FMS dynamic action shake detection method based on deep learning and time-frequency characteristics

Abstract

The invention relates to an FMS dynamic action shake detection method based on deep learning and time-frequency characteristics, and belongs to the technical field of exercise training.The method comprises the steps that action videos in the process that a testee completes FMS dynamic actions are obtained from the front face position and the side face position, and the action videos are obtained; detecting the position change of the trunk center point of the testee in the action video in combination with a deep learning model; and on the basis of the trunk center point position change time sequence data, constructing characteristic quantities from two levels of time domain and frequency domain, and realizing rapid and accurate detection of the body shaking condition according to the characteristic quantity difference of finishing the FMS action without shaking and finishing the FMS action with shaking.

Description

Technical Field

[0001] The present invention relates to the technical field of sports training, and particularly to an FMS dynamic movement jitter detection method based on deep learning and time-frequency features. Background Art

[0002] ‌FMS (Functional Movement Screen)‌, namely Functional Movement Screening, is a test method for detecting the abilities of athletes such as overall movement control stability, body balance ability, softness, and proprioception.

[0003] In the prior art, the FMS assessment usually adopts the traditional manual assessment method. The assessor combines the FMS scoring criteria to score the performance of the subject in the FMS movements. For the hurdle step and the straight-leg lunge squat movements, if the subject's torso moves or sways significantly, points will be deducted. This manual assessment method has problems such as strong subjectivity and low efficiency. To solve the above problems, some studies have proposed an FMS assessment method assisted by computer vision. Specifically, the computer vision algorithm is used to process the movement images, extract the position information of the human body bone key points, and then capture the FMS movement postures. According to the time-domain characteristics analysis of the time-series data of the position information, the body swaying situation is judged.

[0004] The time-domain characteristics analysis mainly analyzes the large-amplitude swaying, but there is also small-amplitude but high-frequency swaying when a person completes the FMS dynamic movement. It is difficult to accurately detect the body swaying only by performing time-domain characteristics analysis. Summary of the Invention

[0005] In view of this, it is necessary to provide an FMS dynamic movement jitter detection method based on deep learning and time-frequency features to solve the technical problem in the prior art that when a person completes the FMS dynamic movement, there is also small-amplitude but high-frequency data jitter, and it is difficult to accurately detect the body swaying only by performing time-domain characteristics analysis.

[0006] To solve the above problems, on the one hand, the present invention provides an FMS dynamic movement jitter detection method based on deep learning and time-frequency features, including: Obtain the movement video of the subject during the process of completing the FMS dynamic movement; Based on a trained complete deep learning model, detect the position of the torso center point in each movement image in the movement video to obtain the time-series data of the change in the position of the torso center point; Extract the frequency-domain features from the time-series data based on the Gabor transform, and extract the time-domain features from the time-series data; Based on the frequency domain features and the time domain features, it is determined whether there is any shaking during the process of the subject completing the FMS dynamic movement.

[0007] In some possible implementation manners, the action video includes action videos of the front and the side.

[0008] In some possible implementation manners, the trained deep learning model is obtained through the following steps: Obtain action image samples during the process of several training personnel completing the FMS dynamic movement, and construct a sample data set based on the action image samples; Obtain an initial deep learning model, where the initial deep learning model includes a first deep network and a second deep network. The input ends of the first deep network and the second deep network are connected to a common input layer, and the output ends are respectively connected to corresponding output layers. The number of layers and the inter-layer connection structures of the first deep network and the second deep network are different; Train the initial deep learning model based on the sample data set. During training, select the first deep network or the second deep network for training according to the loss change situations of the first deep network and the second deep network, and obtain a trained deep learning model.

[0009] In some possible implementation manners, the first deep network is a 30-layer neural network, there is a direct connection operation between every two layers of the first deep network, and the second deep network is a 16-layer neural network.

[0010] In some possible implementation manners, the feature parameters to be extracted during the training of the deep learning model include the distance between the center point of the prediction box and the coordinate origin, the included angle between the center point of the prediction box and the horizontal axis, the width of the prediction box, the height of the prediction box, and the product of the confidence level and the intersection over union, where the confidence level represents the confidence level that the prediction box contains the torso, and the intersection over union represents the intersection over union of the prediction box and the actual box.

[0011] In some possible implementation manners, the loss function during the training of the initial deep learning model is:

[0012]

[0013] In the formula, represents the loss, represents the training personnel number, represents the number of training personnel, represents the grid number of the action image, represents the number of grids, represents the included angle between the center point of the prediction box and the abscissa, represents the abscissa of the center point of the annotation box, represents the distance between the center point of the prediction box and the coordinate origin, represents the distance between the center point of the annotation box and the coordinate origin, represents the width of the prediction box, represents the width of the annotation box, represents the height of the prediction box, represents the height of the annotation box. When the prediction box contains the trunk part, it is 1. When the prediction box does not contain the trunk part, it is 0, represents the intersection over union (IoU) between the prediction box and the annotation box, where the annotation box is the position of the trunk in the annotated image.

[0014] In some possible implementation manners, the extraction formula of the frequency-domain feature is:

[0015]

[0016]

[0017] In the formula, represents the frequency-domain feature, represents the th data in the time-series data, represents the total number of the time-series data. The change amount between the th data and its adjacent data in the time-series data is greater than a preset first threshold, represents a positive integer, represents the imaginary unit, represents the frequency.

[0018] In some possible implementation manners, the extraction formula of the time-domain feature is:

[0019] In the formula, represents the time-domain feature, represents the total number of the time-series data, represents the th data in the time-series data, represents the average value of the time-series data.

[0020] In some possible implementation manners, based on the frequency-domain feature and the time-domain feature, to determine whether there is a shaking situation during the process of the subject completing the FMS dynamic action, it includes: judging whether the frequency-domain feature is greater than a preset second threshold and whether the time-domain feature is greater than a preset third threshold; If the frequency-domain feature is greater than a preset second threshold, or the time-domain feature is greater than a preset third threshold, it is determined that there is a shaking situation during the subject's completion of the FMS dynamic movement; If the frequency-domain feature is not greater than the preset second threshold and the time-domain feature is not greater than the preset third threshold, it is determined that there is no shaking situation during the subject's completion of the FMS dynamic movement.

[0021] On the other hand, the present invention also provides an FMS dynamic movement shaking detection device, including: An image acquisition unit, configured to acquire an action video during the subject's completion of the FMS dynamic movement; An image recognition unit, configured to detect the position of the torso center point in each action image in the action video based on a trained deep learning model, and obtain the time-series data of the position change of the torso center point; A time-frequency domain analysis unit, configured to extract a frequency-domain feature from the time-series data based on Gabor transform, extract a time-domain feature from the time-series data, and determine whether there is a shaking situation during the subject's completion of the FMS dynamic movement based on the frequency-domain feature and the time-domain feature.

[0022] The beneficial effects of the present invention are as follows: The FMS dynamic movement shaking detection method based on deep learning and time-frequency features provided by the present invention first acquires an action video during the subject's completion of the FMS dynamic movement, then combines a deep learning model to detect the position change of the subject's torso center point in the action video, and then constructs feature quantities from two levels of the time domain and the frequency domain based on the time-series data of the position change of the torso center point, and realizes the rapid and accurate detection of the body shaking situation according to the difference in feature quantities between the FMS movement completed without shaking and the FMS movement completed with shaking. Description of the Drawings

[0023] Figure 1 It is a schematic flowchart of an embodiment of the deep learning model training method provided by the present invention; Figure 2 It is a schematic structural diagram of an embodiment of the first deep network provided by the present invention; Figure 3 It is a schematic structural diagram of an embodiment of the second deep network provided by the present invention; Figure 4 It is a schematic flowchart of an embodiment of the FMS dynamic movement shaking detection method based on deep learning and time-frequency features provided by the present invention; Figure 5 For the present invention Figure 4 It is a schematic flowchart of an embodiment of step S404 in the present invention; Figure 6Schematic structural diagram of an embodiment of the FMS dynamic motion shaking detection device provided by the present invention. Detailed implementation manners

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] It should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the present invention illustrate operations implemented according to some embodiments of the present invention. It should be understood that the operations in the flowchart may not be implemented in sequence, and steps without logical context relationships may be reversed or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart or remove one or more operations from the flowchart under the guidance of the content of the present invention. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor systems and / or microcontroller systems.

[0026] The descriptions such as "first" and "second" involved in the embodiments of the present invention are only for the purpose of implicit description, and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Therefore, the technical features defined with "first" and "second" may explicitly or implicitly include at least one such feature. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone.

[0027] Referring to "embodiment" in this article means that the specific features, structures or characteristics described in combination with the embodiment may be included in at least one embodiment of the present invention. The appearance of this phrase at various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art understand explicitly and implicitly that the embodiments described herein can be combined with other embodiments.

[0028] The present invention provides an FMS dynamic motion shaking detection method and device based on deep learning and time-frequency features, which will be described separately below.

[0029] In the FMS dynamic action jitter detection method based on deep learning and time-frequency features provided by the present invention, an image recognition is performed using a trained deep learning model. Before use, the deep learning model needs to be trained. For example, Figure 1 as shown, the deep learning model is trained through the following steps: S101. Obtain action image samples of several training personnel during the FMS dynamic action process, and construct a sample data set based on the action image samples; Considering that the jitter of the FMS dynamic action is not only reflected in the change of the front position, but also in the change of the side position. In some embodiments of the present invention, the action images include the front and side action images to detect the jitter of the sagittal plane and the coronal plane during the process of the tester completing the FMS dynamic action.

[0030] It should be noted that to construct the sample data set, images need to be taken first, including: selecting different venues, allowing several training personnel to complete the FMS action respectively, and simultaneously shooting videos at two positions, the front and the side. An even frame extraction algorithm is used to extract one frame of image at a fixed time interval to form the original action image; in addition, for different venues, background images of the venue where the FMS dynamic action is completed are taken at two positions, the front and the side, to form background image samples, and the number of background image samples needs to be close to the number of original action image samples.

[0031] After the images are taken, the images need to be processed. The processing steps include: first, adjust the aspect ratio of the images to a square, and divide it into S S grids, each grid will be responsible for B prediction boxes, then normalize the position of each pixel point in the image to the range of [0, 1], and at the same time perform standardization processing on each channel, and adjust it using the mean and standard deviation.

[0032] Then, label the trunk part on the processed original action image. Take the lower left corner of the image as the coordinate origin, and label the trunk part in the image. The trunk part is described by a label box, including 4 parameters: the distance between the center point of the label box and the coordinate origin, the angle between the center point of the label box and the horizontal axis, and the width and height of the label box, to form the labeled action image. The original action image and the labeled action image form the action image sample, and the action image sample and the background image sample form the sample data set.

[0033] S102. Obtain an initial deep learning model. The initial deep learning model includes a first deep network and a second deep network. The input ends of the first deep network and the second deep network are connected to a common input layer, and the output ends are respectively connected to corresponding output layers. The number of layers and the inter-layer connection structure of the first deep network and the second deep network are different; To obtain a better deep learning model, in some embodiments of the present invention, the input layer includes a convolutional layer and a pooling layer. Among them, the convolutional layer uses a relatively large convolutional kernel to perform a convolution operation on the input sample action image to extract a preliminary feature map, and then processes it through batch normalization and the ReLU activation function. The pooling layer adopts the maximum pooling method, and uses a sampling kernel smaller than the convolutional layer to perform downsampling on the feature map. The result after downsampling is output to the first deep network and the second deep network; the first deep network is a 30-layer neural network as shown in Figure 2 and each layer performs a convolution operation with a 3 ×3 convolutional kernel, and then processes it through the ReLU activation function. There is a direct connection operation between every two layers; the second deep network is a 16-layer neural network as shown in Figure 3 and each layer performs a convolution operation with a 3 ×3 convolutional kernel, and then processes it through the ReLU activation function; the output layers corresponding to the first deep network and the second deep network both include a pooling layer and a fully connected layer.

[0034] S103. Train the initial deep learning model based on the sample data set. During training, select the first deep network or the second deep network for training according to the loss change situation of the first deep network and the second deep network to obtain a fully trained deep learning model.

[0035] It should be noted that the deep learning model of the front camera is trained with the action image samples of the front sides of several training personnel in the sample data set, and the deep learning model of the side camera is trained with the action image samples of the side sides of several training personnel in the sample data set to detect the front and side action images respectively.

[0036] To better train the model, in some embodiments of the present invention, the prediction box is the position of the torso in the predicted image, the annotation box is the position of the torso in the annotated image, and the feature parameters to be extracted during the training of the deep learning model include the distance between the center point of the prediction box and the coordinate origin, the angle between the center point of the prediction box and the horizontal axis, the width of the prediction box, the height of the prediction box, and the product of the confidence level and the intersection over union. Among them, the confidence level represents the confidence level that the prediction box contains the torso, and the intersection over union represents the intersection over union of the prediction box and the actual box. The prediction box is the position of the torso in the predicted image. Correspondingly, the loss function during the training of the initial deep learning model is:

[0037]

[0038] In the formula, represents the loss, represents the training personnel number, represents the number of training personnel, Indicates the grid number representing the action image, Indicates the number of grids, Indicates the angle between the center point of the prediction box and the abscissa, Indicates the abscissa of the center point of the annotation box, Indicates the distance between the center point of the prediction box and the origin of coordinates, Indicates the distance between the center point of the annotation box and the origin of coordinates, Indicates the width of the prediction box, Indicates the width of the annotation box, Indicates the height of the prediction box, Indicates the height of the annotation box, Indicates the probability that the prediction box contains the torso. When the prediction box contains the torso part, it is 1. When the prediction box does not contain the torso part, it is 0, Indicates the intersection over union of the prediction box and the annotation box.

[0039] Based on the above loss function, when selecting the first deep neural network or the second deep neural network for training according to the loss change situation of the first deep neural network and the second deep neural network, only the network with a loss less than 1 within 5 training epochs is retained for training, and finally a training-complete deep learning model for the front camera position and a training-complete deep learning model for the side camera position are obtained.

[0040] Based on the above obtained training-complete deep learning model, Figure 4 This is a schematic flowchart of an embodiment of the FMS dynamic action jitter detection method based on deep learning and time-frequency features provided by the present invention. As Figure 4 shown, the FMS dynamic action jitter detection method based on deep learning and time-frequency features includes: S401. Obtain the action video during the process of the subject completing the FMS dynamic action; It should be noted that the process of obtaining the action video is specifically as follows: Let the subject complete the FMS action respectively, and at the same time, shoot videos at two positions, the front and the side. The uniform frame extraction algorithm is used to extract one frame of image at a fixed time interval to form the action video.

[0041] S402. Based on the training-complete deep learning model, detect the position of the torso center point in each action image in the action video to obtain the time series data of the change in the position of the torso center point; It should be noted that the training-complete deep learning model of the front camera position is used to detect the image to obtain the front torso center point position, and the training-complete deep learning model of the side camera position is used to detect the image to obtain the side torso center point position.

[0042] Considering that when the subject sways during the completion of an action, the torso will inevitably move horizontally, so the position of the torso center point will surely change. Specifically, the horizontal position change of the torso center point during the action will increase. Therefore, the position of the torso center point is the abscissa of the torso center point, and the abscissa of the torso center point , in the process of the subject completing the hurdle step and the straight lunge squat actions, the time series data can be represented by a time series . It should be noted that there are also corresponding positive time series and side time series for the time series. Detecting sway on either the front or the side represents the existence of a sway situation.

[0043] S403. Extract frequency domain features from the time series data based on the Gabor transform, and extract time domain features from the time series data; It should be noted that considering that the time when sway appears during the completion of the FMS action is phased and uneven, the time series of the change in the position of the torso center point is a non-stationary signal. It is difficult to accurately evaluate body sway by directly extracting frequency domain features through Fourier transform. Therefore, the present invention uses the windowed Fourier transform to extract frequency domain features.

[0044] In order to extract better frequency domain features, the extraction formula for frequency domain features is:

[0045]

[0046]

[0047] In the formula, represents the frequency domain feature, represents the data of the th frame in the time series data, represents the total number of the time series data. The change amount between the rd data and the adjacent data in the time series data is greater than a preset first threshold, represents a positive integer, represents the imaginary unit, represents the frequency.

[0048] In order to extract better time domain features, in some embodiments of the present invention, the extraction formula for time domain features is:

[0049] In the formula, represents the time domain feature, that is, the variance, represents the total number of the time series data, represents the data of the th frame in the time series data, Represents the average value of time series data.

[0050] S404. Based on the frequency domain features and time domain features, determine whether there is any shaking during the process of the subject completing the FMS dynamic movement.

[0051] For better time-frequency domain analysis, in some embodiments of the present invention, as Figure 5 shown, step S404 includes: S501. Determine whether the frequency domain feature is greater than a preset second threshold and whether the time domain feature is greater than a preset third threshold; It should be noted that the second threshold is specifically the maximum value obtained by calculating the FMS dynamic movement when the subject has no shaking and the third threshold is specifically the maximum value of the variance of the trunk center point obtained by calculating the FMS dynamic movement when the subject is shaking. .

[0052] S502. If the frequency domain feature is greater than the preset second threshold or the time domain feature is greater than the preset third threshold, then determine that there is shaking during the process of the subject completing the FMS dynamic movement; S503. If the frequency domain feature is not greater than the preset second threshold and the time domain feature is not greater than the preset third threshold, then determine that there is no shaking during the process of the subject completing the FMS dynamic movement.

[0053] In summary, the FMS dynamic movement shaking detection method based on deep learning and time-frequency features of the present invention first obtains the action videos of the subject during the process of completing the FMS dynamic movement from two positions, the front and the side, combines the deep learning model to detect the position change of the trunk center point of the subject in the action videos, and then constructs feature quantities from two levels, the time domain and the frequency domain, based on the time series data of the position change of the trunk center point, and realizes the fast and accurate detection of the body shaking situation according to the difference in feature quantities between the FMS movement completed without shaking and the FMS movement completed with shaking.

[0054] To better implement a kind of FMS dynamic movement shaking detection method in the embodiments of the present invention, correspondingly, on the basis of a kind of FMS dynamic movement shaking detection method based on deep learning and time-frequency features, as Figure 6 shown, the embodiments of the present invention also provide an FMS dynamic movement shaking detection device 600, including: An image acquisition unit 601, configured to acquire the action videos of the subject during the process of completing the FMS dynamic movement; An image recognition unit 602, configured to detect the position of the trunk center point in each action image in the action video based on a trained deep learning model, and obtain the time series data of the position change of the trunk center point; A time-frequency domain analysis unit 603 is configured to extract frequency domain features from the time series data based on Gabor transform, extract time domain features from the time series data, and determine whether there is any wobbling during the process of the subject completing the FMS dynamic movement based on the frequency domain features and the time domain features.

[0055] The FMS dynamic movement wobbling detection device 600 provided in the above embodiment can implement the technical solutions described in the above embodiment of the FMS dynamic movement wobbling detection method based on deep learning and time-frequency features. The specific implementation principles of the above units can be referred to the corresponding contents in the above embodiment of the FMS dynamic movement wobbling detection method based on deep learning and time-frequency features, and will not be elaborated here.

[0056] The above has introduced in detail a method for detecting wobbling in FMS dynamic movements based on deep learning and time-frequency features. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0057] As mentioned above, the above are only the preferred specific implementation manners of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A FMS dynamic motion shake detection method based on deep learning and time-frequency features, characterized in that: include: Obtain the action video of the subject completing the FMS dynamic action; Based on a well-trained deep learning model, the position of the torso center point in each action image in the action video is detected to obtain time series data of the position change of the torso center point; Extracting frequency domain features from the time series data based on Gabor transform, and extracting time domain features from the time series data; Based on the frequency domain features and the time domain features, it is determined whether there is shaking when the subject completes the FMS dynamic action.

2. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claim 1 is characterized in that: The action video includes front and side action videos.

3. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claim 1 is characterized in that: The fully trained deep learning model is obtained through the following steps: Acquire action image samples of several trainees completing FMS dynamic actions, and construct a sample data set based on the action image samples; Obtain an initial deep learning model, wherein the initial deep learning model includes a first deep network and a second deep network, wherein input ends of the first deep network and the second deep network are connected to a common input layer, output ends are respectively connected to corresponding output layers, and the first deep network and the second deep network have different numbers of layers and inter-layer connection structures; The initial deep learning model is trained based on the sample data set. During training, the first deep network or the second deep network is selected for training according to the loss change of the first deep network and the second deep network to obtain a fully trained deep learning model.

4. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claim 3 is characterized in that: The first deep network is a 30-layer neural network, there is a direct connection operation between every two layers of the first deep network, and the second deep network is a 16-layer neural network.

5. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claim 3 is characterized in that: The feature parameters to be extracted during the deep learning model training include the distance between the center point of the prediction box and the coordinate origin, the angle between the center point of the prediction box and the horizontal axis, the width of the prediction box, the height of the prediction box, and the product of the confidence and the intersection-and-union ratio, wherein the confidence indicates the confidence that the prediction box contains a torso, and the intersection-and-union ratio indicates the intersection-and-union ratio of the prediction box and the actual box.

6. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claim 3 is characterized in that: The loss function during the initial deep learning model training is: In the formula, Indicates loss, Indicates the training personnel number, Indicates the number of trainees, Indicates the grid number of the action image, represents the number of grids, Represents the angle between the center point of the prediction box and the horizontal coordinate, Indicates the horizontal coordinate of the center point of the annotation box. Indicates the distance between the center point of the prediction box and the coordinate origin, Indicates the distance between the center point of the annotation box and the coordinate origin. Indicates the width of the prediction box, Indicates the width of the annotation box. Indicates the height of the prediction box, Indicates the height of the annotation box. When the predicted box contains the torso, is 1, when the predicted frame does not contain the torso part, is 0, It represents the intersection-over-union ratio of the predicted box and the labeled box.

7. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claim 1 is characterized in that: The frequency domain feature extraction formula is: In the formula, represents the frequency domain characteristics, Indicates the time series data data, Represents the total number of time series data, the number of The difference between the data and the adjacent data is greater than the preset first threshold. represents a positive integer, represents the imaginary unit, Indicates frequency.

8. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claim 1 is characterized in that: The extraction formula of the time domain feature is: In the formula, Represents the time domain characteristics, Indicates the total number of time series data. Indicates the time series data data, Represents the average value of time series data.

9. The FMS dynamic motion shake detection method based on deep learning and time-frequency features according to claims 7 and 8 is characterized in that: Based on the frequency domain features and the time domain features, judging whether there is shaking during the subject's FMS dynamic action, including: Determine whether the frequency domain feature is greater than a preset second threshold, and whether the time domain feature is greater than a preset third threshold; If the frequency domain feature is greater than a preset second threshold, or the time domain feature is greater than a preset third threshold, it is determined that the subject is shaking during the FMS dynamic action; If the frequency domain feature is not greater than a preset second threshold, and the time domain feature is not greater than a preset third threshold, it is determined that there is no shaking when the subject completes the FMS dynamic action.

10. An FMS dynamic motion sway detection device, characterized in that: include: An image acquisition unit, used to acquire a motion video of the subject completing a dynamic motion of the FMS; An image recognition unit, configured to detect the position of the center point of the torso in each action image in the action video based on a well-trained deep learning model, and obtain time series data of the position change of the center point of the torso; The time-frequency domain analysis unit is used to extract frequency domain features from the time series data based on Gabor transform, and extract time domain features from the time series data, and determine whether there is shaking when the subject completes the FMS dynamic action based on the frequency domain features and the time domain features.

Images