Research on pattern recognition and sports performance of table tennis game strategy based on big data mining technology

With the rapid development of computer technology and network technology, people have ushered in a new era of networks characterized mainly by information and knowledge. People’s access to information, especially digitized information, is growing exponentially. The continuous strong increase in the amount of information brings people not only convenience, but also serious challenges. Because of the dramatic increase in information, the dynamic balance between information supply and information consumption has been broken, and the gap between the two is getting wider and wider [1-3]. In the mountain of information contains many useful knowledge to be extracted, these knowledge is like a mature crop, if not harvested in time will be wasted. So, how to “timely harvest mature crops” from the information ocean? There is an urgent need for a new generation of technological methods and tools to help mine the treasures in the mountain of information and refine them into useful knowledge [4-6]. As a result, a new research field, data mining (DM), has emerged. With the gradual deepening of knowledge and understanding of data mining technology and the significant economic benefits generated by data mining in practical applications, the application of data mining in various fields has flourished [7-9]. For example, telecommunication companies, credit card companies, insurance companies and securities companies use data mining techniques to detect fraud, medical applications use data mining to determine that certain methods may be effective in treating a certain disease, and supermarkets use data mining to analyze transaction data and arrange the layout of goods placed on the shelves to increase sales. The financial field uses data mining techniques to discover the financial and business interests of a particular customer, consumer group, or organization, and can observe trends in the financial market [10-13].

After recognizing the functions of data mining, data mining techniques have been applied to competitive sports abroad. For example, the coach of American NBA utilizes Advanced Scout, a data mining tool provided by IBM, to assist the decision of replacing players on the spot, and has achieved very good results [14-16]. Currently, about 20 NBA teams use this software system to optimize their tactical combinations. However, in China, the application of data mining in the field of sports is very little, is still in its infancy, but has been increasingly emphasized [17-18]. With the gradual realization of China’s sports informatization, a large amount of data has been accumulated in the field of competitive sports, and how to make full use of these data and discover valuable but neglected important information from them has become an important task for researchers [19-20]. Table tennis is a competitive sport of confrontational nature, which not only requires athletes to have high technical level, but also has high requirements on the use of tactics, and very often the use of tactics in the game will play a decisive role in the victory or defeat of the game [21-22]. Therefore, it is necessary to apply data mining and knowledge discovery in the analysis of table tennis matches, the purpose of which is to analyze the large amount of complicated information provided by the database related to table tennis matches in real time and at a deeper level, and to find out the truly valuable information and knowledge from it, so as to provide reference and basis for coaches to decide on tactics, and their application [23-25].

In this paper, the MHSA module is introduced based on the YOLOv8 algorithm for extracting image features. Considering the bounding box position error, confidence error, and classification error, the performance of the target detection model is evaluated by a weighted combination, and the CIoU function is used to calculate coordinate loss. The default loss function in YOLOv8 is replaced by the EIoU loss function to improve the convergence speed and detection accuracy of the model. Multiple detection heads are added to the model’s network to detect small targets in the dataset by combining the multi-scale strategy. Using stereo matching and parallax method, the 3D coordinates of the target are obtained, and the characteristic corner points of the image are extracted to achieve the transformation from RGB image to HSV model. The motion trajectory of the table tennis ball is reconstructed using the least squares method to represent the trajectory of the table tennis ball and the coordinates are converted to real world coordinates based on the table tennis table. The trajectory of table tennis motion is recognized using the YOLOv8 model, in which the player’s strategy of hitting the ball to change the line is analyzed.

2

Table tennis target detection algorithm based on YOLOv8

2.1

YOLOv8 algorithm

YOLOv8 is a target detection and image segmentation tool that is unique in that it is highly scalable and integrated, allowing users to easily switch between different versions of YOLO and compare their performance. [26]. This is ideal for users who wish to use the latest YOLO technology but still need to use previous versions of the model. In addition, YOLOv8 runs efficiently and flexibly on multiple computational units, such as CPUs or GPUs, and multiple hardware platforms. This allows users to choose the right hardware configuration to run YOLOv8 according to their needs without worrying about performance. Overall, YOLOv8 is a powerful and flexible target detection and image segmentation tool that offers the latest technology with ease of use and the ability to compare all previous versions of YOLO. The default input image size for YOLOv8 is 640*640, which is a three-channel color image (RGB). After passing the image into the neural network, it undergoes a series of processing and this series of processing is done in the convolutional neural network. In one level of processing, the image size will become smaller and smaller, but the number of channels will be more and more, which is also determined by some parameters in the convolutional neural network. If the step size s is set to 1, the size of the image does not change, and if it is set to 2, the height and width of the image is halved. In short the data is processed by the backbone network, the neck network and the head network from front to back, from big to small, etc., and then the intermediate data and the processed data are spliced and processed to finally get the category and location of the detected target, which is an overall a principle and process.

2.2

YOLOv8-improved algorithm

The scale of the model in the YOLOv8 algorithm can be categorized into five, namely n, s, m, l and x. These five scales are designed to deal with different detection scenarios. They have different network widths, depths, and number of channels, but the model structure remains the same. In order to further improve the detection performance of the model, reduce the time for the model to recognize the moving ping-pong balls as well as improve the recognition accuracy of the model, this chapter proposes a YOLOv8-improved algorithm based on YOLOv8n for improvement.

2.2.1

Introduction of the Multi-Headed Attention Mechanism MHSA

In YOLO’s backbone network, basic operations such as convolutional and pooling layers are usually used to extract image features. These operations can effectively capture the local features of an image, but the model’s access to the global context is limited by certain operations in the convolutional neural network, such as the convolution operation. Fig. 1 shows the MHSA module. With the introduction of the MHSA module, the target feature information can be collected and associated by the model from the global large domain, which will help the neural network to learn the relationship between objects more comprehensively and accurately.

2.2.2

Improvement of the loss function

In the sports table tennis detection scenario, because the camera placement cannot interfere with the game, many of the table tennis targets in the final image are relatively small, so it is critical to have a stable detection effect for small targets. Usually, the loss function in YOLO integrates the bounding box position error, confidence error, and classification error, and measures the performance of the target detection model using weighted combination. Different loss functions are suitable for different tasks and model architectures, and by choosing a loss function suitable for a specific task and model, the convergence speed and detection accuracy of the model can be improved, etc. YOLOv8 uses the CIoU function as the coordinate loss calculation by default, and its formula is (1). (1) $L_{I o U} = 1 - I o U + \frac{ρ^{2} (b, b^{g t})}{c^{2}} + α ω$

In the formula, b, b^gt denote the centroid, the former denotes the centroid of the target frame and the latter denotes the centroid of the prediction frame, c, α and v denote the diagonal lengths of the smallest outer rectangles of the real and prediction frames, the weight parameters, and the ratio of the predicted frame to the height and width of the target frame, respectively.

After experiments, the default loss function converges slowly, which may be attributed to the fact that if v presents a linear scale when performing the regression of the prediction frame, then the regression task cannot be well described because the prediction frame height and width cannot be synchronized to change, i.e., the increase or decrease cannot be synchronized. In this paper, according to this characteristic, the EIoU loss function is used to replace the default loss function in YOLOv8, and it is experimentally verified that both the convergence speed and detection accuracy are improved.The EIoU formula is as follows: (2) $\begin{array}{rcl} L_{E I o U} & = & L_{I o U} + L_{d i s} + L_{a s p} \\ = & 1 - I o U + \frac{ρ^{2} (b, b^{g t})}{{(ω^{c})}^{2} + {(h^{c})}^{2}} + \frac{ρ^{2} (ω, ω^{s t})}{{(ω^{c})}^{2}} + \frac{ρ^{2} (h, h^{s t})}{{(h^{c})}^{2}} \end{array}$

In Equation (2), L_IoU denotes the IoU loss, L_dis denotes the center distance loss, L_asp denotes the azimuthal loss, and ω^c, h^c denote the height and width of the minimum closure frame.

2.2.3

Addition of small target detection layer

In the field of target detection, when there are small targets in the dataset, the problem of missed detection or poor detection often occurs [27]. To solve this problem, the YOLO algorithm adopts a multi-scale strategy and uses multiple detection heads in the network. YOLO has three detection heads by default, which are used to detect targets of different scales. This improves the detection capability and accuracy for targets of different sizes. The specific detection size depends on the specific version and configuration of YOLO. The three detection heads of YOLOv8 are as follows: P3/8 corresponds to the detection feature map size of 80*80, which is used for detecting targets of size 8*8 and above; P4/16 corresponds to the detection feature map size of 40*40, which is used for detecting targets of size 16*16 and above; P5/32 corresponds to the detection feature map size of 20*20, which is used for detecting targets of size 16*16 and above. P5/32 corresponds to a detection feature map size of 20*20, which is used to detect targets of size 32*32 and above. Using the default detection head is also enough, but there may be poor detection of tiny target detection, so based on YOLOv8 added a new 160*160 detection feature map, which can be used to detect targets above 4*4. Because the table tennis ball in the picture occupies a relatively small space, it belongs to the category of small targets. Therefore, adding a small object detection head to the small target dataset can improve the detection performance.

2.3

Calculation and analysis of table tennis trajectory

2.3.1

Trajectory reconstruction

1)

Fast three-dimensional localization

At this stage, the most commonly used in binocular vision is to correct the views of the two cameras through stereo correction, so that the corrected images are geometrically aligned, and then use stereo matching algorithms and parallax methods to find the three-dimensional coordinates of the target. However, traditional stereo matching algorithms such as BM, SGBM, and others. They are too inefficient and less practical in real engineering applications. Therefore, this paper combines the target detection algorithm introduced in the previous chapter to directly recognize the table tennis balls in the left and right eye images to get their pixel coordinates, combines the camera calibration parameters and uses the least squares method to obtain the 3D coordinates of the table tennis balls [28].

For a binocular vision system, after the corresponding target points in the image sequence captured by the two cameras are known, their spatial coordinates are calculated as follows: (3) $Z_{C 1} [\begin{matrix} u_{1} \\ v_{1} \\ 1 \end{matrix}] = M_{1} [\begin{array}{l} R_{1} & T_{1} \end{array}] [\begin{matrix} X_{W} \\ Y_{W} \\ Z_{W} \\ 1 \end{matrix}] = [\begin{array}{l} m_{111} & m_{112} & m_{113} & m_{114} \\ m_{121} & m_{122} & m_{123} & m_{124} \\ m_{131} & m_{132} & m_{133} & m_{134} \end{array}] [\begin{matrix} X_{W} \\ Y_{W} \\ Z_{W} \\ 1 \end{matrix}]$ (4) $Z_{C 2} [\begin{matrix} u_{2} \\ v_{2} \\ 1 \end{matrix}] = M_{2} [\begin{array}{l} R_{2} & T_{2} \end{array}] [\begin{matrix} X_{W} \\ Y_{W} \\ Z_{W} \\ 1 \end{matrix}] = [\begin{array}{l} m_{211} & m_{212} & m_{213} & m_{214} \\ m_{221} & m_{222} & m_{223} & m_{224} \\ m_{231} & m_{232} & m_{233} & m_{234} \end{array}] [\begin{matrix} X_{W} \\ Y_{W} \\ Z_{W} \\ 1 \end{matrix}]$

In the above equation, ${[\begin{array}{l} u_{1} & v_{1} & 1 \end{array}]}^{T}$ and ${[\begin{array}{l} u_{2} & v_{2} & 1 \end{array}]}^{T}$ are the chi-square pixel coordinates corresponding to the target point ${[\begin{array}{l} X_{W} & Y_{W} & Z_{W} \end{array}]}^{T}$ in the image sequence captured by the two cameras, and M₁, M₂, $[\begin{array}{l} R_{1} & T_{1} \end{array}]$ , and $[\begin{array}{l} R_{2} & T_{2} \end{array}]$ are the known parameters indicating the internal and external parameters of the two cameras. Equations (3) and (4) can be obtained by eliminating Z_C1 and Z_C2 to simplify: (5) ${\begin{matrix} \begin{array}{l} (u_{1} m_{111} - m_{111}) X_{W} + (u_{1} m_{132} - m_{112}) Y_{W} \\ + (u_{1} m_{113} - m_{113}) Z_{W} = m_{114} - u_{1} m_{134} \end{array} \\ \begin{array}{l} (v_{1} m_{111} - m_{121}) X_{W} + (v_{1} m_{132} - m_{112}) Y_{W} \\ + (v_{1} m_{113} - m_{123}) Z_{W} = m_{124} - v_{1} m_{134} \end{array} \\ \begin{array}{l} (u_{2} m_{231} - m_{211}) X_{W} + (u_{2} m_{232} - m_{212}) Y_{W} \\ + (u_{2} m_{233} - m_{213}) Z_{W} = m_{214} - u_{2} m_{234} \end{array} \\ \begin{array}{l} (v_{2} m_{231} - m_{221}) X_{W} + (v_{2} m_{232} - m_{222}) Y_{W} \\ + (v_{2} m_{233} - m_{223}) Z_{W} = m_{224} - v_{2} m_{234} \end{array} \end{matrix}$

Simplified in the form of $A {[\begin{array}{l} X_{W} & Y_{W} & Z_{W} \end{array}]}^{T} = b$ , the three-dimensional coordinates of the target can be calculated by the least squares method: (6) $[\begin{matrix} X_{W} \\ Y_{W} \\ Z_{W} \end{matrix}] = {(A^{T} \cdot A)}^{- 1} \cdot A^{T} \cdot b$ 2)

Feature corner point extraction

Generally, the camera’s sensor captures light information in the form of RGB, so in this paper we need to convert the captured RGB image to HSV color model. Assuming that a pixel is represented as (r, g, b) and (h, s, v) in the RGB and HSV color models respectively, if the image is converted from the RGB color model to the HSV color model, the conversion relationship of the three components needs to be known h, s, v separately.

First calculate the minimum, maximum and hue in RGB, the hue is determined according to the maximum and minimum values in RGB, if the maximum and minimum values are equal, then the hue is 0. Let MAX = max(r, g, b), MAX = max(r, g, b), then: (7) $H = {\begin{matrix} 0 & M A X = M I N \\ (\frac{60^{°} (g - b)}{M A X - M I N} + 360^{°}) \mod 360^{°} & M A X = r \\ \frac{60^{°} (b - r)}{M A X - M I N} + 120^{°} & M A X = g \\ \frac{60^{°} (r - g)}{M A X - M I N} + 240^{°} & M A X = b \end{matrix}$

Then calculate the saturation, if the maximum value is 0, the saturation is 0, the formula is shown in (8). Finally calculate the brightness, when s = 0, the maximum value is the brightness, which is v = MAX: (8) $s = {\begin{matrix} 0 & M A X = 0 \\ 255 \times \frac{M A X - M I N}{M A X} & M A X \neq 0 \end{matrix}$

2.3.2

Calculation of table tennis trajectory

Given the set of scatter coordinates of the trajectory ${(x_{1}, y_{1}), (x_{2}, y_{2}), \dots (x_{n}, y_{n})}$ , the general expression for the power function model is: (9) $p (x) = a_{0} + a_{1} x + a_{2} x^{2} + \dots + a_{n} x^{n}$

Use least squares to fit these coordinates.

Find the parameters of the above polynomial such that: (10) $\min {\sum_{i = 1}^{k} {(p (x_{i}) - y_{i})}^{2}}$

Equation (9) is solved for all values of a_i by taking the partial derivative of a₀, a₁, …, a_n and making it equal to 0, respectively: (11) $\begin{array}{l} a_{0} * n + a_{1} * \sum_{i = 1}^{n} x_{i} + a_{2} * \sum_{i = 1}^{n} x_{i}^{2} + \dots \dots + a_{k} * \sum_{i = 1}^{n} x_{i}^{k} = \sum_{i = 1}^{n} y_{i} \\ a_{0} * \sum_{i = 1}^{n} x_{i} + a_{1} * \sum_{i = 1}^{n} x_{i}^{2} + a_{2} * \sum_{i = 1}^{n} x_{i}^{3} + \dots \dots + a_{k} * \sum_{i = 1}^{n} x_{i}^{k + 1} = \sum_{i = 1}^{n} x_{i} * y_{i} \\ a_{0} * \sum_{i = 1}^{n} x_{i}^{2} + a_{1} * \sum_{i = 1}^{n} x_{i}^{3} + a_{2} * \sum_{i = 1}^{n} x_{i}^{4} + \dots \dots + a_{k} * \sum_{i = 1}^{n} x_{i}^{k + 2} = \sum_{i = 1}^{n} x_{i}^{2} * y_{i} \\ \dots \dots \\ a_{0} * \sum_{i = 1}^{n} x_{i}^{k} + a_{1} * \sum_{i = 1}^{n} x_{i}^{k + 1} + a_{2} * \sum_{i = 1}^{n} x_{i}^{k + 2} + \dots \dots + a_{k} * \sum_{i = 1}^{n} x_{i}^{2 k} = \sum_{i = 1}^{n} x_{i}^{k} * y_{i} \end{array}$

The expression written as a matrix is: (12) $⌊ \begin{matrix} n & \sum_{i = 1}^{n} x_{i} & \sum_{i = 1}^{n} x_{i}^{2} & \dots & \sum_{i = 1}^{n} x_{i}^{k} \\ \sum_{i = 1}^{n} x_{i} & \sum_{i = 1}^{n} x_{i}^{2} & \sum_{i = 1}^{n} x_{i}^{3} & \dots & \sum_{i = 1}^{n} x_{i}^{k + 1} \\ \sum_{i = 1}^{n} x_{i}^{2} & \sum_{i = 1}^{n} x_{i}^{3} & \sum_{i = 1}^{n} x_{i}^{4} & \dots & \sum_{i = 1}^{n} x_{i}^{k + 2} \\ \dots & \dots & \dots & \dots & \dots \\ \sum_{i = 1}^{n} x_{i}^{k} & \sum_{i = 1}^{n} x_{i}^{k + 1} & \sum_{i = 1}^{n} x_{i}^{k + 2} & \dots & \sum_{i = 1}^{n} x_{i}^{2 k} \end{matrix} ⌋ [\begin{array}{l} a_{0} \\ a_{1} \\ a_{2} \\ \dots \\ a_{k} \end{array}] = ⌊ \begin{matrix} \sum_{i = 1}^{n} y_{i} \\ \sum_{i = 1}^{n} x_{i} y_{i} \\ \sum_{i = 1}^{n} x_{i}^{2} y_{i} \\ \dots \\ \sum_{i = 1}^{n} x_{i}^{k} y_{i} \end{matrix} ⌋$

Hypothesis: (13) $X = [\begin{matrix} 1 & x_{1} & x_{1}^{2} & \dots & x_{1}^{k} \\ 1 & x_{2} & x_{2}^{2} & \dots & x_{2}^{k} \\ 1 & x_{3} & x_{3}^{2} & \dots & x_{3}^{k} \\ \dots & \dots & \dots & \dots & \dots \\ 1 & x_{n} & x_{n}^{2} & \dots & x_{n}^{k} \end{matrix}], Y = [\begin{matrix} y_{1} \\ y_{2} \\ y_{3} \\ \dots \\ y_{n} \end{matrix}]$

The above matrix calculation can be simplified as: X^TXa = X^TY to obtain: $a = {(X^{T} X)}^{- 1} X^{T} Y$ .

Since the acquired scatter coordinates of the table tennis ball are based on the pixel coordinates of the image, it is not possible to accurately determine the instantaneous coordinates of the table tennis ball touching the table, but it can be observed that the slope of the trajectory of the table tennis ball changes before and after it touches the table, and according to this principle, the complete coordinate data of the athlete’s striking process is obtained. (14) ${(x_{1}, y_{1}), (x_{2}, y_{2}), \dots, (x_{n}, y_{n})}$

2.3.3

Coordinate conversion

The drop point coordinates obtained from the analysis are pixel coordinates based on the image, which are transformed into real world coordinates based on the table tennis table. A real coordinate system is established with the edge of the table horizontally to the right as the X-axis positive and the direction of the net vertically down as the Y-axis positive. The coordinate transformation uses a perspective change of the image.

The generalized formula for perspective transformation is: (15) $[x', y', w'] = [u, v, w] ⌊ \begin{array}{l} a_{11} & a_{12} & a_{13} \\ a_{21} & a_{22} & a_{23} \\ a_{31} & a_{32} & a_{33} \end{array} ⌋$

u, v are the original image coordinates.

Transformed coordinates: x = x′/w′, y = y′/w′. (16) $\begin{array}{l} x = \frac{x'}{w'} = \frac{a_{11} u + a_{21} v + a_{31}}{a_{13} u + a_{23} v + a_{33}} \\ y = \frac{y'}{w'} = \frac{a_{12} u + a_{22} v + a_{32}}{a_{13} u + a_{23} v + a_{33}} \end{array}$

Where, $⌊ \begin{array}{l} a_{11} & a_{12} & a_{13} \\ a_{21} & a_{22} & a_{23} \\ a_{31} & a_{32} & a_{33} \end{array} ⌋$ is the perspective transformation matrix, $⌊ \begin{array}{l} a_{11} & a_{12} \\ a_{21} & a_{22} \end{array} ⌋$ represents a linear transformation, $[a_{31} a_{32}]$ is used for translation and ${[a_{13} a_{23}]}^{T}$ produces a perspective transformation [29].

2.4

Table tennis trajectory recognition

2.4.1

Table tennis trajectory analysis

Figure 2 shows the results of table tennis ball trajectory prediction based on YOLOv8, Figure (a) shows the horizontal direction error and Figure (b) shows the vertical direction error. From the results, the horizontal and vertical errors are less than 4.261mm before the rebound, and after the rebound, the horizontal and vertical errors are less than 9.874mm, which is a good prediction accuracy and meets the requirements of trajectory detection.

2.4.2

Localized tracking of table tennis balls

In order to verify the effectiveness of the table tennis motion trajectory detection method based on deep learning and hyperpixel, 20 sets of table tennis video production datasets under different lighting are collected, and different methods are used to detect and compare the results. In the data processing platform, the central processor model is Intel Core i3-9100, and the graphics processor model is NVIDIA Essence GTX 1080Ti.

The training set of table tennis localization module is produced by combining various table tennis game videos. To improve the detection accuracy and training speed of the table tennis ball localization module, optimal training weights are obtained using distributed training. To quantitatively analyze the effect of table tennis ball localization, 1,000 images are selected. The detection accuracy and detection time of the table tennis ball motion trajectory detection method based on deep learning and hyperpixel are compared with other methods, and the results are shown in Table 1.When performing table tennis ball localization, the accuracy of the table tennis ball motion trajectory detection method based on YOLOv8 reaches 98.39%, which is substantially higher than that of traditional methods, especially when compared with the traditional YOLO algorithm, and the accuracy is significantly improved. The color and contour-based detection method focuses on detecting the target appearance and ignores the background, making it difficult to adjust the set correlation threshold to the changing environment in real-time. The YOLOv8-based table tennis ball motion trajectory detection method focuses more on the effective information around the target, resulting in better detection results.

Table 1.

Comparison of table tennis positioning and tracking method

Method	Accuracy rate	Single frame detection time /ms	Model size /Mibyte
Color detection	55.56%	8.15	0.2
Contour detection	64.22%	9.23	0.2
YOLOv4	97.26%	27.01	244.66
YOLOv5s	96.18%	15.21	13.75
Faster regional convolution neural network	75.12%	198.25	113.48
YOLOv8	98.39%	15.23	13.24
Method	Central position pixel error	Detection speed /(frame •s⁻¹)	Model size /Mibyte
YOLOv4	14 pixels	43	244.66
YOLOv5s	36 pixels	75	13.75
Faster regional convolution neural network	-	3	113.48
YOLOv8	11 pixels	64	13.24

From the tracking results, the table tennis ball movement trajectory detection method based on YOLOv8 is able to track the table tennis ball correctly compared to mainstream deep learning algorithms, with fewer cases of misrecognition and non-recognition, and it is more efficient in dealing with blurring and other problems, with no cases of misrecognition of targets outside the target area, and with higher localization precision and accuracy. The error pixel of the center position is 11 pixels, while the size of the model is 13.24M, which is the smallest among the four methods.

2.4.3

Spatial trajectories of table tennis balls in different situations during a match

Figure 3 shows the spatial motion trajectories of table tennis ball in different situations in the game, which are reproduced by using a series of 3D coordinate points reconstructed by the model in this paper, these coordinates represent the positions of the table tennis ball in space at different moments, in order to better display the motion trajectories of the table tennis ball, the reconstructed data will be fitted with nonlinear curves. In this paper, we will extract some clips from table tennis game or training to reproduce the trajectories.Figures (a)~(c) show the motion trajectories of table tennis game clips 1-3 respectively. In the match clip 1, the table tennis ball falls from 390.10181mm on the z-axis to -314.0032mm on the x-axis to the lowest point, and then the trajectory starts to rise up to the highest point of 549.2869mm. While in clip 2, the table tennis ball firstly rises up to 332.58339mm on the z-axis, and then continues to fall down. In clip 3, the ping pong ball rises to 302.72132mm and then drops to 200mm left, followed by continuing to rise to 300mm.

3

Analysis of the change of line strategy of table tennis players

3.1

Research Objectives and Methodology

3.1.1

Subject of the study

This paper focuses on the batting routes of FZD in international matches in 2021-2022, and uses the model constructed in this paper to analyze the players’ table tennis trajectories and study the application of the players’ line-change strategies in matches.

3.1.2

Research methodology

1)

Video research method

Referring to the research ideas of related scholars on the use of video observation method to observe relevant objects, this study collected videos of 30 150-game international table tennis singles matches in FZD from 2021 to 2022, and repeatedly observed and counted data on the match videos with the incoming line of the last board of the opposing player in each round as the base point of observation.

2)

Mathematical and statistical method

The division of the table tennis table landing area was divided into 6 areas, and at the same time, through the use of Excel, the data were counted and the required data were imported into SPSS to establish a database, and then the data were implemented to statistically analyze and test.

The evaluation indexes were the utilization rate, scoring rate, and contribution rate of the batting line form. By analyzing the usage rate, scoring rate, and contribution rate, it is conducive to understanding Fan Zhendong’s batting line, the scoring effect, and the contribution to the game. Its specific calculation form is as follows:

Scoring rate = Score of batting line form / Score of batting line form (Score + Loss) * 100%

Utilization rate = batting line form (points scored + points conceded) / each batting line form (total points scored + total points conceded) * 100%

Contribution rate = batting line form score / each batting line form (total score + total points conceded) * 100%

3)

Comparative Analysis Method

By comparing the characteristics of batting line application in 30 international matches of FZD 2021-2022 at different levels of variations and non-variations, the recorded data were objectively analyzed and elaborated, and the advantages and disadvantages of the two line forms applied in the matches were pointed out.

3.2

Research on Table Tennis Match Strategies

3.2.1

Parameters of variable and non-variable lines

In terms of the utilization of the lane-changing and non-lane-changing strategies, as shown in Table 2, the utilization rate of lane-changing, i.e., the points gained and lost mainly caused by the implementation of the lane-changing strategy, accounted for 40.02% of the total number of points scored by both sides of the game, which is slightly higher than 33.3% but much lower than that of the non-lane-changing strategy, which accounted for 59.98%. Meanwhile, the results of the independent samples t-test also showed that the utilization rates of the two strategies were highly significantly different. This suggests that in table tennis, the non-variable line strategy, i.e., the desire to win by the momentum factor, is the main form of athletes’ technical and tactical utilization, while the variable line strategy, i.e., the desire to win by the spatial factor, is the auxiliary form of athletes’ technical and tactical utilization.

Table 2.

The application of the variation and the non-linear strategy

/	Frequency	Score	Lose point	Scoring rate	Usage rate	Contribution rate
Overall data	6154	3077	3077	50.00%	100.00%	50.00%
Variation line	2463	1536	927	62.36%	40.02%	24.96%
Invariant line	3691	1625	2066	44.03%	59.98%	26.41%
T				21.756	14.564	1.485
P				0.000	0.000	0.135

In terms of the effectiveness of the use of lane changing and non-lane changing strategies, the scoring rate of lane changing was 62.36%, which was much higher than the scoring rate of non-lane changing which was 44.03%. Also, the results of the independent samples t-test indicate that there is a highly significant difference between the two scoring rates. This indicates that the scoring effect of the line-change strategy is significantly better than that of the non-line-change. This also reveals that constantly increasing the utilization rate of the line-change strategy should be an important consideration for athletes’ tactics under the premise of guaranteeing absolute stroke quality.

In terms of the contribution of variance and non-variance strategies to match winning, variance contributes 24.96% and non-variance contributes 26.41%, and although the contribution of non-variance is slightly higher than that of variance, the results of the independent samples t-test show that there is no significant difference between the two. This indicates that the two play almost equally important roles as far as their contribution to winning the game is concerned. If we also consider the fact that the scoring rate of the lane changing strategy is much higher than that of the non-lane changing strategy, as well as the fact that the lane changing strategy is characterized by trickery and economy of effort. It is important to further develop and strengthen athletes’ awareness and ability to change lanes in training, whether it is for winning a point, a game, or a match, or for the long-term development of athletes.

3.2.2

Parameters of the various variant forms

Table 3 shows the parameters of the various variations. In terms of the utilization of the three types of variation strategies, the variation to forehand position strategy has the highest utilization rate of 37.60%, the variation to backhand position strategy has the second highest utilization rate of 37.43%, and the variation to intermediate position strategy has the lowest utilization rate of 24.97%. Meanwhile, the results of ANOVA also showed that, consistent with the results of Fang Yuping’s study, overall, there was a significant difference in the utilization rates of the three types of line changing strategies. The results of multiple comparisons of means showed that there was a significant difference in the utilization rates of the two types of strategies, namely, the line changing to forehand position and the two types of strategies, namely, the line changing to backhand position and the middle position, and there was no significant difference in the utilization rates of the two types of strategies, namely, the line changing to backhand position and the line changing to the middle position. This suggests that breaking from the forehand is the main orientation choice of the world’s best table tennis players when using the changeover strategy, while breaking from the backhand and breaking from the middle are the secondary orientation choices when using the changeover strategy. There are three reasons why breaking from the forehand is the primary change of direction strategy: (1) Due to the influence of the body structure and the threat of hitting the ball, it is generally believed that the threat of hitting the ball from the backhand is less than that of the forehand, and considering the fact that it is easier for the player to move to the same side of the racket holder than to the opposite side of the racket holder during the match, in order to protect the backhand, the stance of the two sides of the opposing team has been favored to be the backhand position for a long time. In order to protect the backhand position, both players in a table tennis match have long favored the backhand position, which also determines that the match is usually played from the backhand position. Because of this, the change of line forehand position, especially from our forehand position or the middle position to the opponent’s forehand position diagonally, can not only mobilize the opponent in the maximum range, but also produce the greatest killing power because the player will use more forehand techniques in this position. (2) Compared with the backhand, the forehand attack is easier to power, the threat of attack is also greater, which makes both athletes psychologically are not quite dare to return the ball to the opponent’s forehand, but the greater the risk, the greater the opportunity, it is on the basis of a clear understanding of the psychological limitations, most of the outstanding athletes will be through the reverse thinking, the opposite of what they are doing, will be the changeover to the forehand as the first choice of changeover form. (3) Even if they can’t score directly when changing to forehand position, they can still take the initiative because they can mobilize the opponent by changing to forehand position, and then seize the weakness of the opponent’s movement to the opposite side of the racket-carrying hand when attacking the backhand position.

Table 3.

Various forms of variation

/	Frequency	Score	Lose point	Scoring rate	Usage rate	Contribution rate
Variation line (total)	2463	1536	927	62.36%	100.00%	50.00%
Meet... Intermediate level	615	452	163	73.50%	24.97%	18.35%
Meet... Become forehand	926	592	334	63.93%	37.60%**	24.04%**
Meet... Variable backhand bit	922	452	470	49.02%	37.43%	18.35%
F				0.785	21.264	11.265
P				0.422	0.000	0.000

Note: “Change the middle position” means that the ball changes the middle position every different line, and the other two types are the same.

3.3

Batting line characterization

In this paper, the specific batting line combination forms of FZD in the game were classified into 27 kinds, among which there were 18 kinds of specific line combination forms of variable line and 9 kinds of specific line combination forms of non-variable line. Through the statistics, comparisons, and analyses on the three aspects of the utilization rate, scoring rate, and contribution rate of the variable and non-variable line combination forms, it is conducive to the discovery of the characteristics of the application of the batting line combination forms of FZD in the game.

3.3.1

Route-specific characterization of the change of line

Statistical analysis of the 18 specific variant line combination forms is helpful to find out the advantages and disadvantages of using line combination forms of FZD variant in order to improve the use of advantageous variant combination forms.Detailed information on the scores, utilization rates, and contribution rates of the 18 specific variant line combination forms are shown in Table 4.

Table 4.

Analysis of the specific route characteristics of the variation line

Number	Variant line combination form	frequency	score	Lose point	Usage rate(%)	Scoring rate(%)	Contribution rate(%)
1	Reverse skew - anyway straight	150	105	45	16.59%	70.00%	11.62%
2	Middle reverse oblique - reverse straight	23	16	7	2.54%	69.57%	1.77%
3	Anti-center oblique – center Positive and oblique	58	45	13	6.42%	77.59%	4.98%
4	Medium straight - medium positive oblique	23	18	5	2.54%	78.26%	1.99%
5	Anti-integrity – right Positive and oblique	134	90	44	14.82%	67.16%	9.96%
6	\| medium oblique – Positive and oblique	49	40	9	5.42%	81.63%	4.42%
7	Anticlinal. - Anticlinal	36	30	6	3.98%	83.33%	3.32%
8	Straight forward and backward mesoblinia	7	6	1	0.77%	85.71%	0.66%
9	· Anti-center oblique - center straight	15	13	2	1.66%	86.67%	1.44%
10	Median oblique - median straight	9	7	2	1.00%	77.78%	0.77%
11	Positive oblique - median oblique	25	22	3	2.77%	88.00%	2.43%
12	Counterup rightness - median oblique	17	17	0	1.88%	100.00%	1.88%
13	Positive oblique - positive straight	65	53	12	7.19%	81.54%	5.86%
14	Medium oblique - straight forward and backward	25	20	5	2.77%	80.00%	2.21%
15	Straight forward and backward	112	80	32	12.39%	71.43%	8.85%
16	Middle antiskew - antiskew	83	50	33	9.18%	60.24%	5.53%
17	Median straight - median anticlinal	38	25	13	4.20%	65.79%	2.77%
18	Median oblique - median anticline	35	24	11	3.87%	68.57%	2.65%

In this paper, the 18 variant combination forms are divided into three parts based on their utilization rate, the first part is the 6 variant combination forms numbered 1, 5, 15, 16, 13 and 3, the second part is the 8 variant combination forms numbered 6, 18, 7, 17, 2, 11, 14 and 4, and the third part is the 4 variant combination forms numbered 12, 9, 8 and 10.

The first part of the 6 combinations of variable line forms had a utilization rate between 6.42% and 16.59%, with a total utilization rate of 66.59% and an average utilization rate of 11.098%. The second part of the 8 combinations of variable line forms has a utilization rate between 2.54% and 5.42%, with a total utilization rate of 28.09% and an average utilization rate of 3.511%. In the third part, the utilization rate of the 4 types of variation forms ranged from 0.77% to 1.88%, with a total utilization rate of 5.31% and an average utilization rate of 1.328%, which shows that the three parts of the game have different emphasis in their application. Therefore, from an overall point of view, it reflects the characteristics of FZD in the match with various forms of line changes but with clear priorities, and it also reflects the importance of line changes in the match from the side. Some scholars have proposed that the winners and losers of racket holders’ net confrontation events depend on the power, speed, rotation, landing point and various line changes of each hit of the athletes, and the present study is completely consistent with it in terms of the importance of the changes of lines in the match. In addition, by analyzing the first part of the six line changes with the highest usage rate, five of them are line changes to change the opponent’s big angle, which are 1 reverse backslash - straight, 5 straight - forward forward slash, 15 forward backslash - reverse backslash, 13 forward slash - forward backslash, 3 anti-center slant - center-square slant.

3.3.2

Chi-square test

Figure 4 shows the chi-square test of the usage rate of FZD’s lane change line combination forms, and the chi-square test results show that there is a very significant difference in the usage of the 18 lane change combination forms in the game (P=0.000<0.01), and the vast majority of the 6 lane change forms in the first part of the game have a very significant difference with the second part of the game, and there is a very significant difference with the third part of the game for each form, and the second and the third part of the game also have a very significant difference in the usage the vast majority of them also have very significant differences, indicating that all 18 forms of lane changing were applied to different degrees by FZD in the game, thus fully reflecting the characteristics of FZD’s strong awareness of lane changing and flexible and diverse forms of lane changing.

Note: Indicates that there is a very significant difference between the two line combination forms, P<0.01, * indicates that there is a significant difference between the two line combination forms, P<0.05.

4

Conclusion

In this paper, based on the YOLOv8 algorithm, the MHSA module is introduced in the multiple attention mechanism, and the EIoU loss function is used to replace the default loss function in YOLOv8. Add a small target detection layer to reconstruct the trajectory of the table tennis ball movement. Analyze the spatial motion trajectory of the table tennis ball in different situations during the game, and study the player’s strategy of changing lines. 1)

The results of table tennis ball trajectory prediction based on YOLOv8 show that the errors in the horizontal and vertical directions before rebound are less than 4.261mm, and after rebound, the errors in the horizontal and vertical directions are less than 9.874mm. It shows that the model’s trajectory recognition accuracy is good and meets the requirements for trajectory detection.

2)

Sorting out the parameters of the table tennis game strategy in which athletes use lane changing and non-lane changing strategies, the utilization rate of lane changing of the subject athletes was 40.02%, meanwhile, the scoring rate of lane changing was 62.36%, which was much higher than the scoring rate of non-lane changing which was 44.03%. Under the assumption that absolute stroke quality is guaranteed, the utilization rate of the lane changing strategy can be increased, thus improving the overall score.

3)

Categorizing the 18 specific line-change route combinations and analyzing the line-change specific routes, the first part of the six combination line-change forms had a usage rate between 6.42% and 16.59%, with a total usage rate of 66.59% and an average usage rate of 11.098%.

Lingua:: Inglese

Frequenza di pubblicazione:: 1 volte all'anno
Argomenti della rivista:: Scienze biologiche, Scienze della vita, altro, Matematica, Matematica applicata, Matematica generale, Fisica, Fisica, altro

Feed RSS della rivista

Research on pattern recognition and sports performance of table tennis game strategy based on big data mining technology

Ku Duan

Pubblicato online: 21 mar 2025

Ricevuto: 30 ott 2024

Accettato: 30 gen 2025

DOI: https://doi.org/10.2478/amns-2025-0694

Parole chiaveMHSA module, Loss function, YOLOv8 algorithm, Table tennis ball trajectory, Line change strategy

© 2025 Ku Duan, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Parole chiave
MHSA module, Loss function, YOLOv8 algorithm, Table tennis ball trajectory, Line change strategy