Introduction

With people’s love for football and the broadcast of major football events, football has become the world’s largest competitive sport. In addition to talent and hard work, football players also need scientific training. On the other hand, traditional technical training of football players relies on the guidance of football coaches. Still, that guidance is based on coaching experience and the technical judgment of football players from the naked eye. The technical training of football players lacks a systematic evaluation standard for assessing their specific movements. Different players have different physical characteristics, technical abilities, and behavioral habits. If people only rely on football coaching experience to guide them, it will be challenging to see exceptional players improve1. With the continuous development of machine vision, the LBP algorithm has a good recognition effect in image texture analysis2. The Internet of Things (IoT) terminal analyzes football players’ movements, identifying strengths and weaknesses and proposing strategies to enhance their technical skills.

Things terminal and corresponding advancement strategies are proposed to enhance the technical movements of football players. Artificial intelligence technology has been extensively utilized in sports training due to the advancement and use of computer technology. Improving training quality is essential to enhance players’ competitive abilities in football. Analysis of extensive training data reveals a need to strengthen the consistency of basic training practices. As a result, the lack of personalized training techniques and content has made it difficult to adapt to the tendency of football advancement in the context of improved performance at the global football level and the diversification of the training process. This method will improve athletes’ abilities, and it is critical to incorporate artificial intelligence technology into football. Deep learning, when applied to recognizing football player behavior, can aid in creating a high-quality team. Deep learning can identify the shooting action while also analyzing the regulations and procedures for goals of scoring. It greatly aids in advancing football exercise techniques and the correctness of football shooting.

Besides the above, identifying the technical action behavior of football players can help football players conduct scientific football training. Preceding researchers have proposed emphasizing techniques for various athletic competitions, from the most basic to the most specialized. Among them, Gu et al.3 research showed that intelligent technologies such as machine vision could analyze and identify the behavior of moving objects, which can be used to analyze football players’ sports technology. Smm et al.4 improved the players’ skills by studying the football players’ technical movements in the 2108 World Cup and training the corresponding football players. Luvizon et al.5 classified the technical action features of football players and can accurately identify different technical actions. The research of Liu et al.6 applied the moving target tracking algorithm to the action recognition of football players, and it has a high recognition rate under different light. Jiang et al.7 used intelligent technology to identify the skills of football players and used terminal equipment to display specific technical action data. Although the recognition of the action behavior of football players can effectively analyze the accuracy of the action, the recognition method used is not optimal.

The combination of PCA and LBP has an efficient image recognition effect, and more and more people apply the PCA–LBP algorithm to the recognition of the technical action behavior of football players. Among them, Zhang et al.8 research indicated that using PCA to reduce the dimension of the LBP algorithm can enhance the recognition accuracy of the algorithm. The research of Werghi et al.9 pointed out that LBP can detect and recognize motion through texture analysis. Hwang and Abebe10 said that the LBP algorithm can deeply analyze the technical actions of football players and compare the recognized actions with the action library, which has a high degree of recognition. Kazak and Koc11 used the combination of PCA and LBP to identify the technical movements of football players. PCA can reduce the complexity of the identification data and shorten the identification time. Although the PCA–LBP algorithm can effectively and accurately identify the technical action behavior of football players, it lacks the comparison of the recognition accuracy with the single LBP algorithm.

Identifying football players’ technical action behaviors can help players understand their action data and provide scientific training methods. In the traditional LBP algorithm recognition process, PCA elements were added to simplify the processing of technical action recognition feature data to improve the efficiency of action feature recognition. The innovation of this paper was to use the PCA–LBP algorithm to identify football players’ technical action behavior and compare and analyze it with the traditional LBP identification algorithm. This paper uses deep learning to identify and evaluate football players’ technical actions based on verified movement images, allowing for precise analysis and standardization. It can also efficiently accurately the player’s action behavior to standardize the player’s action.

Method of technical action behavior recognition

There is still a lack of scientific training for football, and football coaches can only train players based on their stadium experience. At the same time, football players do not fully understand their sports movements. The exercises are usually performed as the coach directs12. Accurate identification of football players’ technical action behavior can obtain precise data on football players and provide targeted scientific training methods13. Feature extraction and classification techniques have been extensively studied for pattern recognition in various domains. Kumar et al.14 proposed biometric-based methods such as muzzle point pattern recognition for individual cattle identification, emphasizing hybrid feature extraction and classification approaches. Other studies15,16 further explored real-time recognition techniques using animal biometrics, achieving high accuracy and practical applicability. These approaches align with the objectives of this study, where feature descriptors like LBP and dimensionality reduction using PCA are employed for accurate action recognition.

Advancements in video-based action recognition and player detection have been extensively explored in the context of sports analytics. For instance, Mazzeo et al.17 proposed techniques for detecting and tracking players in soccer matches, while Gadde and Jawahar18 introduced weakly-supervised methods for player detection in broadcast videos. Recent works, such as those by Xue and Chen19, focused on recognizing football shot actions using deep learning. Giancola et al.20 explored active learning for action spotting in soccer videos. Surveys by Wu et al.21 and Yin et al.22 provide comprehensive overviews of datasets, methodologies, and applications in sports action recognition, highlighting the field’s growing importance.

The model for recognizing football players’ technical action behavior is shown in Fig. 1. While this study focuses on recognizing four foundational actions—kicking, dribbling, stopping, and fake actions—these were chosen as they represent a spectrum of frequently used technical movements with varied levels of complexity. Future research will incorporate a broader range of actions, including passing, tackling, and heading, to test the PCA–LBP model’s robustness further. This expansion will validate the model’s adaptability to diverse and dynamic football performance.

Fig. 1
Fig. 1
Full size image

Model diagram of football player’s technical action behavior recognition.

Figure 1 shows many kinds of technical actions in football, such as header, pass, dribble, etc. Accurate recognition of football players’ technical movements can obtain their technical movement data and provide scientific training for the athletes through intelligent analysis of the data23.

Technical action recognition preprocessing

It is easy to be affected by external factors when detecting the technical movements of football players. Among them, light is the most significant interference in detection and recognition. Whether the light is too intense or too dark, the quality of technical motion detection would be reduced. The technical action preprocessing can improve action detection, which is helpful for the subsequent recognition of the technical action behavior of football players.

In preprocessing football players’ technical movements, converting the original technical movement pictures into grayscale images is necessary. This is because the grey image is more beneficial to the analysis of the player’s action by eliminating the influence of color, and the original player’s technical action image is usually represented by red–green–blue (RGB)24. The RGB model is shown in Fig. 2.

Fig. 2
Fig. 2
Full size image

RGB model diagram.

In Fig. 2, red, green, and blue represent the original image’s three color tones. A point in the RGB model can represent any color in a pixel. Converting the original color image to grayscale enhances computational efficiency by reducing data complexity25.

The conversion from RGB to grayscale calculates the brightness value of each pixel, expressed as:

$$V=0.3R+0.59G+0.11B$$
(1)

In Formula (1), the converted pixel brightness is composed of 0.3 red, 0.59 green, and 0.11 blue of the original pixel.

The image rotation of the converted grayscale image of the football player is beneficial in improving the analysis of the football player’s action. Let the coordinate of a point in the original image be \({S}_{0}({x}_{0},{y}_{0})\), the inclination angle of the coordinate to the court ground is a, and the image is rotated clockwise by an angle b. \({x}_{0},{y}_{0}\) ensuring they accurately represent the intended transformations concerning the origin points. The pixel coordinates in the original image can be expressed as:

$$\left\{\begin{array}{c}{x}_{0}=d\mathit{cos}a\\ {y}_{0}=d\mathit{sin}a\end{array}\right.$$
(2)

The coordinates of the pixel points after rotating the angle \(b\) are expressed as:

$$y={x}_{0}\text{sin}\left(b\right)+{y}_{0} cos(b)$$
(3)

A matrix represents the rotated pixels.

$$\left[\begin{array}{c}x\\ y\\ 1\end{array}\right]=\left[\begin{array}{c}\mathit{cos}b \, \, -\mathit{sin}b \, \, \, \, 0\\ \mathit{sin}b \, \, \, \, \, \mathit{ }\mathit{cos}b \, \, \, \, 0\\ \, \, \, 0 \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, 0 \, \, \, \, \, \, \, \, \, \, \, 1\end{array}\right]$$
(4)

The image preprocessing process also needs to filter the converted image. The filtering process can remove the noise in the action image of the football player, which can better highlight the action information of the football player26. The filtering process is simple and efficient. Let the pixels in the rotated image be \(h(a,b)\), the processed pixels are \(f(a,b)\), and the center pixel of the image is \(h(c,d)\). The filtering process requires a weighted analysis of each pixel’s value:

$$f(a,b)=\frac{\sum_{c,d}h(c,d)v(a,b,c,d)}{{\sum }_{c,d}v(a,b,c,d)}$$
(5)

The weighting parameter (\(v\)) was determined through empirical testing to balance noise reduction with feature preservation, optimizing image clarity for accurate action recognition. In Formula (5), \(v(a,b,c,d)\) represents the pixel point weighting parameter.

Preprocessing the original football player’s technical action image includes image grayscale processing, image rotation, and filtering processing. The preprocessed image has more direct image information and reduces the influence of other image impurities.

Traditional LBP algorithm

LBP is an operator that analyzes the characteristics of local images. It has a superhigh recognition effect on grayscale images. LBP is the most common recognition algorithm in computer vision and is widely used in face recognition and detection of moving objects.

The LBP operator mainly performs texture analysis on grayscale images. The LBP algorithm usually analyzes nine adjacent pixel rectangles and takes the center pixel as the standard so that the surrounding eight pixels are compared with the pixel value of the center point. If the value of the surrounding pixel is larger than the value of the central pixel, it is represented by 1, and if the value of the surrounding pixel is smaller than the value of the central pixel, it is represented by 027. The LBP operator is converted into a binary group of eight bits. For the convenience of calculation, the resulting eight-bit binary number is usually converted to a decimal number. The conversion process of the LBP algorithm is shown in Fig. 3.

Fig. 3
Fig. 3
Full size image

LBP algorithm conversion process diagram.

In Fig. 3, the pixel values in the original pixel are converted to 0 or 1, where 88, 32, 121, and 54 are converted to 0, and 157, 168, 187, and 246 are converted to 1.

Since there is no clear definition in the description of the image texture, the local texture in the action image of the football player can be regarded as the joint density of all gray-scale pixels in the local pixel. The local texture representation process is:

$$R=r({f}_{k},{f}_{1},{f}_{2},\cdots ,{f}_{n-1})$$
(6)

In Formula (6), \({f}_{k}\) represents the gray value of the local pixel center point, and \(R\) represents the local texture. \(\text{R}\) represents the radius of the neighborhood in the local texture analysis, and the symbol ^ denotes an estimated or approximated value of the local pixel’s gray intensity.

Subtract the pixel value of the center pixel from the pixel value of the point around the center in the local pixel to get:

$$R=r({f}_{k},{f}_{1}-{f}_{k},{f}_{2}-{f}_{k},\cdots ,{f}_{n-1}-{f}_{k})$$
(7)

So that the value of the center pixel and the difference between the surrounding points and the center point do not affect each other, Formula (7) can be converted into:

$$R\approx r({f}_{k})r({f}_{1}-{f}_{k},{f}_{2}-{f}_{k},\cdots ,{f}_{n-1}-{f}_{k})$$
(8)

In the LBP transformation, the difference between the surrounding pixels and the pixel value of the center point would be affected by the pixel value of the center point in essence, and the result would be affected if the center point value is too large or too small. Since \(r({f}_{k})\) represents the brightness of the local image, and the texture processing has nothing to do with the image’s brightness, \(r({f}_{k})\) can be discarded.

The converted pixels can be expressed as:

$$R\approx r({f}_{1}-{f}_{k},{f}_{2}-{f}_{k},\cdots ,{f}_{n-1}-{f}_{k})$$
(9)

Then the process of image conversion can be regarded as the process of seeking \({f}_{i}-{f}_{k}\), where the value range of i is \((1,n-1)\). Since a pixel’s brightness is invariant among local pixels, the representation of local textures is also invariant over grayscale translation.

Apply Formula (9) to complete the conversion process. :

$$R\approx r(g({f}_{1}-{f}_{k}),g({f}_{2}-{f}_{k}),\cdots ,g({f}_{n-1}-{f}_{k}))$$
(10)

In Formula (10), \(g()\) represents a binary conversion function.

The specific expression of \(g()\) is as follows:

$$g(s)=\left\{\begin{array}{c}1,s\ge 0\\ 0,s<0\end{array}\right.$$
(11)

Let each converted pixel value of the surrounding pixels have a set weight value, and the weight value is binary magnitude \({2}^{i}\), and the LBP code is obtained by adding the converted pixel point weights.

$$LB{P}_{n}={\sum }_{i=1}^{n-1}g({f}_{i}-{f}_{k})\cdot {2}^{i}$$
(12)

In Formula (12), LBP encoding converts binary codes to decimal numbers.

In the traditional LBP algorithm identification process, probability analysis is performed on the image converted by LBP. In the probability calculation, \(s\) is a random variable. If \(\text{E}[{\text{s}}^{2}]\) exists, the variance of the random variable \(\text{s}\) is given in Eq. (13), where \({\upsigma }^{2}\) represents variance and \(\text{E}[\text{s}]\) denotes the expected value.”

The variance is expressed as:

$$D(s)=E{\left|s-E(s)\right|}^{2}$$
(13)

In Formula (13), \(D(s)\) represents the variance and \(E(s)\) represents the expected value.

The expected value is expressed as:

$$E(s)={\sum }_{j=1}^{n}{s}_{j}{p}_{j}$$
(14)

Applying probabilistic analysis to the analysis of the technical movements of football players to analyze image sequences:

$${H}_{k}=\frac{1}{n}\cdot {\left[{\sum }_{j=1}^{n}d({E}_{j}^{k},{E}^{k})\right]}^{2}$$
(15)

In Formula (15), \({H}_{k}\) represents the variance of the \({k}^{th}\) image sequence, \(d\) represents the distance measure used in evaluating the similarity between features in the model.

All football players’ technical action image analysis has E sub-images, then:

$$E=\frac{1}{n}{\sum }_{j=1}^{n}{E}_{j}$$
(16)

Therefore, after analyzing the LBP algorithm through the probability variance, \({E}^{k}\) can be used to represent the average value of the local image and \({E}_{j}^{k}\) can be used to represent the kth sub-image of the jth football player. The recognition of similarity to the technical action behavior of football players is represented by \(d({E}_{j}^{k}-{E}^{k})\).

PCA-LBP algorithm

The analysis of the technical action characteristics of football players is the most difficult point in the whole recognition model. The LBP algorithm can describe the technical action information of football players well, but the data dimension is very high, and the calculation is complicated28. PCA is a linear combination recognition mode. According to the spatial distribution of picture pixels, the sample direction with the most significant variance is used as the standard, and other samples are projected onto it to extract feature vectors.

The PCA method was applied to reduce the high-dimensional dataset into a lower-dimensional space while preserving the most significant variance. The covariance matrix \(\Sigma\) was calculated as:

$$\Sigma =\frac{1}{m}\sum_{i=1}^{m}{({X}_{i}-\mu )({X}_{i}-\mu )}^{T}$$
(17)

where \(X\) is the dataset, and \(\mu\) is its mean. The eigenvalues \(\lambda\) and eigenvectors \(v\) were derived by solving \(\Sigma v=\lambda v\). The top \(k\) eigenvectors corresponding to the largest eigenvalues were selected to form the principal components, projecting the data into a reduced feature space \(Y\) as:

$$Y=XW$$
(18)

where \(W\) is the matrix of eigenvectors. This reduced feature space captures the key discriminatory features required to classify actions accurately.

Let the image set converted by LBP be \(S=\left[{s}_{1},{s}_{2},\cdots ,{s}_{n}\right]\), assuming that the data in the image set is divided into A types, each type of data has m pieces, and the samples after PCA projection are \(Y=[{y}_{1},{y}_{2},\cdots ,{y}_{n}]\), then the function with the most significant variance is expressed as:

$$\underset{y}{max}={\sum }_{i=1}^{n}{\Vert {y}_{i}-\overline{y}\Vert }^{2}$$
(19)

In Formula (19), \(\overline{y}\) represents the average feature vector of low-dimensional data.

Let the covariance of the soccer player technical action picture sample be denoted as \({S}_{a}\).

$${S}_{a}={\sum }_{i=1}^{n}({s}_{i}-\overline{s})({s}_{i}-\overline{s}{)}^{T}=S{S}^{T}$$
(20)

If the first \(c\) largest eigenvalues of \(S{S}^{T}\) are expressed as \({v}_{j}\), then the corresponding eigenvectors are expressed as:

$${x}_{j}=\frac{1}{\sqrt{{v}_{j}}}S\mu$$
(21)

In Formula (21), \(\mu\) represents the eigenvector corresponding to \(S{S}^{T}\).

Then, the feature space of the technical action behavior of football players is expressed as:

$$X=({x}_{1},{x}_{2},\cdots ,{x}_{c})$$
(22)

Project the sample into the feature space:

$$Q={X}^{T}({s}_{i}-\overline{s})$$
(23)

Videos were processed to isolate the football player’s technical action behavior and exclude non-relevant environmental factors. This involved filtering signals to retain only the desired activities, enhancing the model’s predictive accuracy. A process known as isolation of activity was established to ensure that the framework learns to identify football-related motion precisely. It planned the recordings for the exercise stage by separating the critical tasks from the small action intervals pointed out earlier. Small action monitoring is distinguished by the low magnitude and variability signals, as opposed to the signal behavior throughout a high-activity motion. It is particularly accurate when we concentrate on values of acceleration. The lower body accelerates rapidly whenever a football player jumps from a standing or comfortable position. This is why individual accelerometer signals were utilized during the activity isolation phase. During the study, you will observe this substantial acceleration switch between low and high activity periods in all tested body parts. Players adopt different initiation points for foot positioning when standing according to their preference. The system works equally well for tracking movements in all three directions \(X\), \(Y\) and \(Z\). Although jumping mostly creates vertical movement, the longitudinal movements become more obvious as players pass the ball. Identifying a transition from low to high movement level can be detected through any analyzed body part across all measurement axes. Measurements use the standard values from all three axis positions at each tracking point. Each sensor runs separately and gives its results during the conclusion of the operations.

The system identifies movement changes by determining the average signal value for each measurement component. When a player stands still, or moves about, the signal values typically remain lower than their average measurement. When the player performs physical work beyond their previous level, their signal pattern demonstrates values above their typical baseline value. Our algorithm identifies the beginning of a high-activity period by monitoring when the signal norm crosses its median threshold. Our algorithm selects 50 consecutive time steps of 0.1 s to find meaningful magnitude levels achieved in this domain. As we move the window backwards, we detect the end of a high-activity interval.

The system detected sprints and jogs when the mean passed through the defined threshold but worked best for shots and moves when using 1.5 times the mean as the threshold. Studying each activity type produced two main findings, leading us to name one activity type as monitoring compliance for regular activities and the other as exploding actions for single non-rehearsed events. Research shows that the IQR feature works well for identifying differences between explosive actions that happen just once and follow no pattern versus periodic movements that are done repeatedly. The analysis used the Euclidean distance method to test accelerometer signals and calculated the IQR afterwards. The graph proves the IQR shows strong results for differentiating one movement type from another when a threshold value exists for this method. The recording fell into the periodic motion category when the IQR measurement exceeded 0.12, yet it showed an explosive movement type below that value. Algorithm 1 presents all the steps required to separate movement types. It successfully recognized times of high movement from times of light activity.

Algorithm 1:
Algorithm 1:
Full size image

Activity isolation of football player’s technical action behavior using threshold.

Discover the initial timestep in which the signal and the next 50 timesteps exceed the previously established threshold. Consider that timestep the beginning of the action for that body part \(t\). In addition, gain the last timestep in which the signal and the preceding 50 timesteps exceed the before-established threshold. Consider that timestep to be the conclusion of the action for that body part t.

Take the smallest of the preceding step’s action starts. Subtract 250 timesteps, which marks the activity as a whole beginning. Assume this; take the maximum value from the action endpoints identified in the previous step. Add 250 timesteps (if possible). This marks the end of the action as a whole.

Experiment design of PCA–LBP algorithm recognition

Experimental data

It is essential to perform a statistical analysis of the relevant technical actions of football players to identify their technical action behavior effectively. For statistics, this experiment investigated 200 football-related people, including football players, coaches, and trainers29. The investigators on the technical movements of football are shown in Table 1.

Table 1 Football investigator form.
Full size table

In Table 1, it is described that there are three groups of people for the investigation of football technical movements, and each group of people is investigated separately for different genders. The maximum number of male football trainers is 44, and the minimum number of female football players is 24.

The everyday technical movements in football are obtained through the statistical investigation of these three types of football crowds. The difficulty of various technical movements is counted. The difficulty index is represented by the numbers 1–10. The statistical results of the leading technical actions in football are highlighted in Table 2.

Table 2 Technical action table of football.
Full size table

Table 2 lists the main technical movements of football and the difficulty of various movements. The most difficult action is heading the ball, followed by intercepting the ball, which has a difficulty index of 7. The most effortless technical football action is kicking the ball30.

This paper calculates statistics on the frequency of statistical football technical actions. The field’s most frequently used technical actions are extracted as an indicator for recognizing football players’ technical actions. The data for this statistic comes from the average data of players in a football league in 2019. The statistical results of the frequency of football technical actions are shown in Table 3.

Table 3 Data table of frequency of use of football technical actions.
Full size table

In Table 3, football players’ most frequently used technical action is kicking the ball, followed by dribbling. The actions of tackle and header are the least frequently used, accounting for 4% and 8%, respectively. Therefore, the four technical actions of kicking, dribbling, stopping, and fake action are used as indicators to evaluate the accuracy of technical action recognition31.

Experimental design

The experiment would compare and analyze recognition with the traditional LBP algorithm to explore the PCA–LBP algorithm’s recognition effect on football players’ technical action behavior.

The data comparing the two algorithms is based on data from 200 football players at a football match in 2020. This paper mainly counts the specific stadium information of football players, and the data samples of football technical action recognition are shown in Table 4.

Table 4 Sample table of football technical action recognition data.
Full size table

Table 4 describes the experimental data used to compare the two recognition algorithms. The most significant number of wingers is 12, and the fewest number of goalkeepers is 12.

The experiment would perform technical action recognition on the game images of football players in Table 4. Due to the different technical characteristics and body shapes of male and female athletes, comparative experiments on male and female players are necessary to ensure the experiment’s validity. By recognizing different technical actions, the difference in recognition accuracy between the two football technical action recognition algorithms was analyzed32.

Results and discussion of PCA–LBP algorithm identification

Recognition of football kicks

The kicking action in football is the most frequently used and the most basic technical action used by football players. The recognition model based on the PCA–LBP algorithm is compared with the traditional LBP algorithm recognition, and the identification data statistics of male players and female players are carried out, respectively33. The recognition test of players’ technical movements takes 50 sample data as one statistic. The comparison results of the two football players’ technical action behavior recognition algorithms in the kicking action are shown in Fig. 4.

Fig. 4
Fig. 4
Full size image

Recognition and comparison results of kicking action.

In Fig. 4a, the comparison of the recognition accuracy of two football technical action recognition algorithms for male players is depicted. When the number of recognitions is 50, the recognition accuracy rates of the two recognition algorithms are 82% and 84%, respectively, and the recognition accuracy of the two recognition algorithms is not much different. When the number of recognitions increases, the recognition accuracy of the LBP algorithm begins to decline. The recognition accuracy of the PCA–LBP algorithm would increase, and the recognition accuracy of the two recognition algorithms is 72% and 96%, respectively, when the number of recognition times is 300. Figure 5b compares the recognition accuracy of the two recognition algorithms for female players, and the overall change trend is similar to that in Fig. 4b. The recognition of the LBP algorithm would decrease when the number of recognitions increases, while the recognition accuracy of the PCA–LBP algorithm always maintains a high level. The recognition accuracy of the two recognition algorithms is 76% and 96%, respectively, when the number of recognition times is 300. Therefore, the PCA–LBP algorithm can improve the accuracy of kicking action recognition in football technology by reducing the dimension of the recognition data.

Fig. 5
Fig. 5
Full size image

Performance comparison of PCA-LBP and LBP algorithm.

Table 5 shows each action’s key metrics—accuracy, precision, recall, F1-score, and average latency. Results indicate that the PCA–LBP algorithm consistently outperforms the traditional LBP algorithm in terms of accuracy, precision, recall, and F1-score across all actions. For instance, in kicking actions, PCA–LBP achieves an accuracy of 96%, while LBP reaches only 82%. Additionally, PCA–LBP demonstrates lower latency, suggesting improved computational efficiency and better suitability for real-time applications.

Table 5 Performance metrics for PCA-LBP and traditional LBP algorithms.
Full size table

Figure 5 compares each action’s accuracy, precision, recall, F1 scores (%) between the PCA–LBP and LBP algorithms. The results indicate that the PCA–LBP algorithm achieves consistently higher accuracy across all actions, demonstrating its enhanced performance in correctly recognizing each type of action compared to the LBP algorithm.

The PCA–LBP algorithm consistently achieves higher precision across all actions, indicating its superior ability to reduce false positives compared to the LBP algorithm. The PCA–LBP algorithm consistently achieves higher recall across all actions, demonstrating better performance in correctly identifying relevant actions than the LBP algorithm. The PCA–LBP algorithm consistently achieves higher F1-scores across all actions, indicating improved accuracy and balance between precision and recall in action recognition compared to the LBP algorithm.

Figure 6 compares latency (in milliseconds) between the PCA–LBP and LBP algorithms for different football actions. The results indicate that PCA–LBP generally has lower latency, enhancing its suitability for real-time applications.

Fig. 6
Fig. 6
Full size image

Latency comparison of PCA-LBP and LBP algorithm.

All the figures visualize these metrics through bar charts, highlighting the performance differences between PCA–LBP and LBP for each action. Each metric is displayed on a separate plot, allowing for a clear comparison. Notably, the PCA–LBP algorithm maintains higher metric scores across all actions, particularly in more complex movements like dribbling and fake actions, where the LBP’s performance shows a significant drop. The latency chart further emphasizes PCA–LBP’s advantage, with consistently faster processing times across actions, supporting its potential for real-time implementation.

Recognition of football dribbling actions

In football, dribbling technical movements can drive players’ movements on the court to play tactical cooperation. For male and female players, the technical movements of dribbling are recognized, respectively, and the comparison of the two recognition algorithms can be more comprehensively analyzed by analyzing the dribbling of different types of players34. Among them, traditional technical action recognition adopts LBP algorithm recognition, and the improved recognition algorithm is PCA–LBP recognition. The comparison results of the two recognition algorithms for the recognition of football players’ dribbling actions are shown in Fig. 8.

In Fig. 7a, two recognition algorithms are used to identify and compare the dribbling skills of male football players. Among them, the recognition accuracy of the LBP algorithm decreases continuously after the number of recognitions increases, and the initial 82% accuracy drops to 56%. However, the recognition accuracy of the PCA–LBP algorithm does not change much, and the accuracy is 92% when the number of recognition times is 50. In Fig. 7b, the dribbling skills of female football players are identified. The accuracy of the PCA–LBP recognition algorithm is significantly higher than that of the LBP algorithm. When the number of recognitions reaches 300, the two algorithms have 90% and 64% accuracy in recognizing the dribbling movements of female players, respectively. Therefore, the PCA–LBP algorithm has a higher recognition accuracy than the LBP algorithm in recognizing the dribbling technique of football players of different genders.

Fig. 7
Fig. 7
Full size image

Comparison results of recognition and comparison of dribbling actions.

Fig. 8
Fig. 8
Full size image

Recognition and comparison results of the ball stop action.

Recognition of football stop action

Football stops have many types, such as foot, thigh, chest, etc. The stop action has one thing in common so that the football can stop. The PCA–LBP and LBP algorithms are compared for the action recognition of the ball stop technique, and the recognition times increase. The recognition accuracy of the two algorithms under different recognition times is observed35. In addition, the bodies of female and male players are different, which would also interfere with the poor recognition accuracy. It is necessary to identify and analyze players of different genders separately. Figure 9 shows the comparison results of the recognition accuracy of the two algorithms for the football stop action.

Fig. 9
Fig. 9
Full size image

Comparison result of recognition of fake action.

In Fig. 8a, two algorithms identify the stop action of male football players. When the number of recognitions is 50, the recognition accuracy of the two algorithms is the same, 92%. However, when the number of recognitions continues to increase, the recognition accuracy of the LBP recognition algorithm for stopping the ball continues to decline, reaching a minimum of 66%. In contrast, the recognition accuracy of the PCA–LBP algorithm does not change much. In Fig. 8b, the stop action of the female player is identified, and it can be seen that the identification accuracy of the LBP algorithm has been decreasing. When the number of recognitions reaches 250 times, it decreases to the minimum, and the recognition accuracy is 62%. However, the recognition accuracy of the PCA–LBP algorithm would not drop significantly due to the increase in the number of recognitions. The overall average accuracy of the PCA–LBP algorithm is 86%. Therefore, the PCA–LBP algorithm has higher stability and recognition accuracy in recognizing the ball-stopping action.

Recognition of football fakes

Football fake is a technique that confuses opponents and is a more advanced action on the football field. The PCA–LBP and LBP algorithms were compared to recognize fake football action. This paper counts the recognition accuracy of the two algorithms for football fake action, and the comparison results are shown in Fig. 9.

Figure 9a is the recognition of a male football player’s fake action; it can be seen intuitively that the recognition accuracy of the PCA–LBP algorithm is higher than that of the LBP algorithm. The difference between the two recognition accuracies is more evident with the increase in recognition times. When the number of recognitions reaches 250, the difference in recognition accuracy between the two algorithms is at most 22%. Figure 9b shows the recognition result of the female player’s fake action. The recognition accuracy of the LBP algorithm is higher than that of the PCA–LBP algorithm only when the number of recognition times is 100, and the accuracy of the two algorithms is 80% and 82%, respectively. With the increase in the number of recognitions, the recognition accuracy of the two algorithms decreases to varying degrees. Among them, the accuracy of the LBP algorithm decreases more obviously. When the number of recognitions is 300, the accuracy of the LBP algorithm is 58%. The PCA–LBP algorithm has higher recognition accuracy in the football fake action recognition process.

To assess the robustness and generalizability of the proposed PCA–LBP model, we conducted a k-fold cross-validation process. This technique divides the dataset into k equal parts (folds), using k − 1 parts for training and the remaining part for validation. The process is repeated k times, with each fold being the validation set once. The average performance across all folds provides a more reliable estimate of the model’s accuracy, precision, recall, and F1 score. Additionally, to evaluate the model’s generalization capabilities, we tested it on an external dataset collected from different matches. This testing phase aimed to simulate real-world variations in player behavior, environmental conditions, and match dynamics, providing insights into how well the model performs beyond the initial dataset.

Figure 10 shows the accuracy scores across each fold in the tenfold cross-validation. The consistency of scores across folds demonstrates the model’s stability and robustness in recognizing actions accurately.

Fig. 10
Fig. 10
Full size image

Cross-validation accuracy performance.

Figure 11 displays precision scores for each fold in the tenfold cross-validation. Precision stability across folds indicates the model’s effectiveness in minimizing false positives.

Fig. 11
Fig. 11
Full size image

Cross-validation: precision performance.

Figure 12 illustrates recall scores across the 10 folds, reflecting the model’s sensitivity and ability to correctly identify relevant actions without missing instances.

Fig. 12
Fig. 12
Full size image

Cross-validation: recall performance.

Figure 13 compares accuracy, precision, and recall when testing the model on an external dataset, providing insight into its generalizability and reliability in new conditions beyond the initial dataset.

Fig. 13
Fig. 13
Full size image

Performance on external dataset: accuracy, precision, and recall.

Table 6 compares the performance of the PCA–LBP algorithm from our study with various action recognition models, including traditional LBP, CNN, RNN, Hybrid CNN-RNN, and 3D-CNN models.

Table 6 Comparison of PCA-LBP algorithm with other action recognition models.
Full size table

The accuracy, precision, recall, F1-score, and average latency metrics highlight each model’s effectiveness and efficiency. The PCA–LBP algorithm demonstrates competitive accuracy (90%) and precision (89%) with low average latency (27 ms), making it well-suited for real-time applications. In contrast, advanced models like the Hybrid CNN-RNN (94% accuracy) and 3D-CNN (91% accuracy) achieve higher accuracy but have increased latency, suggesting that while they excel in precision and recall, they may be less feasible for real-time use due to their computational demands.

Conclusions and future work

This paper investigates the application of deep learning techniques, particularly the PCA–LBP algorithm, for recognizing football players’ technical actions. The study highlights the significant potential of the PCA–LBP model, which outperforms the traditional LBP algorithm regarding recognition accuracy, stability, and computational efficiency. By analyzing the four most frequently used technical actions—kicking, dribbling, stopping, and fake actions—this research demonstrates the effectiveness of dimensionality reduction via PCA in enhancing recognition performance while maintaining computational simplicity. Despite these promising results, the dataset’s size limits the study, comprising 200 samples from a single match. This limited sample may not fully reflect the diverse player behaviors and demographics range. Future research will address this limitation by incorporating a more extensive and varied dataset, encompassing players from different age groups, skill levels, and field positions. Expanding the dataset will improve the model’s generalizability and ensure it can be applied to diverse real-world scenarios. Additionally, future work will include recognizing more complex and varied technical actions, such as passing, tackling, and heading, to enhance further the model’s applicability in comprehensive player training and performance analysis. These advancements aim to solidify the PCA–LBP framework as a robust tool for sports analytics, contributing to more effective and scientific training methods for football players.