Application of action and posture feature information matching algorithm in college aerobics

Abstract

With the increasingly rich recreational activities of college students, diversified learning needs and complex physical education resources bring challenges to college physical education. In order to optimize the teaching effect of calisthenics in colleges and universities, this paper proposes a matching method of posture features based on dynamic time warping. Firstly, the dynamic time warping algorithm is introduced, and then the matching model of posture features of calisthenics is constructed on this basis. Finally, the application effect of the model is tested and analyzed. The results show that the model can capture the video frame accurately, and its matching accuracy reaches 94.8%, which greatly improves the accuracy of aerobics action recognition. Good posture matching effect is conducive to teachers to obtain a clear learning situation of students, and provide a reference for adjusting the teaching progress and teaching methods of calisthenics. Under the teaching mode of this model, the average professional score of the students in calisthenics reaches 85 points, which is 25 points higher than that under the convolutional neural network model. It also proves the validity and feasibility of this method in the course of calisthenics in colleges and universities, which is beneficial to enhance the physical quality of college students and enrich the content of calisthenics teaching.

Keywords

Posture feature dynamic time regulation aerobics intelligent sports colleges and universities

1. Introduction

Dynamic Time Warping (DTW) is a time series comparison algorithm, which is often used to measure the similarity of time series. For the data points that have not yet been corresponding, select another time series corresponding to the time series where the time series is located, and the two time series meet the conditions corresponding to the start and end points, calculate the distance between the time series, and use it as the basis for measuring the similarity [1, 2]. The DTW algorithm makes up for the lack of Euclidean distance, solves the difficulty of similarity calculation caused by the inconsistent length of time series, reduces the amount of calculation, improves the simplicity of operation, and can process online data at the same time. The problem of data mining provides technical support for data mining in the era of big data [3, 4].

With the development and wide application of Internet technology, intelligent sports have become the development trend of modern sports teaching. Among them, the matching method of action posture characteristics information is an important part of intelligent sports teaching. The matching of action gesture feature information can judge students’ action standard through action gesture recognition, which is conducive to teachers’ accurate grasp of students’ sports learning effect. On this basis, teachers can adjust the teaching plan and promote the effectiveness of sports teaching interaction between teachers and students. However, in the actual application process, due to the environment, equipment and other factors, there are often problems such as low motion attitude matching efficiency and large matching error [5, 6]. For this reason, the research puts forward a method of action posture feature information matching based on DTW algorithm, so as to improve the accuracy of aerobics teaching and promote the intelligent and information-based development of college physical education teaching.

2. Related work

DTW is an algorithm for synchronizing time series data, which is widely used in various fields of society. Switonski et al. [7] aimed to study the application of DTW algorithm in motion capture data classification. In view of the limitations of the average rotation method based on DTW algorithm, they proposed to use differential filtering to solve the problem of rotation speed in the case of unit quaternion. The results show that this method reduces the pose signature and improves the classification effect of the captured data, which proves the effectiveness of this method. Hollerbach et al. [8] expounded the problems of low resolution in the separation of ion mobility spectra, and proposed to use the DTW algorithm to correct the separation changes of ion mobility spectra, and carried out experiments using single-value calibration as a comparison method. The results show that the separation effect of the two methods is similar for the narrow migration range, and the DTW algorithm is better for the separation effect of the wide migration range. Xu et al. [9] analyzed the advantages and disadvantages of Raman spectroscopy. In order to solve the problem that the spectral library is difficult to share due to the nonlinear shift of Raman, a DTW method with variable penalty was proposed to standardize the Raman spectrum, thereby reducing the time cost of building the spectral library. The experimental results show that the spectral similarity under this method reaches 0.97, which is far better than the MWFFT method. Diab et al. [10] explained the importance of security early warning systems in computer networks, and analyzed and studied the detection technology of computer attacks. identification and detection. Experiments show that this method improves the accuracy of recognition and provides a new idea for improving network security. Chen and Gu [11] proposed a mixed-signal DTW acceleration method in order to increase the computational complexity of the DTW algorithm and improve the effect of time series classification. The experimental results show that, compared with the traditional classification method, this method increases the computing throughput to more than 9 times, greatly enhances the scalability of the DTW algorithm, optimizes the performance of the DTW algorithm, and improves the operability of the algorithm. Aiming at the optimization of visual position recognition, Alqaraleh et al. [12] proposed a recognition method based on DTW algorithm and convolutional neural network for the problem of poor visual position recognition effect in the case of appearance and pilot changes. Experiments show that this method greatly improves the accuracy of visual location recognition and optimizes the performance of visual location recognition.

Bodybuilding is one of the sports to improve physical fitness and is an important part of physical education in colleges and universities. In order to improve the accuracy of key frame extraction in aerobics teaching video, Yan and Woniak [13] proposed an HSV space method, which segmented the teaching video by one-dimensional feature vector, constructed dynamic frames of aerobics teaching video, and extracted the video’s dynamic frame. Feature vector. The results show that the accuracy of this method is as high as 99%, which significantly improves the extraction accuracy. Based on the background of the “Internet $+$ ” era, Shi et al. [14] analyzed and studied the optimization of physical education teaching in colleges and universities, and proposed to integrate computer-aided methods and mobile intelligent terminals into physical education teaching in colleges and universities. The experimental results show that this teaching method improves the efficiency of physical education, stimulates students’ enthusiasm for physical education, improves students’ physical education performance, and promotes the reform of physical education teaching mode. Agans et al. [15] aimed to study the correlation between health courses and college students’ physical activity behaviors and attitudes, taking 1473 college students from a university as the research object, and using one-way analysis of variance to understand the differences in the characteristics of students before and after the course learning, the results show that the curriculum does not have a significant impact on the behavior and attitude of college students’ sports activities. In order to improve the quality of sports mixed teaching, Bao and Yu [16] proposed a teaching quality evaluation method combined with mobile edge computing, used class analysis to classify the importance of each index, and established a fuzzy comprehensive evaluation analysis model. The result shows that this method reduces the error of teaching quality evaluation and improves the evaluation efficiency. In order to improve the work performance of physical education teachers in colleges and universities, Muhammad et al. [17] conducted an experimental analysis on the influencing factors including professional attitude, professional pride and professional autonomy. To understand teachers’ teaching psychology through questionnaires, use descriptive statistics and inferential statistics to verify the results. The results show that three influencing factors have a significant impact on teachers’ work performance. Prieto-González [18] studied the effect of physical education on college students’ physical quality in order to optimize the physical education curriculum in colleges and universities and promote the development of physical education in colleges and universities. By adjusting the physical education class cycle, observe the changes in students’ physical quality. The results show that two physical education classes per week can effectively improve the strength and agility of college students, and achieve the purpose of strengthening the body.

It can be seen from the above research results that the DTW algorithm plays an important role in speech recognition, data mining and information retrieval due to its advantages; there are problems such as low student enthusiasm and poor teacher performance in college physical education teaching. It can be seen that the research on intelligent teaching of physical education in colleges and universities is not sufficient, especially the research on the combination of DTW algorithm and college aerobics is still rare. For this purpose, this paper proposes a matching method of aerobics movement and posture characteristics based on DTW algorithm, in order to achieve accurate data analysis and evaluation of physical education teaching, improve the level of physical education teaching in colleges and universities, and create a modern physical education classroom.

3. Aerobics action pose feature matching method based on DTW algorithm

3.1 Dynamic time rounding algorithm

In view of the different lengths of time series, the DTW algorithm adopts the “one-to-many” processing method, and by determining the relationship between each data feature, the minimum cumulative distance value is obtained, which is the best alignment scheme, so as to realize the time series in the time series. correspondence of data points. DTW algorithm uses the concept of dynamic programming to decompose complex multi-variable problems into multiple simple single-variable problems, which reduces the time to search for the optimal distance and improves the efficiency of decision-making problem solving. It is a typical optimization method [19]. Assume that the action pose feature sequence to be matched has a $A=\{{a_{1},a_{2},\ldots,a_{m}}\}$ length of $m$ , where the coordinates are $i$ , which are $p_{im}$ composed $q_{jn}$ of a differential pose feature vector, that is $a_{i}=\{{p_{i1},p_{i2},\ldots,p_{im}}\}$ , the standard action pose feature sequence $B=\{{b_{1},b_{2},\ldots,b_{n}}\}$ , whose length is $n$ , $B$ and the coordinates in the sequence $j$ are vector, ie $b_{n}=\{{q_{j1},q_{j2},\ldots,q_{jn}}\}$ . The action pose score is used as an index to measure the sequence $A$ and $B$ similarity, and its calculation is shown in Eq. (1).

$\displaystyle r(i,j)=\frac{1}{1+d(i,j)}\times 100$ (1)

In Eq. (1), it $r(i,j)$ represents the score of the action pose feature to be matched compared to the standard action pose feature, and $d(i,j)$ represents the distance between the sequence $A$ and $B$ the middle action feature, and its calculation is shown in Eq. (2).

$\displaystyle d(i,j)=\sqrt{(p_{i1}-q_{j1})^{2}+(p_{i2}-q_{j2})^{2}+\ldots+(p_{% im}-q_{jn})^{2}}$ (2)

In Eq. (2), it $d(i,j)$ represents the Euclidean distance. Since the Euclidean distance cannot solve the similarity problem of different sequences when facing sequences of different lengths, it is necessary to use a regular path to convert the Euclidean distance into a score. Based on the difference between the action pose and the standard action pose, the similarity is measured by the action score. The comparison of the changes of the two sequences before and after using the regular path is shown in Fig. 1.

Figure 1.

Comparison of the changes of the two sequences before and after using the regular path.

As can be seen from Fig. 1, under the action of the regular path, the action features between the sequence $A$ and $B$ the middle are optimally corresponding. The core of the regular path is the regular path distance, which matches the similar feature information between the two sequences, and calculate the distance between all similar points, and the sum of the distances is the regular path distance. There is a negative correlation between the regular path distance and the similarity, and the larger the regular path distance, the lower the similarity.

Set the regularized path as $C=\{{c_{1},c_{2},\ldots,c_{k}}\}$ , $c_{k}=(i,j)$ and the number of its elements $k$ as:

$\displaystyle\left\{{\begin{array}[]{l}c_{1}=(1,1)\\ c_{k}=(m,n)\\ \end{array}}\right.$ (3)

In Eq. (3), $c_{1}$ and $c_{k}$ represent $C$ the starting point and end point of the regular path, respectively, to ensure that the regular path $C$ covers all coordinates in the time series $A$ sum $B$ . The matrix is composed of action and posture score $r_{ij}$ elements $R$ , the size of which is $m\times n$ , the shortest path search is performed in the matrix $R$ , and the single correspondence principle should be followed in the search process, that is, the two sequences of feature information should be corresponded in the order from front to back, and at the same time, the corresponding Omissions, to achieve one-to-one correspondence. The calculation of the regular path distance is shown in Eq. (4).

$\displaystyle R(i,j)=r(i,j)+\max[R(i-1),R(i,j-1),R(i-1,j-1)]$ (4)

In Eq. (4), $R(i-1)$ , $R(i,j-1)$ and $R(i-1,j-1)$ represent three search directions. For the wrong actions in the sequence $A$ , it is necessary to match and mark the non-standard features. The standard degree of the action posture is judged mainly through the feature of the center of gravity vector and the feature of the angle between the limbs. The calculation of the centroid vector error degree between the two sequences is shown in Eq. (5).

$\displaystyle T=(a_{i1}-b_{j1})*\frac{180}{\pi}$ (5)

In Eq. (5), $a_{i1}$ and $b_{j1}$ represents $A$ the action pose feature vector of the sequence sum. $B$ The limb angle feature $b_{j}^{v}=(v_{j1},v_{j2},\ldots,v_{jn})$ of the sequence $A$ , $B$ the wrong limb angle feature in the sequence, and the difference between the limb angle features $a_{i}^{v}=(v_{i1},v_{i2},\ldots,v_{im})$ of the two sequences is the limb angle deviation $\textit{Dif}_{il}$ , $\textit{Dif}_{il}=(\textit{difi}_{1},\textit{difi}_{2},\ldots,\textit{difi}_{l})$ . Set the standard limb angle, which is calculated as shown in Eq. (6).

$\displaystyle D=\frac{S}{100}$ (6)

In Eq. (6), the $S$ standard score is represented. Using the feature of the angle between the limbs, the deviation factor of the limb angle in the sequence $A$ is obtained by calculation, and it is used as an index to measure the deviation of the action, and its operation is shown in Eq. (7).

$\displaystyle D_{is}=1-\frac{\textit{dif}_{is}}{y_{js}}$ (7)

In Eq. (7), $s$ represents the number of limb angle features, and $D_{is}$ represents the deviation factor. $D_{is}$ The larger the value, the better the action. Compare the deviation factor $D_{is}$ with the standard limb angle $D$ . If it $D_{is}$ is greater than $D$ that, it means that the deviation of the action pose feature is unqualified, and the $i$ value and $s$ value need to be output and marked. After matching the wrong actions, $A$ all action poses in all sequences need to be calculated to obtain the final score of the action pose features to be matched, and the calculation is shown in Eq. (8).

$\displaystyle G=\frac{R(i,j)}{m}$ (8)

Equation (8), it represents $G$ the score of the sequence. $A$ By comparing $G$ with the standard score $S$ , it is judged whether the action posture meets the standard. Only when it $G$ is greater than the standard score $S$ , the action is standardized. The steps of the DTW algorithm are shown in Fig. 2.

Figure 2.

DTW algorithm steps.

As can be seen from Fig. 2, the regular path distance is the focus of the DTW algorithm. On the basis of matching the action pose sequence and the standard action pose sequence, a cumulative distance matrix is constructed. After filling all the elements of the matrix, the two sequences are searched according to the principle of closest distance. The corresponding relationship between data points, the path with the smallest cumulative distance is taken as the optimal path to complete the optimal correspondence of action features between sequences.

3.2 Aerobics action pose feature matching model based on DTW algorithm

DTW algorithm can improve the efficiency of local feature matching between time series, reduce the difficulty of similarity measurement caused by different sequence lengths, and effectively deal with the phase shift problem of time series in complex environments. Amplitude changes and other situations have strong robustness, and optimize the effect of time series similarity measurement [20]. Incorporating the DTW algorithm into the aerobics courses in colleges and universities is conducive to understanding the learning situation of students and laying a foundation for the targeted teaching of aerobics. After the teacher’s professional action teaching, the students practice and shoot aerobics action videos. The teacher uses the DTW algorithm to match the students’ aerobics movements and postures, judge the students’ learning effect, and adjust the teaching plan based on this. The basis of its realization is computer technology, and relevant systems are developed through a software development platform, which has strong operability. The object of the DTW aerobics action and posture feature matching model is the students’ aerobics action videos. Due to the differences in the students’ shooting environment, pixels, equipment and other factors, the collected action and posture information to be matched has a large amount of data, complex and diverse and clear. In order to reduce the matching error caused by the data, the collected data needs to be standardized to achieve data unification and normalization. The calculation is shown in Eq. (9).

$\displaystyle A^{\prime}=\frac{A-A_{\min}}{A_{\max}-A_{\min}}$ (9)

In Eq. (9), $A_{\max}$ and $A_{\min}$ represent the maximum and minimum values of the sample data, respectively. Use OpenPose to track the body’s limbs, extract the main body posture features, including limb angle features, center of gravity vector angle features, and limb orientation features. On this basis, the features are fused to form body posture feature information. According to the limb angle feature, according to the human body structure, the angle between the limb block and the spine is calculated, and based on this, the same limb angle is eliminated, the matching deviation caused by the same data is reduced, and the difference of similar posture features is enhanced. Set the skeleton point as $E(q_{1},q_{2},\ldots,q_{m})$ , where $q_{m}$ is the limb block vector, and the center point of the limb block is $q_{0}$ , and its calculation is shown in Eq. (10).

$\displaystyle q_{0}=(x_{0},y_{0})=\left(\frac{x_{1}+x_{2}+\ldots+x_{m}}{m},% \frac{y_{1}+y_{2}+\ldots+y_{m}}{m}\right)$ (10)

In Eq. (10), $(x_{0},y_{0})$ the coordinates of the center block are represented. For the bias feature of the center point of the limb block, it needs to be solved by the cosine theorem. Its calculation is shown in Eq. (11).

$\displaystyle\theta=\arccos\left(\frac{(x_{0}x_{m}+y_{0}y_{m})}{\sqrt{(x_{0}^{% 2}+y{}_{0}^{2})\times(x_{m}^{2}+y_{m}^{2})}}\right)$ (11)

In Eq. (11), $x_{m}$ and $y_{m}$ represent the spine vector coordinates, respectively. For the angle feature of the center of gravity vector, the core is the coordinate of the joint point, and the angle between the direction of the center of gravity and the spine is used to obtain the posture bias information of the human body. The horizontal and vertical coordinates of the joint point are respectively the $X_{i}$ sum $Y_{i}$ . In the video image, the joint coordinates of each frame need to be obtained, and the calculation is shown in Eq. (12).

$\displaystyle\left\{{\begin{array}[]{l}X=\sum\limits_{i=1}^{m}{k_{i}}\cdot X_{% i}\\ Y=\sum\limits_{i=1}^{m}{k_{i}}\cdot Y_{i}\\ \end{array}}\right.$ (12)

Equation (12), $k_{i}$ the coefficient of the th joint point is expressed. $i$ The coordinates of the center of gravity of the human body are obtained by the $X$ sum $Y$ of the coordinates, ie $(X,Y)$ . By calculating the coordinates of the center of gravity of the human body, the center of gravity bias information is obtained, and its expression is shown in Eq. (13).

$\displaystyle\theta_{g}=\arccos\left(\frac{\mathord{\buildrel\lower 3.0pt\hbox% {$\scriptscriptstyle\rightharpoonup$}\over{G}}\cdot\mathord{\buildrel\lower 3.% 0pt\hbox{$\scriptscriptstyle\rightharpoonup$}\over{M}}}{|\mathord{\buildrel% \lower 3.0pt\hbox{$\scriptscriptstyle\rightharpoonup$}\over{G}}|\cdot|\mathord% {\buildrel\lower 3.0pt\hbox{$\scriptscriptstyle\rightharpoonup$}\over{M}}|}\right)$ (13)

In Eq. (13), $\mathord{\buildrel\lower 3.0pt\hbox{$\scriptscriptstyle\rightharpoonup$}\over{% G}}$ and $\mathord{\buildrel\lower 3.0pt\hbox{$\scriptscriptstyle\rightharpoonup$}\over{% M}}$ represent the center of gravity vector and the spine vector, respectively, $\theta_{g}$ representing the angle between them. For the limb orientation feature, it is necessary to use the front and back orientation features and the left and right orientation features to construct the feature. Let the positions of the left and right joints be sum, respectively $(h_{1},o_{1})$ , $(h_{2},o_{2})$ and the position coordinates of the heel and the big toe are $(h_{3},o_{3})$ sum $(h_{4},o_{4})$ , respectively. The calculation of the limb orientation feature vector is as shown in Eq. (14):

$\displaystyle\left\{{\begin{array}[]{l}f=\frac{h_{2}-h_{1}}{|h_{2}-h_{1}|}\\ \theta_{1}=\arctan\left(\frac{o}{h}\right)\\ \end{array}}\right.$ (14)

In Eq. (14), it $f$ represents the direction vector, and when the $f$ value is 1 and $-$ 1, it represents the forward and backward features of the human body, respectively. $\theta_{1}$ It represents the angular feature of the direction of the human foot. In order to more intuitively understand the features of human motion and posture, the extracted action and posture feature vectors need to be fused, and the dimensions of the three kinds of human posture feature vectors are unified by splicing, which is expressed as Eq. (15).

$\displaystyle F=[{V,T,U}]$ (15)

In Eq. (15), $F$ represents the human gesture recognition vector, $V$ , $T$ and $U$ represent the body angle feature set, the center of gravity vector angle feature set, and the limb orientation feature set, respectively. The OpenPose action pose feature extraction process is shown in Fig. 3.

Figure 3.

OpenPose gesture feature extraction process.

As can be seen from Fig. 3, the extraction of action pose features by OpenPose is mainly divided into three steps: extraction of coordinate points, classification and integration. For the collected image information, OpenPose first obtains the skeleton point coordinates of the human body posture, and uses the key coordinate points as the origin to construct the human body space coordinate system. Based on the human body coordinate data in the image, the human skeleton coordinate system is constructed through the mapping operation. From this space, the human body pose features are extracted and saved, and after all the pose features are obtained, the labels are integrated and the final human action pose information is obtained. The structure of the DTW aerobics action and posture feature matching model is shown in Fig. 4.

Figure 4.

Structure of DTW aerobics posture feature matching model.

In Fig. 4, the DTW aerobics action and posture feature matching model is composed of a data preprocessing module, a posture feature extraction module and a posture matching module. After standardizing the data of students’ aerobics action videos, OpenPose is used to extract body angle features, gravity vector angle features and body orientation features, which provide reliable data support for accurately judging students’ movements. The action feature sequence is aligned with the teacher’s standard action sequence to realize the matching between the two sequences, and complete the judgment on the standardness of the students’ aerobics actions.

4. Application of aerobics motion posture feature matching model based on DTW algorithm

4.1 Performance analysis of DTW algorithm

The A and B datasets are used for model training and testing, and the two datasets are derived from the MPII Human Pose dataset and the PoseTrack dataset, respectively. In both datasets, 80% are randomly selected as the training set, 10% as the test set, and 10% as the validation set. Accuracy, recall, and area under the ROC curve (Area Under Curve, AUC) are used as the criteria for evaluating model performance, and support vector machine (SVM), K-Nearest neighbor (K-Nearest Neighbor, KNN) and convolution are added at the same time. The neural network (Convolutional Neural Network, CNN) three algorithm models are used as experimental comparisons. The experimental environment and parameters were set before the experiment, as shown in Table 1.

Table 1
Experimental environment and parameter setting

/	Parameter name	Parameter value
Experimental environment parameters	Operating system	Windows10 64bit
	Application framework	Qt5.9.0
	Memory	DDR4 2400MHz 8GB
	Software program	Visual Studio 2015
	CPU	GTX1660
OpenPose parameter	Frame rate	30fps
	Resolving power	480p
	Software library	OpenCV

It can be seen from Table 1 that the experimental environment is the operating system Windows10 64bit, the memory is DDR4 2400MHz 8GB, the CPU model is GTX1660, and the software platforms include Qt5.9.0, Visual Studio 2015 and OpenCV. The video file has a resolution of 480p and a frame rate of 30fps. The comparison of matching accuracy under different algorithm models is shown in Fig. 5.

Figure 5.

Comparison of matching accuracy under different algorithm models.

Figure 5 shows that the matching accuracy of the DTW model in the two datasets is significantly different from other models. In the training of datasets A and B, the accuracy rates are 94.8% and 93.4%, respectively. Among them, the matching accuracy of the station action in the A dataset is the highest, reaching 98.75%. The matching effects of the CNN model in the two datasets are similar, with 31 and 38 incorrect action pose matching errors, and the accuracy rates are 91.1% and 89.1%, respectively. The matching effect of the KNN model is poor. The matching error rates in datasets A and B are 90.5% and 88%, respectively, which are 4.3% and 5.4% lower than that of the DTW model, respectively. As a result, the DTW model reduces the error of action pose matching, improves the accuracy of model matching, and improves the model’s ability to measure the similarity of action sequences. The comparison of recall rates under different algorithm models is shown in Fig. 6.

Figure 6.

Comparison of matching recall under different algorithm models.

It can be seen from Fig. 6 that in the training of datasets A and B, the recall rate of the DTW model is kept in the interval [0.6, 0.68] and [0.67, 0.75] respectively, and the recall rate changes at a small speed and is relatively stable. The recall rates of the CNN models in the two datasets are kept in the range of [0.4, 0.58] and [0.45, 0.6], respectively, and the recall rates fluctuate greatly with the increase of the number of samples. The training effect of the KNN model in data set B is better than that in data set A, and its recall rate is up to 0.65, which is 0.12 higher than the highest recall rate in data set A. The training effect of the SVM model in dataset A is relatively stable, and its recall rate basically remains at 0.46. As a result, the DTW model improves the model’s ability to distinguish between positive and false actions, and achieves accurate matching of various action poses. The comparison of ROC curves under different algorithm models is shown in Fig. 7.

Figure 7.

Comparison of ROC curves under different algorithm models.

In Fig. 7, in the training and testing of dataset A, the AUC value of the DTW model is 0.865, and its sensitivity is the highest among the four models, at 0.72. The AUC value of the SVM model is 0.805, and the training effect is second only to the DTW model. The AUC values of the CNN model and the KNN model are 0.683 and 0.558, respectively, and the KNN model has the lowest sensitivity at 0.42. In the training and testing of dataset B, the AUC value of the DTW model reached 0.812, and the AUC values of the SVM model, CNN model and KNN model were 0.632, 0.597 and 0.694, respectively. The training effect of the KNN model was much better than that of the dataset A. It can be seen that the AUC values of the DTW model are the highest in both datasets, which indicates that the DTW model has improved the sensitivity to recognition of various actions and gestures, enhanced the flexibility of model matching, and optimized the model performance.

4.2 Model matching results

Select 150 students of aerobics major in a university as the model application object, and integrate the DTW aerobics action and posture feature matching model into the aerobics teaching of the school. The model is used to match the students’ bodybuilding operations. After a semester of teaching, the professional skills of aerobics mastered by the students are tested, and the average matching accuracy and the students’ average professional grades are used as the test of the model application effect. Also add SVM model, KNN model and CNN model as model application comparison. The comparison of the average matching accuracy of action pose features under different models is shown in Fig. 8.

Figure 8.

Comparison of average matching accuracy of motion and posture features under different models.

It can be seen from Fig. 8 that the average matching accuracy of the DTW model remains in the [0.75, 0.8] interval, and the matching accuracy is relatively stable for different action poses. Among them, the accuracy of lunges is the lowest, and the accuracy of head wrapping is the highest. The average matching accuracy of the SVM model remains in the range of [0.6, 0.7], and the matching objects with the highest and lowest accuracy are raising arms and touching the ground, respectively. The average accuracy of the CNN model varies greatly, and the matching accuracy of the shoulder-lifting action is the highest, reaching 0.68. The action pose matching effect of the KNN model is the worst, and its average accuracy is concentrated at the 0.5 level. As a result, the DTW model can perform high-precision matching of various aerobics postures, improve the stability of the model, and provide a reliable data reference for teaching. The comparison of the average aerobics professional performance of students under different models is shown in Fig. 9.

Figure 9.

Comparison of students’ average aerobics professional scores under different models.

In Fig. 9, after applying the DTW model, the average professional score of students in aerobics is about 85 points, which is the highest among several teaching modes; followed by the SVM model teaching mode, whose students’ average professional score reaches 75 points, KNN and CNN models. Under the teaching mode, the average professional scores of the students are about 70 and 60 points, respectively, which are 15 and 25 points lower than that of the students under the DTW model. The worst teaching effect is the traditional aerobics physical education teaching mode, whose average professional score is about 40 points. It can be seen that the combination of the action and posture feature information matching algorithm and aerobics teaching can significantly improve the teaching effect of aerobics and enhance the physical quality of students. Among them, the DTW model has the best teaching effect, the highest teaching quality, and the fastest improvement of students’ professional performance.

5. Conclusion

Aerobics is a kind of aerobic exercise with both exercise and entertainment. Among them, mass aerobics has become a popular sports item due to its low difficulty and strong flexibility. In order to optimize the teaching effect of aerobics in colleges and universities, a DTW aerobics action and posture feature matching model was proposed, and the performance of the model was tested and analyzed. In the model training and testing of datasets A and B, the matching accuracy of the DTW algorithm model is 94.8% and 93.4%, respectively, and the recall rate is kept within the interval of [0.6, 0.68] and [0.67, 0.75]. The change is small and relatively stable, and the AUC values reach 0.8 65 and 0.812, respectively, which are 0.182 and 0.215 higher than the CNN model. In the model application of a college aerobics course as an example, the average matching accuracy of the DTW model remains in the range of [0.75, 0.8], and the average professional score of students under its teaching mode reaches 85 points, compared with traditional teaching methods. Mode improved by 45 points. As a result, the DTW aerobics action and posture feature matching model reduces the error of posture feature matching, improves the application value of DTW intelligent algorithm in college aerobics teaching, and is conducive to improving students’ enthusiasm for sports learning. Experience the joy of participating in sports. Although the research has achieved certain results, there is a limitation of the small amount of experimental data. In future research, it is necessary to collect more data samples for experimental exploration, and to build a more complete aerobics work posture feature information matching model for Provide technical support for university aerobics teaching.

References

Pang

Tian

Tang

. Fault state recognition of wind turbine gearbox based on generalized multi-scale dynamic time warping. Struct Health Monit. 2021; 20(6): 3007-3023.

Thuy

Anh

Chau

. Efficient segmentation-based methods for anomaly detection in static and streaming time series under dynamic time warping. J Intell Inf Syst. 2021; 56(3): 121-146.

Baek

Alhindi

Jeong

Chung

. Real-time fire detection algorithm based on support vector machine with dynamic time warping kernel function. Fire Technol. 2021; 57(2): 2929-2953.

Mohammadzade

Hosseini

Dastjerdehei

MRR

Tabejamaat

. Dynamic time warping-based features with class-specific joint importance maps for action recognition using Kinect depth sensor. IEEE Sens J. 2021; 21(7): 9300-9313.

Men

. Intelligent sports prediction analysis system based on improved Gaussian fuzzy algorithm. Alexandria Eng J. 2022; 61(7): 5351-5359.

Ding

Cheng

. Application of Internet of Things and virtual reality technology in college physical education. IEEE Access. 2020; 8: 96065-96074.

Switonski

Josinski

Wojciechowski

. Dynamic time warping in classification and selection of motion capture data. Multidimensional Syst Signal Process. 2019; 30(6): 1437-1468.

Hollerbach

Conant

Nagy

Monroe

Ibrahim

. Dynamic time-warping correction for shifts in ultrahigh resolving power ion mobility spectrometry and structures for lossless ion manipulations. J Am Soc Mass Spectrom. 2021; 32(4): 996-1007.

Chen

Zeng

Zhang

. Standardization of Raman spectra using variable penalty dynamic time warping. Anal Methods. 2021; 13(30): 3414-3423.

10.

Diab

Assadhan

Binsalleeh

Lambotharan

Address

. Denial of service detection using dynamic time warping. Int J Network Manage. 2021; 31(6): 193-198.

11.

Chen

. High-throughput dynamic time warping accelerator for time-series classification with pipelined mixed-signal time-domain computing. IEEE J Solid-State Circ. 2020; 56(2): 624-635.

12.

Alqaraleh

Hafez

Tello

. Dynamic time warping of deep features for place recognition in visually varying conditions. Arabian J Sci Eng. 2021; 46(4): 3675-3689.

13.

Yan

Woniak

. Accurate key frame extraction algorithm of video action for aerobics online teaching. Mobile Networks Appl. 2022; 27(3): 1252-1261.

14.

Shi

. The application of computer-aided teaching and mobile Internet terminal in college physical education. Comput-Aided Des Appl. 2021; 18(S4): 163-174.

15.

Agans

Wilson

Bopp

. Required health and wellness courses: associations with college student physical activity behavior and attitudes. J Phys Activity Health. 2020; 17(6): 632-640.

16.

Bao

. Evaluation method of online and offline hybrid teaching quality of physical education based on mobile edge computing. Mobile Networks Appl. 2021; 26(5): 2188-2198.

17.

Muhammad

Khan

. Effect of professional determinants upon job performance of lecturers in health and physical education at college level. Sir Syed J Edu Soc Res. 2020; 3(4): 342-350.

18.

Prieto-González

Sedlacek

. Physical education: Practice, cessation and resumption in untrained Saudi college students UDC 796.371.212 (532). Facta Univ Ser Phys Edu Sport. 2020; 18(3): 681-692.

19.

George

Silversides

Zigman

Melkumyan

. Bedding angle identification from BIF marker shales via modified dynamic time warping. Math Geosci. 2021; 53(7): 1567-1585.

20.