RECONSTRUCTION OF PHYSICAL DANCE

TEACHING CONTENT AND MOVEMENT

RECOGNITION BASED ON A MACHINE

LEARNING MODEL

Lei Li

Football College, Wuhan Sports University, Wuhan, Hubei, 430079, China

Tingting Yang*

School of Journalism and Communication, Wuhan Sports University, Wuhan,

Hubei, 430079, China

ytt342784081@163.com

Reception: 14/11/2022 Acceptance: 17/01/2023 Publication: 05/03/2023

Suggested citation:

L., Lei and Y., Tingting (2023). Reconstruction of physical dance teaching

content and movement recognition based on a machine learning model.

3C TIC. Cuadernos de desarrollo aplicados a las TIC, 12(1), 267-285. https://

doi.org/10.17993/3ctic.2023.121.267-285

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

267

ABSTRACT

With the technological development of movement recognition based on machine

learning model algorithms, the content and movements for physical dance teaching

are also seeking changes and innovations. In this paper, a set of three-dimensional

convolutional neural network recognition algorithms based on a machine learning

model is constructed through the collection to recognition of sports dance movement

data. By collecting the skeleton information of typical movements of physical dance, a

typical movement dataset of physical dance is constructed, which is recognized by the

improved 3D convolutional neural network recognition algorithm under the machine

learning model, and the method is validated on the public dataset. The experimental

results show that the 3D CNNs in this paper can produce relatively satisfactory results

for sports dance action recognition with high accuracy of action recognition, which

verifies the feasibility of the 3D convolutional neural network action recognition

algorithm under the machine learning model for the acquisition to recognition of sports

dance actions. It illustrates that the future can be better to open a new direction of

physical dance education content through machine learning models in this form.

KEYWORDS

machine learning model; sports dance movements; DDPG algorithm model; 3D

convolutional neural network movement recognition algorithm; movement skeleton

information dataset

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

268

PAPER INDEX

ABSTRACT

KEYWORDS

1. INTRODUCTION

2. MACHINE LEARNING MODELS

2.1. The DDPG algorithm model used for machine learning

2.2. Observed and rewarded values of machine learning model algorithms

2.3. THREE-DIMENSIONAL CONVOLUTIONAL NEURAL NETWORK

3. METHODS OF RECONSTRUCTING THE CONTENT AND MOVEMENT

IDENTIFICATION OF PHYSICAL DANCE TEACHING

3.1. Physical dance movement data pre-processing

3.2. Experimental data analysis methods

4. EXPERIMENTAL RESULTS AND ANALYSIS

5. CONCLUSION

DATA AVAILABILITY

CONFLICT OF INTEREST

REFERENCES

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

269

1. INTRODUCTION

With the boom of artificial intelligence, academics have begun to explore the

research use of machine learning models in various fields in the hope of reducing

human costs and improving output efficiency through these techniques [1-2]. In the

era of big data, machine learning models that can model relevant variables with

complex relationships will surely become the mainstream of scientific research in the

future [3]. Machine learning models are a top science and technology that specializes

in how computers can simulate or implement human learning behaviors [4]. Machine

learning models are trained to acquire new knowledge or skills and reorganize the

existing knowledge structure to continuously improve their performance [5]. With the

advent of the third wave of artificial intelligence, machine learning models, which are

the core of artificial intelligence, have started to appear frequently in the limelight [6].

Machine learning models are models that make predictions about unknown data

based on known data [7]. Machine learning models usually divide the original data set

into a training set and a test set. The data are then fitted and optimized several times

in the training set to build the model that best reflects the characteristics of the data,

and finally, the performance of the model is evaluated in the test set to verify the

generalization ability and reliability of the model [8-11].

In recent years, machine learning models have been applied to various fields with

remarkable results and their predictive reliability has been widely recognized by

researchers. In the literature [12], two sets of classification models were developed to

predict students' academic performance at graduation using individual course grades

and GPA, respectively. Logistic regression, random forest and plain Bayesian

algorithms in machine learning were used to build prediction models for academic

performance. The literature [13-14] proposed that when machine learning models

were used to predict students' academic performance, the prediction of students'

academic performance was more accurate when combining multiple factors

compared to a single factor. The literature [15] used machine learning methods to

predict changes in major depressive and generalized anxiety disorder symptoms from

pre-treatment to 9-month follow-up. The literature [16] revealed the relationship

between Internet use behaviors and academic performance and used machine

learning models to predict the academic performance of college students with these

behavioral data. The literature [17] studied a review of different machine learning

algorithms and their application in cardiovascular diseases and found that the

application of machine learning can increase the understanding of different types of

heart failure and congenital heart disease. The literature [18] proposed a

multidimensional approach based on GPS measurements and machine learning to

predict injuries in professional soccer. It also provided a simple and practical method

to assess and interpret the complex relationship between sports injury risk and

training performance in professional soccer. The literature [19] analyzed how to

predict and prevent financial losses through public news and historical prices in the

Brazilian stock market through a machine learning model. The literature [20] uses

Recurrent Neural Networks (RNN) for stock market forecasting. RNN is a machine

learning model dedicated to time series, which can take into account past correlation

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

270

series when predicting future trends, and the authors introduced the model to the

analysis of financial time series and tested it on the Nikkei index, proving the

usefulness of the approach. The literature [21] uses oil price movements, gold and

silver price movements, and foreign exchange movements as features to demonstrate

that the KSE-100 index can be predicted by machine learning algorithms and that the

multilayer perceptron outperforms the other algorithms among the machine learning

algorithms used. The literature [22] applied machine learning algorithms such as deep

learning, random forests, neural networks, and support vector regression machines to

stock index prediction in the UK stock market and found that deep learning gave the

best prediction results and support vector machines gave the second best results. The

literature [23] used the XGBoost algorithm for stock index futures forecasting and

compared it with LSTM and traditional autoregressive time series processing

methods, and the empirical results showed that the XGBoost algorithm was superior

in determining the ups and downs. Research on dance moves also abounds, and the

literature [24] proposes a modern deep architecture for C3D that can learn on large-

scale datasets. And the C3D method based on linear classifiers outperforms or

approaches the current state-of-the-art methods in both recognition accuracy of action

videos. The literature [25] uses convolutional neural networks (CNNs) for the

classification of Indian classical dance movements. Two hundred dance poses and

gestures were collected from online videos and offline recordings, respectively, and

the experiments were compared with the results of other classification algorithms on

the same dataset, finally obtaining a recognition rate of 93.33%. In [26], six Greek folk

dance movements were collected using a Kinect sensor, and four common classifiers

were used to directly classify the movements on the raw data to compare the

classification results, and the effect of different body joints on the recognition rate was

also investigated. The literature [27] studied motion data acquisition devices, among

which the Kinect depth vision sensor device has the advantages of high depth map

resolution, low cost, and the ability to directly track the human skeleton motion

trajectory.

In this paper, the typical movement dataset of sports dance was constructed by

collecting the skeleton information of the typical movement of sports dance. This

dataset was identified using a 3D convolutional neural network recognition algorithm

based on a machine learning model, and the original data recorded in this dataset

was checked for missing skeleton point data. In the case that the skeleton point data

was lost in the experiment and could not be recorded, the missing skeleton point data

of a frame was filled with the skeleton data of the previous frame to reduce the

influence of the subsequent machine learning model on the judgment of different

movements. The frame rate of the few skeletal sequences of typical sports dance

movements obtained is 30 fps. The duration of each movement in the dataset is

different, and the corresponding number of frames is also different. The sequence

length of each movement sequence in the dataset is unified, and the maximum

number of frames of the samples in the dataset is obtained. The remaining movement

samples with less than the maximum number of frames are copied to the last frame

and supplemented to the maximum number of frames for subsequent input into the

machine learning model for training Learning. After the dataset is completed, it is input

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

271

into the machine learning model, and the 3D convolutional neural network recognition

algorithm is used to train the dataset. The experimental results show that the 3D

CNNs in this paper have greater recognition advantages and accuracy compared with

several other action recognition algorithms, which can help the innovative

development of physical dance teaching contents and also make another perspective

of physical dance teaching methods. It is verified that the 3D convolutional neural

network action recognition algorithm under the machine learning model is feasible in

the acquisition to recognition of physical dance actions, which can better open a new

direction of physical dance education content through this form of machine learning

model and provide an optional path for the reconstruction of physical dance teaching

content and action recognition.

2. MACHINE LEARNING MODELS

With the development of information technology, intelligent algorithms represented

by reinforcement learning are increasingly used in the field of robot control with their

adaptive characteristics. In recent years, DQN algorithms combining deep neural

networks and reinforcement learning have been proposed to solve the problem of

high-dimensional input to machine learning models. However, DQN is still an

algorithm oriented to discrete control and has insufficient capacity to handle

continuous actions. In the practical control of machine learning models, the angular

output of each joint is a continuous value, and if the range of values taken for each

joint angle is discretized, the number of behaviors grows exponentially with the

number of degrees of freedom. If this accuracy is further improved, the number of

values taken will grow exponentially [28-29].

2.1. THE DDPG ALGORITHM MODEL USED FOR MACHINE

LEARNING

DDPG is an Actor-Critic based algorithm in reinforcement learning, the core idea of

which is to use the Actor-network to generate the behavior policy of the intelligence,

and the Critic network to judge the good and bad actions and guide the updated

direction of the actions.

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

272

Figure 1. DDPG algorithm structure

As shown in Figure 1, the DDPG algorithm structure contains an Actor-network with

parameters and a Critic network with a parameter to compute the deterministic policy

and action-value function, respectively. Since the learning process of a single network

is not stable, the Actor-network and the Critic network are subdivided into a realistic

network and a target network, respectively, drawing on the successful experience of

DQN fixed target networks. The Actor-network and the Critic network are each

subdivided into a realistic network and a target network. The real network and the

target network have the same structure, and the target network parameters are softly

updated by the real network parameters with a certain frequency.

The loss function of the realistic Critic network is:

(1)

Among them:

(2)

is the number of samples, is the weight of the different samples used,

is the action value calculated by the realistic Critic network when

sample j takes action at the state , is the target action value calculated by the

samples and derived from the target Critic network,

is the immediate reward

obtained by sample j for taking action at the state , and γ is the discount factor.

The loss function of a realistic Actor network is :

(3)

Which, finding the minimal value of this loss function by the gradient descent

method is equivalent to the process of maximizing the action value .

The target Critic network and target Actor network parameters are updated in the

following way:

(4)

(5)

is the update coefficient, and the range is 0.01~0.1 to avoid excessive parameter

changes.

When a machine learning model has a high degree of freedom, an additional class

of exploration strategy, maximum a posteriori policy optimization, is needed to

enhance the efficiency of the machine learning model based on the use of the DDPG

algorithm model. The maximum a posteriori policy optimization approach models the

reinforcement learning problem as an inference problem from a probabilistic point of

Q Q 2

( ) ( ( , , ))

j j j j

t t

J y Q S A

θ ω θ

=−

∑

1 1

Q( , , ),

j j

j j j j

t t

R end

yR S A not endγ θ ʹ

+ +

⎧

⎪

⎨+

⎪

⎩

ωj

Q(Sj

t,Aj

t,θQ)

( ) ( , , )

j j

t t

J Q S A

θ θ

=∑

J(θπ)

Q(Sj

t,Aj

t,θQ)

Q Q Q

(1 )θ τθ τ θ

ʹ ʹ

←+−

(1 )

π π π

θ τθ τ θ

ʹ ʹ

←+−

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

273

view. Assuming the probability of completing the task, the probability, according to the

inference problem, is:

(6)

Let the loss function :

(7)

where: is the proposed distribution. the optimization problem can be solved by

the EM method, and the optimal proposed distribution is obtained in step E. The

non-parametric optimization solution in this step is:

(8)

where the optimal temperature term can be optimized according to equation (9):

(9)

In the M-step, the optimal proposal distribution is used to update the neural network

strategy:

(10)

2.2. OBSERVED AND REWARDED VALUES OF MACHINE

LEARNING MODEL ALGORITHMS

To give full play to the performance advantages of the DDPG algorithm, this paper

takes into account the motion state of the machine learning model, the processing

efficiency of the intelligent body, the environment and other factors, and selects some

observations as shown in Table 1.

Table 1. Learning process observations

log ( 1) log ( ) ( 1| )

( )

( ) log ( 1| ) log ( )

p O p p O d

q p O d

π π

τ τ τ

= = =

⎡ ⎤

≥= +

⎢ ⎥

⎣ ⎦

∫

( , ) / ( ( ) || ( ))

q t

J q E r a KL q p

π τ τ

⎡ ⎤

=−

⎢ ⎥

⎣ ⎦

∑

q(τ)

( , )

( | ) ( | , )exp( )

i i

Q s a

q a s a s θ

π θ η

∝

η*

( , )

( ) ( ) log ( | , ) exp( )

Q s a

g s a s dads

η ηε η µπ θ η

= +

∫ ∫

( ) ( | )

max ( , ) max [ [log ( | , )]] log ( )

i S q a s

J q E E a s p

θ θ

θ π θ θ= +

Parameter Symbol

Machine learning model location

Machine learning model pose angle

Machine learning model speed

Contact force

Joint torque output

Last joint moment output

[vx,vy,vz]

[thipl,thipr,tkneel,tkneer,ttirel,ttirel]

[roll,pitch,ya w]

[x,y,z]

[FNl,FNr]

[t′

hipl,t′

hipr,t′

kneel,t′

kneer,t′

tirel,t′

tirel]

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

274

and the following weighted reward values r are designed based on the observed

values, where 1~9 are the weights.

(1) The speed reward value

, which is to encourage the machine learning model

to move forward, is shown in equation (11):

(11)

where: is the velocity of the machine learning model on the axis.

(2) The stability reward value

, lies in rewarding the machine learning model for

completing smooth motions in both instantaneous and global decisions, as shown in

equation (12):

(12)

where: denotes the current position of the machine learning model on the

axis, , denotes the position of the machine learning model on the

axes,

denote the attitude angle (roll angle, pitch angle, yaw

angle) of the machine learning model.

(3) The joint stability bonus value

, which lies in improving the energy utilization

efficiency of the robot, is shown in equation (13):

(13)

where: and represent the output of each joint torque and its torque output at the

previous time.

(4) The value of touchdown reward

, lies in rewarding the machine learning

model for controlling the same contact force between its own two feet and the ground,

reducing the probability of training to generate singular motion poses, as shown in

equation (14):

(14)

where:

denotes the contact forces between the left and right feet of the

machine learning model and the ground, respectively.

(5) The motion duration reward value

, lies in encouraging the machine learning

model to keep moving, as shown in equation (15):

(15)

where is a constant and takes any value.

2.3. THREE-DIMENSIONAL CONVOLUTIONAL NEURAL

NETWORK

Compared with other traditional deep learning methods, 3D convolutional neural

networks are not limited to the input of 2D single-frame images, and can better extract

features from the temporal and spatial dimensions and extract motion information of

1v x

r k v=⋅

2 2 2 2 2

4 5 6

( ) 3( )

s init init

r y y k z z k roll k pitch k yaw=− − − − − − −

y,z

yinit

zinit

y,z

roll,pitch,roolll,pit yaw

( )

js i i

r k t t

=− −

∑

t′

8( ])

l r

F N N

r k F F=− −

FNl,FNr

r k T=

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

275

multiple consecutive frames after reinforcement learning algorithms by machine

learning models.

The three-dimensional deep convolutional neural network used in this paper

includes four convolutional layers, two downsampling layers, two fully connected

layers and one Softmax classification layer, and the downsampling layer uses Max-

pooling with a kernel size of 3×3×3 and a step size of 1, as shown in Figure 2.

Figure 2. Three-dimensional convolutional neural network framework

To capture the sports dance movement information in multiple consecutive frames,

the features are calculated from the spatial and temporal dimensions, the value of the

cell with position coordinates in the th feature map of the th layer is given by

the following equation:

(16)

The time dimension of the 3D convolution kernel is , and the weight value of the

convolution kernel connected to the th feature map at position is .

The ReLU function is the most commonly used activation function in deep machine

learning models, which keeps the original feature value unchanged when the input

feature value is greater than 0, and sets it to 0 when it is less than 0. This is the

unilateral inhibition of the activation function, which allows the model parameters to

become sparse and thus reduces the risk of overfitting to some extent. In addition, the

derivative of the activation function is very simple to compute, which can speed up the

computation to a certain extent, and the derivative is always 1 when the input is

positive, so it can effectively alleviate the problem of gradient disappearance, the

ReLU activation function is defined as:

(17)

Pooling is also known as downsampling, where, unlike the processing of 2D

images, the information of the video in the temporal dimension is taken into account;

(x,y,z)

1 1 1 ( )( )( )

( 1)

0 0 0

i i i

l m n x l y m z n

xyz lmn

ij ij ijr i r

r l m n

V f b v

− − − + + +

−

= = =

⎛ ⎞

= +

⎜ ⎟

⎝ ⎠

∑∑ ∑∑

(l,m,n)

ωlmn

ijr

0,( 0)

( ) max(0, ) , ( 0)

f x x x x

≤

⎧

= = ⎨>

⎩

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

276

by pooling, the feature map is reduced, the dimensionality of the data is reduced, and

the number of calculations is reduced, making it easier to train and improve accuracy:

(18)

The Softmax function is often used in the last layer of a classification task to map

an n-dimensional vector x to a probability distribution such that the probability of the

correct category tends to 1 and the probability of the other categories tends to 0, and

the sum of the probabilities of all categories is 1.

The Dropout strategy is to temporarily discard the neurons in the deep model from

the network and disconnect them according to a certain probability when the model is

trained so that the closed neurons do not participate in the calculation of forward

propagation and the update of weights by the backward gradient. Therefore, Dropout

can be considered as an integration method, i.e., averaging different model

architectures over several iterations, thus significantly reducing the risk of overfitting.

3. METHODS OF RECONSTRUCTING THE CONTENT

AND MOVEMENT IDENTIFICATION OF PHYSICAL

DANCE TEACHING

At present, there are many action recognition methods based on deep learning,

and the commonly used classical action recognition algorithms are the C3D algorithm,

P3D ResNet algorithm, and ConvLSTM algorithm, and these algorithms have a good

effect on action recognition. To realize the recognition of physical dance teaching

content movements, this paper firstly constructs a physical dance movement dataset,

which needs to record the skeleton information of dancers, using the skeleton

information for recognition can be processed faster and reduce the storage space.

Secondly, based on the machine learning model and 3D convolutional neural network,

an improved model is proposed for the recognition of physical dance movements. In

this paper, the Dropout technique is also used to reduce the overfitting phenomenon,

and Dropout is experimentally compared by setting different ratios.

This experiment uses the 5-fold cross-validation method, in which the pre-

processed sports dance typical movement dataset is randomly divided into 5 groups,

of which 4 groups are used as the training set and 1 group is used as the test set, and

then the results of 5 times are averaged as the recognition rate.

3.1. PHYSICAL DANCE MOVEMENT DATA PRE-PROCESSING

To reduce the impact of the subsequent machine learning model on the judgment

of different movements, the missing skeletal data of a certain frame was filled with the

skeletal data of the previous frame to check the missing skeletal data of the recorded

raw data. To reduce the redundant information, only one frame out of every five

frames of the original data was kept, and the maximum number of frames for each

action in the dataset is shown in Table 2, The remaining action samples with less than

1 2 3

, , , ,

0 ,0 ,0

max ( )

x y z x s i y t j z r k

i s j s k s

Vµ

×+×+×+

≤ ≤ ≤ ≤ ≤ ≤

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

277

the maximum number of frames are copied to the last frame to be added to the

maximum number of frames for subsequent input into the machine learning model for

training and learning.

Table 2. Sports Dance Motion Dataset Motion Name and Maximum Number of Frames

3.2. EXPERIMENTAL DATA ANALYSIS METHODS

Using the test set in the dataset, the recognition results of each method for 20

typical sports dance movements are evaluated in terms of accuracy. the recognition

results of the experiments using the 3D convolutional neural network recognition

method under the machine learning model of this paper are calculated, which can

clearly express the categories and numbers of correct and incorrect recognition of

each movement, and then the recognition accuracy of three different types of

recognition methods are calculated, namely, the UTKinect dataset, the MSRAction3D

Dance type Serial

number The action name The maximum

number of frames

Modern dance

1 Next to the T-step step 62

2 Z-step 37

3 Pendulum 60

4 Side-click steps 30

5 Single step 33

Waltz

6 Step back 50

7 Lift the step 40

8 One-handed one-handed step 42

9 Single step 36

Tango

10 Cats wash their faces 25

11 Alternate cover hands 30

12 Flick your fingers 35

Latin dance

13 Flat step 48

14 Drag 31

15 Shrug 40

16 Crankulated arms 38

Samba

17 High shake hands 35

18 Alternate waving hands 32

19 Wave your hands 50

20 Draw circles with both hands 24

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

278

dataset and the TYYDance dataset of this paper under the recognition algorithm.

MSRAction3D dataset and the sports dance action TYYDance dataset of this paper

are then calculated for the overall dance action recognition accuracy under the

recognition algorithm, and finally, the method of this paper is compared with the

classical algorithm, and the corresponding recognition results and conclusions are

drawn.

4. EXPERIMENTAL RESULTS AND ANALYSIS

First, to alleviate the overfitting phenomenon, the Dropout technique was used, and

the Dropout ratios were set to 0.2, 0.4, 0.5, 0.6, and 0.8 and the experiments were

conducted on the dataset of this paper, and the results of the validation set are shown

in Figure 3. After 50 iterations, the recognition accuracy of the Dropout ratio of 0.5 is

slightly higher than that of the ratio of 0.4.

Figure 3. Experimental results of different Dropout ratios

To verify the effectiveness of the 3D convolutional neural network dance movement

recognition method based on the machine learning model proposed in the text,

experiments were conducted on the public dataset UTKinect dataset, MSRAction3D

dataset and the sports dance movement TYYDance dataset of this paper. The

UTKinect dataset contains 10 types of movements with 220 movement samples, the

MSRAction3D dataset contains 20 types of movements with 540 movement samples,

and the dance movement dataset in this paper contains 20 types of dance

movements with 640 movement samples, The dance movement dataset in this paper

contains 20 types of dance movements, with a total of 640 movement samples.

Figure 4 shows the recognition accuracy of the training set and Figure 5 shows the

recognition accuracy of the test set. 82% recognition rate was obtained on the

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

279

UTKinect dataset, 90% recognition rate was obtained on the MSRAciton3D dataset,

and 96% recognition rate was obtained on the TYYDance dataset. The reasons for

the better recognition results are as follows: (1) in the selection process of the typical

movements of sports dance, the dance movements with relatively large differences in

movements are selected; (2) more dance movements are collected. ) more samples

of movements were collected, and more deep learning data were obtained; (3) the

collected dance movements had more duration than the general movements, and

more movement representations were also obtained.

Figure 4. The training recognition rate of three data sets

Figure 5. The test recognition rate of three data sets

In the experimental process, the 3D convolutional neural network action recognition

algorithm under the machine learning model was trained on the training and validation

sets of all subjects, and the loss function curves were obtained as shown in Figure 6.

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

280

In the figure, the vertical coordinate is the loss value, and the horizontal coordinate is

the number of iterations; from Figure 6, we can see that after 45 iterations, the loss

gradually stabilizes at about 0.016, which is the best training effect of the model.

Figure 6. The loss function of the training set and test set

As shown in Figure 7, the graphs of the recognition accuracy functions of the

training and validation sets are obtained experimentally. In the graph, the recognition

rate of the training set reaches 99.74% after 45 iterations of model training, and then

the curve remains stable, reaching the best recognition effect of the model.

Figure 7. The loss function of the training set and test set

Other classical algorithms also use skeleton data input, and compared with the 3D

convolutional neural network action recognition algorithm based on the machine

learning model, the results are shown in Table 3, the model used in this paper has

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

281

higher recognition accuracy in the dataset of this paper, which is slightly higher than

P3D, about 6% higher than ConvLSTM, and about 3.5% higher than C3D(1net).

Table 3. Compare results with other methods

From the experimental results, we can see that using human skeleton information

can occupy less storage space to obtain good recognition results, so each model has

good recognition results in the dataset of this paper. Compared with the C3D

recognition algorithm model, this paper follows the convolutional kernel small and

large 3×3×

3, but there are 8 convolutional layers in C3D, and the data volume is too

large for the data set in this paper, so this paper adjusts the number of layers and the

number of convolutional kernels in the model to obtain better recognition results. This

is one of the reasons why the skeleton information of sports dance movements is

collected in this paper, and the good results obtained from the experiments in this

paper validate the experimental movement collection, as well as the data pre-

processing and the rationality of this data set.

5. CONCLUSION

Each kind of sports dance has its own unique culture and spirit, and dance is a way

for human beings to use body language to express emotions and convey their

feelings, which is a common language for human beings regardless of borders and

race. In this paper, a complete process from dance movement data collection to

movement recognition is realized, and the typical movement data set is constructed

by collecting the typical movement skeleton information of physical dance, excluding

the interference of background, lighting and other factors. The experimental results

show that the 3D CNNs in this paper can produce satisfactory results for sports dance

movement recognition, and validate the feasibility of the 3D convolutional neural

network movement recognition algorithm based on a machine learning model in the

acquisition to recognition of sports dance movements, which can better open up the

content of sports dance education through machine learning model. However, there

are still some deviations in the recognition results of a few movements with high

similarity, and only some typical movement fragments were collected as data samples

for the construction of the dataset, and there is still a lot of research space for the

whole performance. Therefore, the next research direction is to use more optimized

deep learning algorithms to optimize the recognition of sports dance movements,

collect more sports dance movement data to expand the dataset, and to recognize

different dance movements in longer dance performances.

Methods Recognition accuracy

C3D(1 net) 91.27%

P3D ResNet 94.29%

ConvLSTM 90.13%

3D CNNs(ours) 96.91%

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

282

DATA AVAILABILITY

Data for this study are available from the authors upon request.

CONFLICT OF INTEREST

The authors declare that the research was conducted in the absence of any

commercial or financial relationships that could be construed as a potential conflict of

interest.

REFERENCES

(1) Lou, R., Lv, Z., Dang, S., Li, S., Zhang, Y., & Li, L. (2021). Application of

machine learning in ocean data. Multimedia Systems, 1-10.

(2) Guo, J. X., & Huang, C. (2020). Feasible roadmap for CCS retrofit of coal-

based power plants to reduce Chinese carbon emissions by 2050. Applied

Energy, 259, 114112.

(3) Zounemat-Kermani, M., Matta, E., Cominola, A., Giuliani, M., & Castelletti, A.

(2020). Neurocomputing in Surface Water Hydrology and Hydraulics: A

Review of Two Decades Retrospective, Current Status and Future

Prospects. Journal of Hydrology.

(4) Waring, J., Lindvall, C., & Umeton, R. (2020). Automated Machine Learning:

Review of the State-of-the-Art and Opportunities for Healthcare. Artificial

Intelligence in Medicine, 104, 101822.

(5) Bougie, N., & Ichise, R. (2020). Skill-based curiosity for intrinsically

motivated reinforcement learning. Machine Learning, 109(3), 493-512.

(6) Horton, M. B., Brady, C. J., Cavallerano, J., et al. (2020). Practice Guidelines

for Ocular Telehealth-Diabetic Retinopathy, Third Edition. Telemedicine and

e-Health, 26(4).

(7) Shuntaro, C., Kenjirowelq, L., Narin, S., et al. (2021). eSkip-Finder: a machine

learning-based web application and database to identify the optimal

sequences of antisense oligonucleotides for exon skipping. Nucleic Acids

Research.

(8) Mangold, C., Zoretic, S., Thallapureddy, K., et al. (2021). Machine Learning

Models for Predicting Neonatal Mortality: A Systematic Review.

Neonatology, 1-12.

(9) Hindson, J. (2022). Proteomics and machine-learning models for alcohol-

related liver disease biomarkers. Nature Reviews Gastroenterology &

Hepatology, 19(8), 488.

(10) Hasan, M. S., Kordijazi, A., Rohatgi, P. K., et al. (2022). Machine learning

models of the transition from solid to liquid lubricated friction and wear in

aluminum-graphite composites. Tribology International, 165.

(11) Sun, Z. Y., Herold, F., Cai, K. L., et al. (2022). Prediction of Outcomes in Mini-

Basketball Training Program for Preschool Children with Autism Using

Machine Learning Models. International Journal of Mental Health Promotion,

24(2), 143-158.

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

283

(12) Tatar, A. E., & Distegor, D. (2020). Prediction of Academic Performance at

Undergraduate Graduation: Course Grades or Grade Point Average?

Applied Sciences, 10(14), 4967.

(13) Lau, E. T., Sun, L., & Yang, Q. (2019).

Modelling, prediction and classification

of student academic performance using artificial neural networks. SN

Applied Sciences, 1(9), 1-10.

(14) Oyedeji, A. O., Salami, A. M., Folorunsho, O., et al. (2020). Analysis and

prediction of student academic performance using machine learning.

Journal of Information Technology and Computer Engineering, 4(1), 10-5.

(15) Jacobson, N. C., & Nemesure, M. D. (2021). Using artificial intelligence to

predict change in depression and anxiety symptoms in a digital

intervention: Evidence from a transdiagnostic randomized controlled trial.

Psychiatry research, 259, 113618.

(16) Xu, X., Wang, J., Peng, H., et al. (2019). Prediction of academic performance

associated with internet usage behaviors using machine learning

algorithms. Computers in Human Behavior, 98, 166-173.

(17) Mathur, P., Srivastava, S., Xu, X., et al. (2020). Artificial intelligence, machine

learning, and cardiovascular disease. Clinical Medicine Insights: Cardiology,

14, 1179546820927404.

(18) Rossi, A., Pappalardo, L., Cintia, P., et al. (2018). Effective injury forecasting

in soccer with GPS training data and machine learning. PloS one, 13(7),

e0201264.

(19) Duarte, J. J., Gonzalez, S. M., & Jr JC C. (2020). Predicting stock price falls

using news data: Evidence from the Brazilian market. Computational

economics, 57(2), 311-340.

(20) Yoshihara, A., Fujikawa, K., Seki, K., et al. (2014). Predicting stock market

trends by recurrent deep neural networks. In Pacific Rim International

Conference on Artificial Intelligence (pp. 759-769). Springer, Cham.

(21) Usmani, M., Adil, S. H., Raza, K., et al. (2016). Stock market prediction using

machine learning techniques. In 2016 3rd International Conference on

Computer and Information Sciences (ICCOINS) (pp. 322-327). IEEE.

(22) Nikou, M., Mansourfar, G., & Bagherzadeh, J. (2019). Stock price prediction

using deep learning algorithm and its comparison with machine learning

algorithms. Intelligent Systems in Accounting, Finance and Management, 26(4),

164-174.

(23) Hah, D. W., Kim, Y. M., & Ahn, J. J. (2019). A study on KOSPI 200 direction

forecasting using XGBoost model. The Korean Data & Information Science

Society, 30(3), 655-669.

(24) Tran, D., Bourdev, L., Fergus, R., et al. (2015). Learning spatiotemporal

features with 3D convolutional networks. In Proc of IEEE International

Conference on Computer Vision (pp. 4489-4497). Washington DC: IEEE

Computer Society.

(25) Kishore, P. V. V., Kumar, K. V. V., Kiran Kumar, E., et al. (2018). Indian classical

dance action identification and classification with convolutional neural

networks. Advances in Multimedia, 2018.

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

284

(26) Protopapadakis, E., Grammatikopoulos, A., Doulamis, A., et al. (2017). Folk

dance pattern recognition over depth images acquired via Kinect sensor.

The International Archives of Photogrammetry, Remote Sens.

(27) Li, G., & Li, C. (2020). Learning skeleton information for human action

analysis using Kinect. Signal Processing: Image Communication, 84, 115814.

(28) Frayssinet, M., Esenarro, D., Juárez, F. F., & Díaz, M. (2021). Methodology

based on the NIST cybersecurity framework as a proposal for

cybersecurity management in government organizations. 3C TIC.

Cuadernos de desarrollo aplicados a las TIC, 10(2), 123-141. https://doi.org/

10.17993/3ctic.2021.102.123-141

(29) Chen, S., & Ren, Y. (2021). Small amplitude periodic solution of Hopf

Bifurcation Theorem for fractional differential equations of balance point in

group competitive martial arts. Applied Mathematics and Nonlinear Sciences,

7(1), 207-214. https://doi.org/10.2478/AMNS.2021.2.00152

https://doi.org/10.17993/3ctic.2023.121.267-285

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

285