Football teaching and training based on video surveillance using deep learning

Abstract

BACKGROUND:

The objective performance evaluation of an athlete is essential to allow detailed research into elite sports. The automatic identification and classification of football teaching and training exercises overcome the shortcomings of manual analytical approaches. Video monitoring is vital in detecting human conduct acts and preventing or reducing inappropriate actions in time. The video’s digital material is classified by relevance depending on those individual actions.

OBJECTIVE:

The research goal is to systematically use the data from an inertial measurement unit (IMU) and data from computer vision analysis for the deep Learning of football teaching motion recognition (DL-FTMR). There has been a search for many libraries. The studies included have examined and analyzed training through profound model construction learning methods. Investigations show the ability to distinguish the efficiency of qualified and less qualified officers for sport-specific video-based decision-making assessments.

METHODS:

Video-based research is an effective way of assessing decision-making due to the potential to present changing in-game decision-making scenarios more environmentally friendly than static picture printing. The data showed that the filtering accuracy of responses is improved without losing response time. This observation indicates that practicing with a video monitoring system offers a play view close to that seen in a game scenario. It can be an essential way to improve the perception of selection precision. This study discusses publicly accessible training datasets for Human Activity Recognition (HAR) and presents a dataset that combines various components. The study also used the UT-Interaction dataset to identify complex events.

RESULTS:

Thus, the experimental results of DL-FTMR give a performance ratio of 94.5%, behavior processing ratio of 92.4%, athletes energy level ratio of 92.5%, interaction ratio of 91.8%, prediction ratio of 92.5%, sensitivity ratio of 93.7%, and the precision ratio of 94.86% compared to the optimized convolutional neural network (OCNN), Gaussian Mixture Model (GMM), you only look once (YOLO), Human Activity Recognition- state-of-the-art methodologies (HAR-SAM).

CONCLUSION:

This finding proves that exercising a video monitoring system that provides a play view similar to that seen in a game scenario can be a valuable technique to increase selection accuracy perception.

Keywords

Football training deep learning video surveillance skills detection

1. Introduction

1.1 Overview of football teaching and training based on video surveillance

Objective performance, identification, and tracking have evolved as critical topics in deep Learning (DL), with applications in various fields, including automatic identification, classification of football teaching, and training exercises [1]. The most obvious advantage of employing video analysis in practice is providing players with immediate feedback. Video allows coaches to identify and correct flaws quickly. Coaches can use video to monitor their players’ progress and skill development [2]. Athletes use video analysis to assess their performance to enhance their abilities and avoid injury. When filming a player’s performance, you can catch small details and techniques that overlook watching them live. The sports business likewise embraces new technologies for capturing and evaluating athletes’ performance [3]. It is critical to use video surveillance to help keep fans, players, workers, and the facilities themselves secure. In recent years, deep learning (DL) and teaching listening have gotten much attention in football. DL becomes connected with football because it predicts match outcomes [4]. A player must check the analyzed video regularly to avoid bad habits from the past and keep consistency in any acts that have been remedied. During video analysis, comparison video can be employed [5]. According to research, gymnasts who compared their video analysis to professional gymnasts performing the same skills improved faster than they would have with only standard practice and instruction. Deep learning algorithms can forecast the success or failure of football games based on the analysis of large amounts of data [6, 7]. Due to the nature of camera settings utilized to capture the full football action, player detection is one of the most challenging tasks. Ground truth annotations are critical for successful player detection using any deep learning-based model [8, 9]. There are usually 22 players from two teams, the essential targets spotted in each football match, and they are captured using Stat metric camera setups. Educational leaders can employ technology to improve players’ capabilities and offer more value to their teaching [10]. In sporting situations, computers are increasingly used to simulate the dynamics of sports. Wearable sensors have piqued interest recently due to their potential for monitoring users’ health, fitness, and surroundings. In addition, virtual environments enable sports practice in remote locations and transfer habits to real-life settings [11, 12].

Recent breakthroughs in sports events are captured using single, multiple, and customized camera setups. As a result, football managers find it challenging to design effective plans to counter their opponents [13]. Viewing a moment in multiple surveillance cameras means looking at different settings and looking for inappropriate actions or the emergence or existence of inappropriate behavior [14]. Video surveillance is designed to monitor the public during sports events and outside of secure organizations, particularly those directly bordered by communal spaces. Identifying Concerned locations and particular cameras or groups of cameras capable of viewing such regions is part of the video surveillance process [15, 16]. As a result, seeing the chosen photographs at the proper periods becomes possible. Any assembly of many individuals is a potential target for criminal activity. Sporting events are an excellent illustration [17]. Authorities must ensure that such events are safe, the people are kept under control, and any threat of large-scale disruption, such as terrorism, is avoided [18]. Skilled evaluators are utilized to post-process the captured sports event data to pinpoint crucial acts and track individuals. They offer critical data for coaches to study the team’s strategy and provide individual player ratings on their performance during games [19]. The conventional methodology is time-consuming due to the repeated nature of the technique and the requirement to annotate nearly all of the frames. Some of the functions performed by human annotations can be automated, speeding up the process, thanks to developments in computer vision data (CVD) and Deep Learning (DL) [20].

For many years, performance analysis of sporting events, such as football, has been done, and many innovative methodologies are used to obtain quick and precise findings [21]. Player tracking, ball tracking, and action detection are immediate actions that must be annotated in a football match. The nature of the game and the camera utilized to capture the game make these duties challenging [22]. The study presents a dataset that combines various components and gives a performance ratio of 94.5%. The review highlights challenges and opportunities for innovation in e-healthcare risk prediction systems. Many studies have been conducted on recognizing numerous objects of varied forms and sensitivities. As a result of recent breakthroughs in AI, many deep-learning trackers have been built. Most monitoring algorithms employ a score of temporal features matching as a significant criterion, along with other measures such as Intersection of Union, Neural Network predicting, and multi-separation, to produce tracking results. When objects are spatially distinguishable, these strategies function well [23]. When the things being monitored are spatially similar, and their locations overlap, multi-object tracking becomes difficult. Due to the nature of football, the players to be monitored wear the same uniform, and player overlapping (same team or opposite team) is joint and can last for several frames. To create fewer swaps and switches while keeping high confidence in the short term, leverage overlap between all detected bounding boxes and feature matching to only candidates of the individual track [24].

The main contribution of the paper is,

Video surveillance is essential for recognizing and preventing or minimizing undesirable human behavior in real-time. The digital content of the video is categorized according to its relation to those specific acts.

The study combines the inertial measurement unit (IMU) and computer vision data processing systematically for deep Learning of football teaching motion recognition (DL-FTMR).

The findings are interpreted as proof of exercising using a video monitoring system. It provides a play view similar to that seen in a game scenario, which can be an essential part of improving selection accuracy perception.

The paper is structured as follows: Section 1.1 introduces football teaching and training based on video surveillance. Section 2 discusses related works and corresponding debates. Section 3 elaborates on the DL-FTMR, a systematic approach to using inertial measurement unit (IMU) and computer vision data analysis for football games. Section 3.2 explains deep learning technologies for playing football games, while Section 3.3 outlines the architecture of the projected IMU. Section 3.4 focuses on football teaching motion recognition, while Section 3.5 discusses multi-action detection and recognition. Section 4 presents the results and compares them to existing methods. Finally, Section 5 concludes the paper based on the analysis presented in the previous sections and discusses future research directions.

2. Literature review

This study executed a comprehensive literature work, and the most significant works among them are summarized below:

Song et al. [25] discussed a model to provide successful diagnosis and risk assessments of sports-medicine disorders, an optimized convolutional neural network (OCNN) based on a deep-learning model. It used SAR (Self-Adjustment Resizing) combined with a convolutional neural network (SCM). Two convolutional layers, two pool layers, a fully connected layer, and a softmax structure made up the proposed OCNN classification, which could be utilized to categorize sport-related medical data. Experiments have shown that this strategy could give technical support and a roadmap for developing a specific cloud-based fusion system.

Girdhar et al. [26] deliberated an exhaustive and complete analysis of state-of-the-art methodologies (SAM). They discussed the identification of problematic regions of the HAR ecosystem named Human Activity Recognition-state-of-the-art methods (HAR-SAM). The current evaluation effectively examined trends in research outflowing using the vision-based approach. This study presented a survey of the HAR training datasets currently available and discussed the most popular publicly available datasets. Readers could benefit from thoroughly examining the substantial work done at HAR and its vision-based approaches.

Gochoo et al. [27] described the Gaussian Mixture Model (GMM) and a saliency map to extract a human silhouette. Then, a single pseudo-2D stick model was employed for calculating adaptive posture and detecting human body parts. Further, a single dataset was created by combining the energy, 3D Cartesian view, angular geometric, skeleton zigzag, and moveable body components. The UCF aerial action dataset, the UT-interaction dataset, and sports videos in the wild (SVW) were used to identify complex events.

Feddersen et al. [28] discussed how two high-level fencing and football coaches regarding issues developed during a three-month coach training program could be handled. This study aimed to assist them in providing psychological skills training to athletes. According to the findings, the trainers needed help to define their duties. Coaches needed help knowing whether to include psychological skills training in their programs and refer athletes to sports psychology experts due to the psychological effects.

Bamunif et al. [29] introduced the minimax algorithm to compute each game utility calculation from the game tree. The report started with a schematic showing how two viewpoints come together. It recognized a subset of AI applications utilized in games nowadays. According to the investigation, huge fans were the primary beneficiaries of AI development jobs. Furthermore, sports-related organizations are rapidly expanding due to the high level of care provided by technology. Because of the correct reforms in the appropriate directions, the future will be even more secure. As a result, it’s possible that staying rejuvenated is beneficial for sports organizations.

Fu et al. [30] discussed the basketball court videos as inputs in the suggested approach. Then, using real-time object detection, the You Only Look Once (YOLO) model was employed to locate the position of the basketball hoop. After that, motion detection based on frame difference was used to detect an object moving in the hoop region to determine the basketball scoring condition. Furthermore, numerous intelligent basketball analysis systems based on the proposed method were installed at many Beijing and proven applicable basketball courts.

Falloon et al. [31] aimed to improve student’s skills in using ‘educational’ apps and digitally sourced material and successfully understanding blends of pedagogical content. This article introduced a conceptual framework for a broader understanding of Digital teacher skills (TDC). It demanded a systemic and thorough knowledge that acknowledges the increasingly diverse expertise and skills that young people need in a wide range of digital media environments to work with legal, safety, and production.

Xiao et al. [32] designed an Internet-based intelligent sports wristband, a tracking device to detect the user’s changes in heart rate. It was intended to address the drawbacks of present healthcare systems’ huge sizes that are difficult to wear. Data were sent to cellular phones via the Internet of Technology Services, ZigBees wireless sensors, Bluetooth, and other networking technologies for real-time tracking, storage, and review. The Internet was used to relay data via communication network technologies to the monitoring system platform. After processing and timed data review for heart problems monitored by mobile staff in the movement process, abnormal data were alerted.

Bin et al. [33] described a system that consists of three main subsystems. It provided a lightweight and inventive control system plan depending on the Bluetooth sensor. Sensors focused on Bluetooth devices, wireless communications terminals, and sportspeople’s physical data acquisition systems. This model proposed an innovation choice for producing Bluetooth-related correspondence-dependent products in space. The system was low-cost and flexible, allowing several devices to monitor small changes at the heart of this system. Password authentication was used so approved users could only access the computer from home.

Poulose et al. [34] Research in human activity recognition (HAR) has gained importance in healthcare systems. Conventional HAR systems use wearable sensors but must be more accurate with complex activities. To address this, researchers proposed a human image threshing (HIT) machine-based HAR system that uses an image dataset from a smartphone camera for activity recognition. HIT machine uses a mask region-based convolutional neural network (R-CNN) for human body detection, a facial image threshing machine (FIT) for image cropping and resizing, and a deep learning model for activity classification. The proposed HIT machine achieved 98.53% accuracy when the ResNet architecture was used as its deep learning model.

Poulose et al. [35] Advances in deep Learning have made it possible to develop complex models that can be applied in various fields. Human activity recognition (HAR) is one area where the goal is to improve predictive accuracy for devices with limited computational resources. Recent research proposes iSPLInception, a deep learning model that achieves high predictive accuracy but uses fewer device resources. It outperforms existing models on several accuracy metrics, cross-entropy loss, and F1 score on four public HAR datasets. The proposed model’s performance is validated on the UCI HAR using the smartphone, Opportunity activity recognition datasets, Daphnet freezing of gait datasets, and PAMAP2 physical activity monitoring datasets.

Based on the survey through some existing OCNN, HAR-SAM, GMM, and YOLO methods, several parameters have been considered: behavior analysis, performance ratio, athletes’ energy level, F1 Measure, sensitivity ratio, etc., interaction ratio, and specificity rate are the significant aspects. Therefore, this research presents deep Learning of football teaching motion recognition (DL-FTMR). Here, there has been a hunt for several libraries. The studies have examined and studied training using advanced model construction techniques, which helps overcome the abovementioned conventional challenges.

3. DL-FTMR has been utilized systematically to analyze the inertial measurement unit (IMU) and computer vision data

This paper discussed the DL-FTMR used to systematically analyze computer vision data and employ the inertial measurement unit (IMU). Deep Learning is used to build a fully automated sports performance that appears nearly comparable to a professional football presentation, including movement, camera zoom-ins on the action, etc. A decently automated sports program should determine the athletes and the football team. An inertial measuring unit (IMU) is an electronic device that detects and reports a body’s special force, angular rate, and sometimes orientation using a combination of accelerometers, sensors, and sometimes magnetometers [36]. In particular, inertial measuring units (IMUs) have played an increasingly important role in human movement analysis, allowing for measuring various movement-related metrics outside of instrumented laboratories. Based on these features, IMUs have been applied to quantify balance and gait metrics in pathological subjects to determine disease progression and therapeutic effects. Considering motion sensors in numerous sports tracking technologies, inertial measurement units (IMU) have grown in the previous decade. An IMU is a set of sensors that collect data based on the unit’s movement, including an accelerometer, gyroscope, and magnetometer. Inertial measuring equipment collects raw data, which many sports technology businesses enhance using algorithms and in-depth computations to provide understandable results. The algorithms and software applied to the raw data from the IMU give real value to the end-user, even if the hardware is very inexpensive to produce. Computer vision data analysis (CVDA) is a branch of automation that focuses on developing methods for teaching computers to interpret and comprehend the contents of videos. This might be applied to videos because a video is essentially a series of sequential images or frames. Deep learning models correctly recognize and classify items in a constantly changing physical environment. Computer vision tries to reproduce some of the intricacies in the human visual system and visual processing.

Figure 1.

Framework for recognizing sports activity.

Figure 1 depicts how athletic sports action identification has gotten a lot of interest in computer vision data analysis because of its wide range of applications. It tackles problems in various domains, including video surveillance, human-computer interface, and rehabilitation. Despite a lot of research, there are still a lot of issues with action recognition. Variations in viewpoint, occlusion, subject body size, spatiotemporal placement of actions, and inter-and intra-class variance are all factors to consider. The Computer Vision community has been captivated by recent advances in activity recognition from sports videos.

On the other hand, the problem of activity recognition from football video sequences is relatively unexplored in the literature. This research aims to develop a sports activity detection model based on deep Learning. As a result of the recent breakthrough of deep learning methods in image processing, researchers are encouraged to use deep learning methods to find dynamic video frames. Some studies have attempted to convert posture information into coloring or texture space and use the produced images as a deep learning input in this method. Based on their findings, the authors suggest that comparable results can be attained using only a portion of joint-line distance data. The proposed technique is compared to those methods, even though deep representation learning is not used in this work. The framework achieves equivalent and competitive performance according to the outcomes. With the evolution of deep understanding, there has been an increase in the desire for greater specificity in distinguishing and analyzing sports’ daily activities. Furthermore, research into applications such as monitoring athletics and sports activities to identify suspicious people and commodities left in public places continues. Wearable sensors have been created to recognize athletic motions, and the promise of machine learning and profound understanding has been clear.

Due to the obvious way the game is played, players’ directions vary frequently, and their actions modify the bounding box dimensions, resulting in Kalman filter prediction ( $m_{n}$ ) errors. As a result, assigning based just on Bayesian predictions is not possible. Compared to Kalman filter predictions, the method gives more weight to preceding frame detection, i.e., the bounding box of accessible tracks. As a result, It presents the gating distance (AB) for each assignment, which provides candidates ( $c$ ) for each track with a high probability. Because c is a subset of $D$ , it must be assigned to the ways.

$\displaystyle\textit{AB}\left({C,D}\right)=\sqrt{\left({xy_{z}-xy_{p}}\right)^% {2}+\left({ij_{z}-ij_{p}}\right)^{2}}$ (1) $\displaystyle m_{n}\left(D\right)=\max\left({a_{b}c_{d}}\right)$

Equation (3) shows the AB and denotes the Euclidean distance between the center ( $xy_{z}-xy_{p}$ ) of the bounding box of a given track ( $t$ ) and a detection point ( $a$ ). ( $a_{b}c_{d}$ ) is the symbol for the bounding boxes. The Gd is the maximum value between a track’s bounding box’s width and height. Equation (3) calculates the number of detections attributed to a path.

$\displaystyle T\left(\textit{mission}\right)\left({\tau,i}\right)=a_{n}\left\{% {\frac{\textit{AB}\left({k,i}\right)\leqslant m_{n}\left(d\right)}{0}}\right\}$ (2) $\displaystyle T\left(\textit{mission}\right)\left(\tau\right)=\left\{{a_{1}a_{% 2}\ldots\ldots.a_{n}}\right\}\in B$

Equation (3) creates a list of possibilities with a high likelihood of matching a specific track inside the database. Create Feature matching scores (FM) and Track overlap scores ( $\tau$ ) to find a person who fits away. Figure 2 shows the Split Feature matching approach; calculate the spatial similarity of bounding boxes.

$\displaystyle\textit{track area}=\left({\tau,i}\right)\left({\textit{weight}% \ast\textit{height}}\right)$ (3) $\displaystyle\textit{candidate area}=\left({c,i}\right)\left({\textit{weight}% \ast\textit{height}}\right)$

Equation (3), using a Split Feature matching approach, calculates the spatial similarity of two bounding boxes, the track $\left(t\right)$ and the person detection $\left(c\right)$ . Furthermore, when two or more players overlap, the bounding box dimensions alter from when only one person is present. Calculate the overlap area between the track and candidate detection to quantify the similarity of the bounding box area, Area Confidence (Ac), to use this feature in the tracking.

Figure 2.

Split Feature matching approach, calculate the spatial similarity of bounding boxes.

Equation (3) is utilized as a criterion for determining which candidates the user should track. The spatial correlations for all $\left\{{a_{1}a_{2}\ldots}\right\}$ and the feature matching result determine the persons under consideration.

$\displaystyle A_{b}\left({\tau,i}\right)=\left\{\begin{array}[]{l}{1,\textit{% if}\ \frac{\min\left({\textit{track area, candidate area}}\right)}{\max\left({% \textit{track area, candidate area}}\right)}>0.75}\\ \textit{otherwise}\\ \end{array}\right\}$ (4) $\displaystyle\textit{FM}\left({\tau,i}\right)=m^{\textit{mxm}}\ast A_{b}$

Track confidence ( $\tau$ ), (FM) helper function, as shown in Eq. (3), finds a player assigned to a track with no bounding box nearby. The low matching scores of the feature matching model can be attributable to two issues. Even though detection should have been an excellent match to a specific track, feature matching scores are predicted when player detection is not compared correctly with the bounding box details contained in existing ways

3.1 Deep learning technologies to play football games

Most classic deep-learning approaches focus on developing skilled experts who can act in complex domains. The experts’ nature can be based on wording such as consistency and ampleness. All aspects of validity are one of the fundamental variables of viable DL, where all of the space of the DL and PC games and open play are essential. The study of deep Learning (DL) in video games has a long history. It looks into how deep learning technology can achieve human-level gaming performance. It focuses on the complex interactions that occur between agents and gaming environments in general. Video games are suitable environments for DL research because they present agents with exciting and challenging challenges to solve. These online communities are safe and easy to control. As a result, a successful DL specialist for the game should be ready to demonstrate distinct playstyles to persuade a fun opponent to battle. In addition, if a virtual PC-controlled player has a genuine model, the DL structure should be prepared to mimic their playing style. Show how it used a technique to generate a sensible and believable expert for a 3D football PC game in the most recent research post. The Deep Learning system combines learning through discernment and case reasoning pathways. The PC system first alerts an athlete’s expert, who then demonstrates the required leadership by playing the game and following the knowledge base’s instructions.

Figure 3.

IMU architecture.

3.2 The projected inertial measurement unit’s (IMU) architecture

Figure 3 depicts interior placement with the suggested IMU effective process, divided into football coaching, offline training, and online location approach. IP (Internet Protocol) cameras take videos of many people and send them to a server during the offline phase. These videos construct a model’s training dataset that detects multiple users using a deep learning method. When a user submits a positioning request, cameras take real-time images of the user and send them to the server, which uses the trained model to recognize the Athletes in the picture. In the meantime, the user end transmits the server its own IMU’s projected pose estimations. The fusion filter combines. The suggested DL-FTMR technique combines the IMU’s predictions with the user’s ranges and angles relative to these cameras. Finally, the user is redirected to the revised estimations and sensor biases. Integrating IMU and video data for training neural networks involves synchronizing the data streams, extracting features from each modality, fusing the features, training models, evaluating performance, and deploying the model for activity recognition. The process enhances the understanding of observed activities by leveraging complementary information from both sensor modalities.

Figure 4.

IMU detection error models.

The most popular IMU detection error models are illustrated in Fig. 4. The following equations validate this figure statistically.

$\displaystyle\delta\rho=-\omega_{ab}\times\delta\rho+\delta\mu,$ $\displaystyle\delta\mu=-\left({2\omega_{ab}+\omega_{cd}}\right)\times\delta\mu% -\delta\varphi\times f+\delta f^{b}$ $\displaystyle\delta\varphi=-\left({\omega_{ab}+\omega_{cd}}\right)\times\delta% \varphi+\delta\varphi^{b}_{ab}$ (5)

As Eq. (3.2) shows, IMU detection error models and the variable define the $\delta\mu,\delta\varphi,\delta\rho$ refer to the coordinates system position, velocity, and orientation distortion parameters, respectively. The angular rate of the technological characteristics of the athlete’s position is denoted by $\omega_{ab}$ ; the athlete’s angular rate is denoted by $\omega_{cd}$ , and the special force is denoted by. $\delta\varphi^{b}_{ab}$ Drifts and accelerometer distortions are represented by $f^{b}$ and $\delta f^{b}$ , correspondingly.

A dimensional state vector is defined as follows for any single user $i$ :

$\displaystyle y_{xi}=\left[{\delta pu_{i},\delta v_{ui}^{n},\varphi_{ui},% \delta\omega_{ui}^{b},\delta f_{ui}^{b}}\right]^{T}$ (6)

Equation (6) describes the $y_{xi}$ as the conditional probability model, $\delta pu_{i}$ as the control model, and $\delta v_{ui}$ . Gaussian normal distribution with zero means is believed to be the source of the process noise vector. The integrated system’s measurements include users’ ranges and angles of their visible anchors. A simple model is created with a specific vendor and an attachment. The user $d_{u_{i}}$ ranging function $\propto u_{ij}$ and angulation function $\theta n_{ij}$ are related to anchor $u_{i}$ and are as follows in Eq. (7),

$\displaystyle d_{u_{i}}=\sqrt{\Delta e^{2}_{ij}}+\Delta n_{ij}^{2}+\Delta u_{% ij}^{2}$ (7) $\displaystyle d_{u_{i}}=\left[{\frac{\propto u_{ij}}{\theta n_{ij}}}\right]=% \left[{\frac{\frac{\Delta u_{ij}}{\Delta v_{ij}}}{\sqrt{\Delta n_{ij}^{2}+% \Delta u_{ij}^{2}}}+\delta pu_{i}}\right]$ (8)

As Eqs (7) and (8) define, the $\propto u_{ij}$ , $\theta n_{ij}$ , and $\delta pu_{i}$ Remain observational distortions derived from a Gaussian normal distribution with a zero mean. $\Delta e=\Delta u_{ij}$ , $\Delta_{Vij}$ is the difference in user $d_{u_{i}}$ locations in IMU synchronizes concerning anchor $u_{i}$ . The $u_{i}$ observation model with $n_{i}$ is created by:

$\displaystyle u_{ij}n_{ij}=\left[{\frac{d^{\text{FTMR}}_{u_{ij}}-d^{\text{IMU}% }_{U_{ij}}}{d^{\text{FTMR}}_{n_{ij}}-d^{\text{IMU}}_{n_{ij}}}}\right]$ (9)

Where $d^{\text{FTMR}}_{u_{ij}}$ and $d^{\text{FTMR}}_{n_{ij}}$ represent ranges and angles estimated using the FTMR method, and $d^{\text{IMU}}_{u_{ij}}$ and $d^{\text{IMU}}_{n_{ij}}$ Represent the user’s IMU’s expected range and orientations. Several anchors can increase an athlete’s measurement ability and an announcer’s communication ability. Assume $u_{ij}n_{ij}=\frac{d^{\text{FTMR}}_{u_{ij}}}{d^{\text{IMU}}_{n_{ij}}}$ is the number of cameras that can view user $u_{ij}$ at the same time, and the observation model is extended to take in Eq. (10),

$\displaystyle u_{ij}n_{ij}=\left[\begin{array}[]{c}\frac{d^{\text{FTMR}}_{u_{% ij}}}{d^{\text{IMU}}_{n_{ij}}}\\ \frac{d^{\text{FTMR}}_{u_{ij}}}{-d^{\text{IMU}}_{n_{ij}}}\\ \vdots\\ \frac{d^{\text{FTMR}}_{u_{ij}}}{d^{\text{IMU}}_{n_{ij}}}\\ \frac{d^{\text{FTMR}}_{u_{ij}}}{-d^{\text{IMU}}_{n_{ij}}}\\ \end{array}\right]$ (10)

Extend the dynamical and observation models to a group of users $u_{ij}=[d^{\text{FTMR}}_{u_{ij}}]$ and a group of cameras $n_{ij}=[d^{\text{IMU}}_{n_{ij}}]$ in an IMU/ FTMR combined organization. Finally, determine an initial video distance for each camera by doing the activation.

Figure 5.

Detection method for football pitches and players.

3.3 Football teaching motion recognition

This section explains the proposed strategy, divided into four phases, as shown in Fig. 5. It extracts recognized and tracked humans to build primers from the videos in the first stage (i.e., the preprocessing phase). The proposed Deep learning design then includes both 3-dimensional and 2-dimensional networks. The 3D counterpart is used to acknowledge actions from generated sequence videos of each person, and the 2D companion is used to believe actions based on Image Images of the detected motion. The proposed methodology leverages the proposed 3D without any preprocessing of videos acquired from the web and YouTube in the last part (activity classification), which is notably helpful in dealing with more complicated video content. Following the recognition stage, each person is tasked with summarizing the scene’s events. Each person’s recognized action at each point in the movie is used to translate the film. Frames can be used to depict this summarization (each frame illustrates an effort made). Players and coaches can use Sports Performance Analysis to gain objective data that will help them better understand and improve the team and individual performance. Tagged events and actions can either focus on the overall game, which aids in understanding team performance, or individual players, which aids in understanding individual player performance.

3.4 Multi-action detection and recognition

Figure 6.

Retrieved sequence of the detected person.

The retrieved sequence of the detected person is shown in Fig. 6 as the input to the proposed Deep learning algorithm. Multi-action detection and recognition are challenging to choose, depending on the applications. Before feeding action, detecting the suggested deep learning-based methodology entails preprocessing them. Preprocessing entails extracting the target region containing the player’s activity and tracking and then scaling the data before generating the video. There has been implemented a 3-D, which is supervised learning with a multistage deep learning network. From the videos in the input, 3D can learn various invariant features. A method for gathering data from moving objects for training and testing. The motion analysis methodology typically employs a high-speed camera and a computer with software that allows frame-by-frame replay of the video. It can extract and construct the sequence of each individual based on the tracking results, which reflects sports player actions. Recognizing the activities of multiple people in a single scenario has the added benefit of allowing the creation of a sequence for each of them. These acts can be recognized with the given methodology and proposed preprocessing. The segregated human bodies in the video extract a sequence or clip of body motions throughout their action and presence in the movie to prepare data for deep Learning for training. The detection is carried out using the method outlined, which involves background modeling and subtraction before segmenting the moving objects.

Finally, they intend to show in a future study that this strategy is effective. Performing Ratios, Behavior Processing Ratios, Athletes Energy Level Ratios, Interaction Ratios, Prediction Ratios, Sensitivity Ratios, and Precision Ratios, DL-FTMR has improved the data analysis using an inertial measurement unit (IMU) and computer vision systematically.

4. Results and discussion

The study aimed to use the inertial measurement unit (IMU) and computer vision data analysis to systematically employ the IMU and computer vision data analysis for deep Learning of football teaching motion recognition (DL-FTMR). The findings revealed that response filtering accuracy was enhanced without sacrificing response time. This finding proves that exercising a video monitoring system that provides a play view similar to that seen in a game scenario can be a valuable technique to increase selection accuracy perception. Different deep-learning models were constructed and trained to detect digits, classify, and monitor players in research. The video monitoring model was modified to see most of the person class objects in the movie due to camera settings and player resolution. This work can change most deep learning models for Stat Matrix video inputs, resulting in outputs that reduce operator effort and are not compared to other tracking systems.

On the other hand, these models can create tracking data and identify individuals based on their t-shirt numbers for any other athletic video. We plan to apply split-feature matching to different tracking situations to see how it compares to traditional full-body matching.

Consequently, this study thoroughly examines well-known available public datasets for video sequences and complete descriptions of each dataset. The current comprehensive view of the tools and methodologies for Vision-based person identification can be further researched in the future with player action recognition. Finally, they intend to show in a future study that this strategy is effective. Performing Ratios, Behavior Processing Ratios, Athletes Energy Level Ratios, Interaction Ratios, Prediction Ratios, Sensitivity Ratios, and Precision Ratios, DL-FTMR has improved the data analysis using an inertial measurement unit (IMU) and computer vision systematically. The proposed model has been tested and compared to current models on various performance metrics. Finally, the comparative study yields the highest classification precision.

Table 1
Comparisons of performance metrics

Parameters	OCNN	HAR-SAM	GMM	YOLO	DL-FTMR
Performance ratio (%)	45.7	55.67	65.6	47.7	94.5
Behavior analysis ratio (%)	55.6	50.5	65.4	67.8	92.4
Athletes energy level ratio (%)	62.3	52.3	42.3	64.3	92.5
Interaction ratio (%)	23.5	43.5	53.5	25.5	91.8
Sensitivity ratio (%)	58.2	68.2	38.2	59.2	93.7
Prediction ratio (%)	58.5	38.5	68.5	57.5	92.5
Precision ratio (%)	55.5	65.5	55.5	54.5	94.86

Table 1 shows the Comparisons of performance metrics, which are graphs and data that depict how a sports athlete’s behavior, capabilities, and overall performance are represented. Organizations must choose the highest standards of success to guide and measure college student effectiveness. College students’ mental health can benefit from DL-FTMR. The study’s purpose is to systematically use data from an inertial measurement unit (IMU) and data from computer vision processing for deep Learning of football motion recognition (DL-FTMR). Many libraries have been sought out. For sport-specific video-based decision-making assessments, investigations reveal the ability to differentiate between the efficiency of competent and less qualified officers. Performance metrics help to enhance analysis and predictability, which leads to better results.

4.1 Performance ratio (%)

Figure 7.

(a) Performance ratio; (b) Behavior analysis ratio.

Figures 7(a) and (b) show the performance and Behavior analysis ratios. Coaches can use player surveillance to improve their teams’ performance by instantly analyzing how individual players move on the field and the general configuration of their squad. Finally, action recognition and categorization can automatically provide performance data such as shot kinds, passes, and possession throughout a match or training session. It automatically creates highlight videos by indexing films by predetermined themes depending on their contents. The importance of performance assessment in this process must be considered. For observation, the coach or a video camera can be deployed. Because coaches can only recall about half of the critical events during a game, it is ideal to employ a video camera to record the actual events (actions and motions) for future analysis. The term analysis refers to analyzing data, which includes data management. Video can be used in coaching in a variety of ways. The majority of coaches should have access to video systems. Athletes can use camera regression to determine their performance and develop their skills while avoiding harm. When one records a player’s performance, one can pick up on minor aspects and techniques that one might miss when watching them live. A video camera and a specialist to evaluate the actions can assist athletes in seeing mistakes, discovering strengths, and often providing the player with a completely new viewpoint on their performances. The coach’s job is to make the most of the techniques to increase the athletes’ and team’s performance. The challenge for the sports scientist is determining how the coach should employ technology. The software developer aims to construct intelligent tools that let coaches analyze and evaluate the player and team skills as near to real-time as possible. The raw player tracking data can be enhanced by applying machine learning and data mining techniques to the results acquired by a computer vision system. After identifying essential features in an image or video frame, semantic information can be generated to provide context for the players’ activities.

4.2 Athletes energy level ratio (%)

Figure 8.

(a) Athletes energy level ratio; (b) Interaction ratio.

Figure 8(a) shows the Athletes’ Energy level Ratio. Football is a highly intermittent activity that requires a strong capacity for resistance due to high-intensity racing and very specialized moves. Maximum effort actions are interspersed with active or passive running or walking recovery periods. It is known that several metabolic pathways are engaged in energy production in this fashion. According to the research, high and efficient athletes’ energy in the DL-FTMR systems results in tall rushes during play, making the DL-FTMR system more critical. The increased activity ratio can be attributed to this. Figure 8(b) depicts the obtained Interaction Ratio. Player position detection has traditionally relied on various data-collecting methods, from live monitoring to video analysis. Player performance patterns are manually documented and evaluated to assess efficiency. Because manually managing and evaluating such data took so long, the investigation focused primarily on a small number of participants in specific areas. Even though notational analysis is a convenient, realistic, and typically cost-effective method, its validity and reliability might vary based on a range of factors, such as the number of observers involved, their competence, and the nature of their viewing viewpoint. Inadequate video and computer infrastructure on sporting grounds has undoubtedly hampered the use of autonomous monitoring technology in team sports. However, resolving the fluid quality of effort inherent in many sports activities is a huge challenge. Athletes are known for their speed, agility, and ability to adapt to various situations. Interactions with other players are frequent and predictable.

4.3 Sensitivity ratio (%)

Figure 9.

Sensitivity ratio.

Figure 9 shows the Sensitivity Ratio. Technical action in football is diverse from the standpoint of action analysis. As a critical assurance, successfully implementing these procedures necessitates high sensitivity. As a result, sensitivity training for players is particularly crucial. This study conducts relevant research on related topics in conjunction with this element to reflect the ultimate usefulness of the training and testing models. First, it should be linked with the unique characteristics of linear motion and the professional and personal structure goals to build the particular investigation. The footballer sensitivity proper training has a solid theoretical foundation. The characteristic velocity performances of footballer sensitivity quality are built on a foundation of explosive beginnings, quick stops, acceleration changes, and start-up speed, allowing the football player to execute more complex technical moves at high speeds. The beginning force can be defined as displaying the greatest extent of a smaller mass in the shortest time, and it is the foundation of the acceleration forces. It can overcome external resistance to the greatest extent possible in the shortest time, resulting in good power for high-speed running – commonly called explosive power. Video motion analysis is used by sports professionals, athletes, and coaches to gather data by capturing moving images with digital movie cameras and then analyzing the images frame by frame with software.

4.4 Prediction ratio (%)

Fast and precise action is prevalent in most significant football leagues, making it challenging for coaches and analysts to observe and analyze in great detail. Using wearable tracking equipment and sensors to enhance data collection isn’t possible, which is highly problematic. During training sessions and specific matches, performance analyzers can only collect a limited number of angles of video footage, especially if they are not aired. Coaches and managers can now make informed judgments based on the facts they’ve received rather than relying just on their intuition or projections. If a key player is injured, the team manager, coach, or captain can use the predetermined data to choose a suitable replacement. The assistance supplied by intelligent systems is beneficial to the players. They can grow as individuals by recognizing their flaws. Players, coaches, and teams always seek critical information that might influence performance. It benefits the referees, umpires, and decision-makers because decisions are now made in seconds rather than minutes. The Prediction Ratio (%) is shown in Table 2.

Table 2
Prediction ratio (%)

Number of users	Prediction ratio (%)
	OCNN	HAR-SAM	GMM	YOLO	DL-FTMR
10	61.3	75.9	77.8	79.2	87.6
20	62.4	76.6	77.7	79.3	88.7
30	63.0	74.5	81.4	82.6	89.1
40	64.0	74.2	78.5	83.9	90.3
50	65.8	76.4	79.7	84.5	90.7
60	66.8	72.1	78.2	83.4	90.9
70	67.7	74.2	80.3	85.6	91.5
80	70.9	77.0	82.6	86.2	92.0
90	71.5	78.1	83.8	87.5	92.2
100	72.2	79.1	82.1	88.9	92.5

4.5 Precision ratio

Cognitive function factors (power, velocity, stamina, flexibility, cooperation, and precision), social factors (mental, conative, social), recognition program, teaching and training methods, a variation of external influences (playground, referees, equipment, public, etc.), and error factors all influence the outcome in football. Those players with these variables in an ideal balance have a better chance during the sport. Many elements effectively changed throughout the training process with players, notably psychomotor ones, particularly on endurance, where the teaching and training methods are attractive and crucial for the procedure. The situation and the element of chance significantly impact one’s level of gaming success. The importance of coincidence in success in football is due to the game’s changeable, complex, and unpredictable nature. When this is applied to a football game, precision is determined by the game’s accuracy. A player who lacks this aptitude is unlikely to fit into any tactical scheme. However, precision is only one of the factors that must be considered when hitting, passing the ball, or making shots. It is primarily influenced by how well technical components are implemented, the speed and precision with which movements are coordinated, and agility (football is an agility sport). The Precision Ratio (%) is shown in Table 3.

Table 3
Precision ratio (%)

Number of users	Precision ratio (%)
	OCNN	HAR-SAM	GMM	YOLO	DL-FTMR
10	31.36	65.37	77.52	91.22	87.62
20	30.01	66.65	77.27	92.11	88.71
30	39.94	64.57	77.94	92.15	89.15
40	38.24	64.22	78.51	93.18	89.31
50	38.82	66.42	78.79	94.24	90.72
60	47.52	72.12	79.34	94.48	90.64
70	47.27	74.54	80.45	95.27	91.72
80	45.94	63.04	81.65	96.90	92.92
90	55.57	66.14	82.87	97.32	93.32
100	55.128	64.124	83.111	95.76	94.86

Thus, the experimental results DL-FTMR show that these parameters – performing Ratios, Behavior Processing Ratios, Athletes’ Energy Level Ratios, Interaction Ratios, Prediction Ratios, Sensitivity Ratios, and Precision Ratios – have been proposed when compared to the optimized convolutional neural network (OCNN), Gaussian Mixture Model (GMM), you only look once (YOLO), and Human Activity Recognition – state-of-the-art methodologies (HAR-SAM).

5. Conclusion

The study aims to systematically apply the inertial measurement unit (IMU) and computer vision data analysis for deep learning of football instructional motion recognition (DL-FTMR). Reaction filtering accuracy improved without reducing response speed, according to the data. The findings revealed that response filtering accuracy was enhanced without sacrificing response time. This finding proves that exercising a video monitoring system that provides a play view similar to that seen in a game scenario can be a valuable technique to increase selection accuracy perception. We published on the website ‘Football instruction and training based on Video surveillance utilizing deep Learning,’ which is a complete package of everything a hopeful individual wants to know about sports and its associated data. It provides all of the necessary information regarding football games.

Additionally, it enables users to participate in the various clubs associated with it. The initial page of our project provides a basic description of the games and their experiences. It includes information on the top players’ profiles in multiple sports and an update on current game developments. It is quite beneficial to someone who requires information about various sports. Online dialogue has taken advantage of invaluable consumer analysis and created a positive image for product testing that appeals to online customers and keeps them in mind for business and enrollment purposes. As a result of this Flash-based game, a global web business focusing on consumer needs has been established. Our website is simple to navigate. Thus, the experimental results of DL-FTMR showed a performance ratio of 94.5%, a behavior processing ratio of 92.4%, an athlete energy level ratio of 92.5%, an interaction ratio, a prediction ratio of 92.5%, a sensitivity ratio of 93.7%, and a precision ratio of 94.86%.

Funding

None of the authors received funds or grants.

Author contributions

All Authors contributed to the design and methodology of this study, the assessment of the outcomes, and the writing of the manuscript.

Data availability

All data generated or analyzed during this study are included in the manuscript.

Code availability

Not applicable.

Footnotes

Conflict of interest

There is no conflict of interest among the authors.

References

Leilei

Rajendiran

Gayathri

. An emergency response system created to combat injuries during physical education training in a university using deep learning. The Electronic Library. 2021 Nov 4; 39(4): 505-25.

Arshad

Khan

Sharif

Yasmin

Tavares

Zhang

Satapathy

. A multilevel paradigm for deep convolutional neural network features selection with an application to human gait recognition. Expert Systems. 2022 Aug; 39(7): e12541.

Parthiban

Pazhanivel

Baskar

Sridhar

. Experimental Analysis to Reduce the Emissions from Exhaust of Petrol Engines Using Nano Catalytic Coating with Heterogeneous Nano Structured Sensor. Sensor Letters. 2018 Dec 1; 16(12): 931-7.

Berk

. Multimedia teaching with video clips: TV, movies, YouTube, and mtvU in the college classroom. International Journal of Technology in Teaching & Learning. 2009 Jun 1; 5(1).

Zhang

Chen

Wei

. A sport athlete object tracking based on deep sort and yolo V4 in case of camera movement. In 2020 IEEE 6th international conference on computer and communications (ICCC) 2020 Dec 11 (pp. 1312-1316). IEEE.

Ngo

Bui

Pham

Thai

Nguyen

. Image deconvolution for optical small satellite with deep learning and real-time GPU acceleration. Journal of Real-Time Image Processing. 2021 Oct; 18(5): 1697-710.

Nguyen

Phan

Hoang

Nguyen

So-In

. Efficient SDN-based traffic monitoring in IoT networks with double deep Q-network. In International conference on computational data and social networks 2020 Dec 11 (pp. 26-38). Cham: Springer International Publishing.

Abidi

Alkhalefah

Moiduddin

Alazab

Mohammed

Ameen

Gadekallu

. Optimal 5G network slicing using machine learning and deep learning concepts. Computer Standards & Interfaces. 2021 Jun 1; 76: 103518.

McDole

Gupta

Abdelsalam

Mittal

Alazab

. Deep learning techniques for behavioral malware analysis in cloud iaas. Malware analysis using artificial intelligence and deep learning. 2021: 269-85.

10.

Manogaran

Shakeel

Fouad

Nam

Baskar

Chilamkurti

Sundarasekar

. Wearable IoT smart-log patch: An edge computing-based Bayesian deep learning network system for multi access physical monitoring system. Sensors. 2019 Jul 9; 19(13): 3030.

11.

Zhou

Yang

Kim

Cimato

. Introduction to the special issue on deep learning for real-time information hiding and forensics. Journal of Real-Time Image Processing. 2020 Feb; 17: 1-5.

12.

Huifeng

Kadry

Raj

. Continuous health monitoring of sportsperson using IoT devices based wearable technology. Computer Communications. 2020 Jul 1; 160: 588-95.

13.

Chang

Lin

Chu

. Effects of training using an active video game on agility and balance. The Journal of sports medicine and physical fitness. 2015 Sep 1; 55(9): 914-21.

14.

Khan

Kabir

Hassan

Ahmed

. A deep learning framework using passive WiFi sensing for respiration monitoring. In GLOBECOM 2017-2017 IEEE Global Communications Conference 2017 Dec 4 (pp. 1-6). IEEE.

15.

Mohammadi

Amoozegar

Jolfaei

Mirghadri

. Enhanced adaptive bandwidth tracking using mean shift algorithm. In 2011 IEEE 3rd International Conference on Communication Software and Networks 2011 May 27 (pp. 494-498). IEEE.

16.

Elhoseny

. Multi-object detection and tracking (MODT) machine learning model for real-time video surveillance systems. Circuits, Systems, and Signal Processing. 2020 Feb; 39(2): 611-30.

17.

Devi

Priyan

Gokulnath

. Wireless camera network with enhanced SIFT algorithm for human tracking mechanism. International Journal of Internet Technology and Secured Transactions. 2018; 8(2): 185-94.

18.

Kumaran

Radha Rammohan

Nagarajan

Prathik

. Fusion of mel and gammatone frequency cepstral coefficients for speech emotion recognition using deep C-RNN. International Journal of Speech Technology. 2021 Jun; 24(2): 303-14.

19.

Zhou

Montenegro-Marin

Hsu

. RETRACTED ARTICLE: Wearable IoT based cloud assisted framework for swimming persons in health monitoring system. Current Psychology. 2022 May 1: 1.

20.

Qiu

Liu

Muthu

Sivaparthipan

. Design of an energy-efficient IoT device with optimized data management in sports person health monitoring application. Transactions on Emerging Telecommunications Technologies. 2022 Oct; 33(10): e4258.

21.

Christina

Barresi

Shaffner

. The development of response selection accuracy in a football linebacker using video training. The Sport Psychologist. 1990 Mar 1; 4(1): 11-7.

22.

Cocca

Marciano

Alberti

. Video surveillance systems to enhance occupational safety: A case study. Safety Science. 2016 Apr 1; 84: 140-8.

23.

Janarthanan

Doss

Baskar

. Optimized unsupervised deep learning assisted reconstructed coder in the on-nodule wearable sensor for human activity recognition. Measurement. 2020 Nov 1; 164: 108050.

24.

Manogaran

Shakeel

Fouad

Nam

Baskar

Chilamkurti

Sundarasekar

. Wearable IoT smart-log patch: An edge computing-based Bayesian deep learning network system for multi access physical monitoring system. Sensors. 2019 Jul 9; 19(13): 3030.

25.

Song

Montenegro-Marin

Krishnamoorthy

. Secure prediction and assessment of sports injuries using deep learning based convolutional neural network. Journal of Ambient Intelligence and Humanized Computing. 2021 Mar; 12: 3399-410.

26.

Girdhar

. Vision based human activity recognition: a comprehensive review of methods & techniques. Turkish Journal of Computer and Mathematics Education (TURCOMAT). 2021 Apr 28; 12(10): 7383-94.

27.

Gochoo

Akhter

Jalal

Kim

. Stochastic remote sensing event classification over adaptive posture estimation via multifused data and deep belief network. Remote sensing. 2021 Feb 28; 13(5): 912.

28.

Feddersen

Keis

Elbe

. Coaches’ perceived pitfalls in delivering psychological skills training to high-level youth athletes in fencing and football. International Journal of Sports Science & Coaching. 2021 Apr; 16(2): 249-61.

29.

Bamunif

. Sports information & discussion forum using artificial intelligence techniques: A new approach. Turkish Journal of Computer and Mathematics Education (TURCOMAT). 2021 May 10; 12(11): 2847-54.

30.

Yue

Pan

. Camera-based basketball scoring detection using convolutional neural network. International Journal of Automation and Computing. 2021 Apr; 18(2): 266-76.

31.

Falloon

. From digital literacy to digital competence: the teacher digital competency (TDC) framework. Educational technology research and development. 2020 Oct; 68(5): 2449-72.

32.

Xiao

Han

. Wearable heart rate monitoring intelligent sports bracelet based on Internet of things. Measurement. 2020 Nov 1; 164: 108102.

33.

Bin

Xiaoyun

. RETRACTED: Home intelligent sports action automation system based on bluetooth.

34.

Poulose

Kim

Han

. HIT HAR: human image threshing machine for human activity recognition using deep learning models. Computational Intelligence and Neuroscience. 2022; 2022(1): 1808990.

35.

Ronald

Poulose

Han

. iSPLInception: an inception-ResNet deep learning architecture for human activity recognition. IEEE Access. 2021 May 7; 9: 68985-9001.

36.

Gokulakrishnan

Jarwar

Ali

Kamruzzaman

Meenakshisundaram

Jaber

Kumar

. Maliciously roaming person’s detection around hospital surface using intelligent cloud-edge based federated learning. Journal of Combinatorial Optimization. 2023 Jan; 45(1): 13.

Football teaching and training based on video surveillance using deep learning

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1. Introduction

1.1 Overview of football teaching and training based on video surveillance

2. Literature review

3. DL-FTMR has been utilized systematically to analyze the inertial measurement unit (IMU) and computer vision data

3.4 Multi-action detection and recognition

Table 1 Comparisons of performance metrics

Table 2 Prediction ratio (%)

Table 3 Precision ratio (%)

Funding

Author contributions

Data availability

Code availability

Footnotes

Conflict of interest

References

Table 1
Comparisons of performance metrics

Table 2
Prediction ratio (%)

Table 3
Precision ratio (%)