056] In a second aspect, the application provides a computer readable storage medium storing a computer program capable of being loaded and executed by a processor to perform the method described above.

[0057] The application has the following beneficial effects:

[0058] The application uses UWB positioning technology to track the three-dimensional coordinates of the UWB tags worn by each participant at each shooting moment, which helps to quickly calculate the positions of the participants in the image and facilitates the identification of the identity of each participant in the image. The N key point coordinate time series of each participant are input into N pre-trained scoring models, and the average value of the N scores obtained is taken to obtain the final score of the participant. This way, the participant can be observed and evaluated from N different angles, and the score obtained is more accurate. Therefore, the application solves the problem of scoring each participant in a multi-person Tai Chi movement using image recognition methods. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a main step schematic diagram of the multi-person Tai Chi movement image recognition scoring method embodiment of the application;

[0060] Figure 2 is a main step schematic diagram of the human shape detection model training in the embodiment of the application;

[0061] Figure 3 is a main step schematic diagram of the key point detection model training in the embodiment of the application;

[0062] Figure 4 is a main step schematic diagram of the N scoring model training in the embodiment of the application. DETAILED DESCRIPTION

[0063] The preferred embodiments of the application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the application and are not intended to limit the protection scope of the application.

[0064] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0065] It should be noted that, in the description of the present application, the terms "first", "second" are only for the convenience of description, and do not indicate or imply the relative importance of the devices, elements or parameters, and therefore cannot be understood as a limitation on the present application. In addition, the term "and / or" in the present application is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper, if not specially stated, generally represents a "or" relationship between the front and rear associated objects.

[0066] Figure 1 is the main step schematic diagram of the multi-person taijiquan movement image recognition and scoring method embodiment of the present application. As Figure 1 shown, the scoring method of the present embodiment includes:

[0067] Step S10, synchronously capturing N-way global images of multi-person taijiquan movement in the target area from N different angles by using N cameras, and tracking the three-dimensional coordinates of the UWB tags worn by each participant at each shooting moment by using UWB positioning technology.

[0068] In the present embodiment, N is a preset value; the target area is a rectangular sports field, and the outer frame is drawn in a prominent color. According to the actual situation, it can also be an elliptical or circular field; N cameras are fixedly arranged above or obliquely above the target area, and the position of each camera is higher than all participants to avoid image overlap (maintain appropriate distance between people), and the shooting range of each camera is sufficient to cover the target area; each participant wears a UWB positioning tag in front of the chest, and each UWB positioning tag has an ID number as the ID of the corresponding participant; four UWB positioning base stations are arranged directly above the four vertices of the target area, and the positions are all higher than all participants; the N cameras, all UWB positioning tags and four UWB positioning base stations are pre-uniformly calibrated time.

[0069] Step S20, according to the three-dimensional coordinates of the UWB tags worn by each participant at each shooting moment, cutting out N-way local images corresponding to each participant from N-way global images.

[0070] This step specifically includes steps S21-S24:

[0071] Step S21, pre-process the N-way global image to obtain an N-way pre-processed image.

[0072] Specifically, each image in the N-way global image is denoised and scaled to a preset second size to obtain an N-way pre-processed image.

[0073] Step S22, according to the N-way pre-processed image and the three-dimensional coordinates of the UWB tag of each participant at each shooting time, the two-dimensional projection coordinates of the UWB tag of each participant in each image are calculated.

[0074] This step can specifically include steps S221-S223:

[0075] Step S221, according to the positions of the four UWB positioning base stations and the three-dimensional coordinates of the UWB tag of each participant at each shooting time, the vertical projection coordinates of the UWB tag of each participant in the target area at each shooting time are calculated.

[0076] In this step, the UWB tag position at each shooting time is vertically projected onto the plane where the target area is located, which roughly determines the position of each participant in the target area at each shooting time.

[0077] Step S222, using OpenCV to detect the four vertex coordinates of the target area image in each image.

[0078] Step S223, according to the projection coordinates of the UWB tag of each participant in the target area at each shooting time, and the four vertex coordinates of the target area image in each image, the two-dimensional projection coordinates of the UWB tag of each participant in the image are calculated.

[0079] In this step, the approximate positions of each participant in the image are determined, but each participant corresponds to only one point position, and the complete image of each participant needs to be divided out by the following step S23 using a rectangular frame.

[0080] Step S23, according to the N-way pre-processed image and the two-dimensional projection coordinates of each participant, a pre-trained human detection model is used to detect the rectangular frame position of each participant in each pre-processed image.

[0081] Step S24, according to the rectangular frame position, the N-way pre-processed image is cropped, and the cropped rectangular image is scaled to a preset first size to obtain an N-way local image corresponding to each participant.

[0082] Step S30, according to the N-way local image corresponding to each participant, the pre-trained key point detection model is used to detect the human body key point coordinates of the participant in each local image, and then N key point coordinate time sequences of the participant are generated.

[0083] In the embodiment, the human body key points of each participant include: head, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left middle finger tip, right middle finger tip, left hip, right hip, left knee, right knee, left ankle, right ankle, left foot and right foot, a total of 17 key points.

[0084] From the first local image of a participant, one image is selected each time to input into the pre-trained key point detection model, and two-dimensional human body key point coordinates of the participant on each image in the first local image are obtained (individual detection can be filled with zero), and the key point coordinates are arranged in time sequence to generate a first key point coordinate time sequence (each value in the sequence is spliced by 17 point coordinates).

[0085] From the second local image of the participant, one image is selected each time to input into the key point detection model, and two-dimensional human body key point coordinates of the participant on each image in the second local image are obtained, and a second key point coordinate time sequence of the participant is generated.

[0086] Repeat the execution until the Nth key point coordinate time sequence of the participant corresponding to the Nth local image is generated, and N key point coordinate time sequences of the participant are obtained.

[0087] Step S40, inputting the N key point coordinate time sequences of each participant into N pre-trained scoring models respectively to obtain N scores of the participant.

[0088] Step S50, averaging the N scores of each participant to obtain the final score of the participant.

[0089] In an optional embodiment, the human shape detection model is designed by modifying the YOLOv3 network structure, and the purpose of the modification is to fuse the image features and the two-dimensional projection coordinate features of each participant (UWB tag), and guide the model to pay attention to the two-dimensional projection coordinates of each participant, so as to speed up the detection speed and distinguish the corresponding relationship between the detected human image and the participant ID.

[0090] Figure 2 is the main step diagram of the human shape detection model training in the embodiment of the application. As shown in Figure 2 The main steps of the human shape detection model training can include A10-A50:

[0091] Step A10, collect M times of multi-person Taijiquan movement data.

[0092] Wherein, the multi-person Taijiquan movement data of each time includes N global sample images synchronously photographed by the N cameras from N different angles on the multi-person Taijiquan in the target area and three-dimensional coordinates of UWB tags worn by P participants in this time at each shooting time; M and P are preset values, and the participants in different times can be the same, partially the same or different.

[0093] Step A20, pre-process each global sample image to obtain a set of pre-processed global sample images.

[0094] Step A30, according to each image in the set of pre-processed global sample images and the three-dimensional coordinates of the UWB tags of each corresponding participant at the image shooting time, calculate the two-dimensional projection coordinates of the UWB tags of each corresponding participant in each pre-processed global sample image.

[0095] Step A40, for each pre-processed global sample image, use OpenCV to obtain the rectangular frame position of the human body image of each corresponding participant.

[0096] Step A50, select a sample image from the set of pre-processed global sample images, and input the selected sample image and the two-dimensional projection coordinates of each corresponding participant in the image into the human body detection model to detect the rectangular frame position of the human body image of each corresponding participant in the sample image.

[0097] Step A60, calculate a first loss function according to the rectangular frame position obtained by openCV and the detected rectangular frame position, and adjust the parameters of the human body detection model.

[0098] Step A70, go to step A50 until the first loss function no longer decreases, and obtain the trained human body detection model.

[0099] In another optional embodiment, the key point detection model uses OpenPose, and its training method can use the set of pre-processed global sample images obtained by steps A10-A20, and then perform steps B10-B50 as shown in Figure 3 .

[0100] Step B10, crop each image in the set of pre-processed global sample images according to the manually obtained rectangular frame position, and scale the cropped rectangular image to a preset first size, thereby obtaining a set of local sample images.

[0101] Step B20, for each local sample image in the set of local sample images, use OpenCV to label human body key points and obtain two-dimensional key point coordinates.

[0102] Step B30, select a sample image from the local sample image set, and input the selected sample image into the key point detection model to detect two-dimensional key point coordinates in the sample image;

[0103] Step B40, calculate a second loss function according to the two-dimensional key point coordinates obtained by OpenCV and the detected two-dimensional key point coordinates, and adjust the parameters of the key point detection model;

[0104] Step B50, go to step B30 until the second loss function no longer decreases, and obtain the trained key point detection model.

[0105] In another optional embodiment, the structures of the N scoring models are the same, and each scoring model can include an input layer, an LSTM layer, a full connection layer, and an output layer. The N scoring models correspond one-to-one to the N shooting angles. The training method can use the local sample image set generated in step B10, and then perform steps C10-C50 as shown to train the N scoring models together: Figure 4

[0106] Step C10, generate N key point coordinate time series of each participant in each session according to the two-dimensional key point coordinates in the N local sample images corresponding to each participant in each session in the local sample image set, thereby obtaining a time series sample set; the time series sample set includes MxP time series samples, and each time series sample includes N key point coordinate time series.

[0107] Step C20, score the standardization of the action of each participant in each session according to the N local sample images corresponding to each participant in each session in the local sample image set by a human, thereby obtaining MxP scores.

[0108] Step C30, select a time series sample from the time series sample set, and input the N key point coordinate time series in the sample into the N scoring models respectively, and output N predicted scores.

[0109] Step C40, calculate N third loss functions respectively according to the artificial scores and the N predicted scores, and adjust the corresponding N scoring models respectively.

[0110] When calculating each third loss function, the artificial score is used as the true value, but the predicted value is the predicted score output by the corresponding model.

[0111] Step C50, go to step C30 until a preset number of iterations is reached, and obtain the trained N scoring models.

[0112] ​Although each step is described in the above-described order in each of the above-described embodiments, it is understood by those skilled in the art that, in order to achieve the effects of the embodiments, the different steps do not have to be executed in such an order, and can be executed simultaneously (in parallel) or in a reversed order, and these simple changes are within the protection scope of the present application.

[0113] Based on the above-described scoring method, the present application further provides an embodiment of a computer-readable storage medium, and the storage medium of the embodiment stores a computer program capable of being loaded and executed by a processor to perform the method described above.

[0114] The computer-readable storage medium can include a U disk, a mobile hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes.

[0115] Those skilled in the art should be able to realize that the method steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both, and in order to clearly illustrate the interchangeability of electronic hardware and software, the components and steps of each example have been described in the above description in general terms. Whether the functions are executed by electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0116] So far, the technical solution of the present application has been described in combination with the preferred embodiments shown in the drawings. However, those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without deviating from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

Claims

1. A multi-person taijiquan movement image recognition scoring method, characterized in that, The method comprises: synchronously capturing N global images of multiple Tai Chi movements in a target area from N different angles by using N cameras, and tracking three-dimensional coordinates of UWB tags worn by each participant at each shooting time by using UWB positioning technology; cropping N local images corresponding to each participant from the N global images according to the three-dimensional coordinates of the UWB tags worn by each participant at each shooting time; detecting human body key point coordinates of each participant in each local image by using a pre-trained key point detection model according to the N local images corresponding to each participant, and further generating N key point coordinate time sequences of the participant; inputting the N key point coordinate time sequences of each participant into N pre-trained scoring models respectively to obtain N scores of the participant; averaging the N scores of each participant to obtain a final score of the participant; wherein N is a preset value; the N scoring models correspond to the N shooting angles one by one; the N scoring models have the same structure, and the training steps comprise: generating N key point coordinate time sequences of each participant in each session according to two-dimensional key point coordinates in N local sample images corresponding to each participant in each session in a local sample image set, thereby obtaining a time sequence sample set; the time sequence sample set comprises MxP time sequence samples, and each time sequence sample comprises N key point coordinate time sequences; scoring the standardization of the action of each participant by a person according to N local sample images corresponding to each participant in each session in the local sample image set, thereby obtaining MxP scores; selecting a time sequence sample from the time sequence sample set, and inputting N key point coordinate time sequences in the sample into N scoring models respectively to output N predicted scores; calculating N third loss functions respectively according to the artificial scores and the N predicted scores, and adjusting the corresponding N scoring models respectively; repeating the steps of selecting the time sequence sample, inputting into the N scoring models, calculating the N third loss functions, and adjusting the N scoring models until a preset iteration number is reached, thereby obtaining the N trained scoring models.

2. The multi-person Tai Chi movement image recognition scoring method according to claim 1, wherein: the target area is a rectangular playing field, and the outer frame is drawn in a prominent color; the N cameras are fixedly arranged above or obliquely above the target area, and the position of each camera is higher than all the participants to avoid image overlap, and the shooting range of each camera is sufficient to cover the target area; each participant wears a UWB positioning tag on the chest, and each UWB positioning tag has an ID number as the ID of the corresponding participant; four UWB positioning base stations are arranged directly above the four vertices of the target area, and the positions of the four UWB positioning base stations are higher than all the participants; the N cameras, all the UWB positioning tags, and the four UWB positioning base stations are pre-uniformly calibrated in time.

3. The multi-person taijiquan movement image recognition scoring method according to claim 2, characterized in that, The step of "cutting out N local images corresponding to each participant from the N global images according to the three-dimensional coordinates of the UWB tags worn by each participant at each shooting time" comprises: Preprocessing the N global images to obtain N preprocessed images; According to the three-dimensional coordinates of the UWB tags of each participant at each shooting time, calculating the two-dimensional projection coordinates of the UWB tags of each participant in each image according to the N preprocessed images and the three-dimensional coordinates of the UWB tags of each participant at each shooting time; According to the N preprocessed images and the two-dimensional projection coordinates of each participant, a pre-trained human body detection model is used to detect the rectangular frame position of each participant in each preprocessed image; According to the rectangular frame position, the N preprocessed images are cut out and scaled to a preset first size to obtain N local images corresponding to each participant.

4. The multi-person taijiquan movement image recognition scoring method according to claim 3, characterized in that, The step of "calculating the two-dimensional projection coordinates of the UWB tags of each participant in each image according to the N preprocessed images and the three-dimensional coordinates of the UWB tags of each participant at each shooting time" comprises: According to the positions of the four UWB positioning base stations and the three-dimensional coordinates of the UWB tags of each participant at each shooting time, the vertical projection coordinates of the UWB tags of each participant in the target area at each shooting time are calculated; Using OpenCV to detect the four vertex coordinates of the image corresponding to the target area in each image; According to the projection coordinates of the UWB tags of each participant in the target area at each shooting time and the four vertex coordinates of the image corresponding to the target area in each image, the two-dimensional projection coordinates of the UWB tags of each participant in the image are calculated.

5. The multi-person taijiquan movement image recognition scoring method according to claim 1, characterized in that, The step of "using a pre-trained key point detection model to detect the human body key point coordinates of each participant in each local image, and then generating N key point coordinate time series of the participant according to the N local images corresponding to each participant" comprises: From the first local image of a participant, one image is selected each time and input into the pre-trained key point detection model to obtain the two-dimensional human body key point coordinates of the participant on each image in the first local image, and then generate the first key point coordinate time series of the participant; From the second local image of the participant, one image is selected each time and input into the key point detection model to obtain the two-dimensional human body key point coordinates of the participant on each image in the second local image, and then generate the second key point coordinate time series of the participant; Repeat until the Nth key point coordinate time series of the participant corresponding to the Nth local image is generated.

6. The multi-person taijiquan movement image recognition scoring method according to claim 3, characterized in that, The training steps of the human body detection model comprise: Collecting M times of multi-person Taijiquan movement data; wherein, the multi-person Taijiquan movement data of each time includes: N global sample images synchronously photographed by the N cameras from N different angles on the target area and three-dimensional coordinates of UWB tags worn by P participants in each shooting moment in this time; M and P are preset values, and the participants in different times are the same, partially the same or different; Preprocessing each global sample image to obtain a set of preprocessed global sample images; According to each image in the set of preprocessed global sample images and the three-dimensional coordinates of the UWB tags of each corresponding participant at the shooting moment of the image, the two-dimensional projection coordinates of the UWB tags of each corresponding participant in each preprocessed global sample image are calculated; For each of the preprocessed global sample images, the OpenCV is used to obtain the rectangular frame position of the human body image of each corresponding participant; Selecting a sample image from the set of preprocessed global sample images, and inputting the selected sample image and the two-dimensional projection coordinates of each corresponding participant in the image into the human body detection model to detect the rectangular frame position of the human body image of each corresponding participant in the sample image; According to the rectangular frame position obtained by OpenCV and the detected rectangular frame position, a first loss function is calculated, and the parameters of the human body detection model are adjusted; The steps of selecting a sample image, inputting a model, calculating the first loss function and adjusting parameters are repeatedly executed until the first loss function no longer decreases, and a trained human body detection model is obtained.

7. The multi-person taijiquan movement image recognition scoring method according to claim 6, characterized in that, The training steps of the key point detection model include: For each image in the set of preprocessed global sample images, the rectangular frame position is obtained by artificial acquisition, and the rectangular image obtained by cutting is scaled to the preset first size, thereby obtaining a set of local sample images; For each local sample image in the set of local sample images, the human body key points are labeled and the two-dimensional key point coordinates are obtained by using OpenCV; Selecting a sample image from the set of local sample images, and inputting the selected sample image into the key point detection model to detect the two-dimensional key point coordinates in the sample image; According to the two-dimensional key point coordinates obtained by OpenCV and the detected two-dimensional key point coordinates, a second loss function is calculated, and the parameters of the key point detection model are adjusted; The steps of selecting a sample image, inputting a model, calculating the second loss function and adjusting parameters are repeatedly executed until the second loss function no longer decreases, and a trained key point detection model is obtained.

8. The multi-person taijiquan movement image recognition scoring method according to any one of claims 3-4, 6-7, characterized in that, The human body key points of each participant include: head, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left middle finger tip, right middle finger tip, left hip, right hip, left knee, right knee, left ankle, right ankle, left foot and right foot; The human body detection model is a modified design of YOLOv3 network structure, and the guide model pays attention to the two-dimensional projection coordinates of each participant; The key point detection model is OpenPose. The scoring model comprises: an input layer, an LSTM layer, a full connection layer, and an output layer.

9. A computer-readable storage medium, characterized in that, Computer programs are stored which can be loaded and executed by the processor to perform the method of any of claims 1-8.

Images