Abstract
The study was aimed at realizing the identification of athletes’ actions in badminton teaching. The teaching process is segmented into many independent actions to help learners standardize their movements in badminton play, improving the national physical quality. First, the principle and advantages of machine vision sensing are introduced. Second, the images and videos about the action decomposition of badminton teaching are collected, and the image data are extracted by Haar-like. Subsequently, badminton players’ actions are recognized and preprocessed, and a dataset is constructed. Furthermore, a new algorithm model is implemented and trained by using Haar-like and Adaptive Boosting (AdaBoost). Finally, the badminton players’ action recognition algorithm is tested and compared with the traditional hidden Markov model (HMM) and support vector machine (SVM). The results show that action images improved by machine vision can process the captured actions effectively, making the computer better identify different badminton teaching actions. The proposed method has a recognition rate of more than 90% for each action, the average recognition accuracy of actions reaches 95%, the average recognition rate of the same person’s actions is 96.5%, and the average recognition rate of different people’s actions is 94.8%. The badminton teaching action recognition model based on Haar-like and AdaBoost can recognize and classify badminton actions and improve the quality of badminton teaching. This study shows that the image processing technology can effectively process the players’ static images, which gives the direction for physical education (PE) under artificial intelligence (AI).
1. Introduction
Due to the development of Internet technology, artificial intelligence (AI) [1] technology is widely used in all aspects of peoples’ lives, such as production, education, and research fields. Nowadays, it is also used for national physical fitness and professional physical training.
Vision sensors can provide information to the machine vision sensing system, and computer vision conduction focuses on the simulation of animal vision by using computer technology. The simulation involves a variety of technologies, including image processing technology, mechanical engineering technology, mechanical and electronic control technology, lighting technology, optical image capture, processing technology and sensor technology, analogy and digital video technology, computer software and hardware, and support technology (for image enhancement and algorithm analysis). Its main task is computer processing and recognition of images [2, 3]. The application system of machine vision includes the image acquisition module, light-source processing module, image digitization module, digital image processing module, decision-making module, and hardware control implementation module. Machine vision is realized by improving intelligent automation and the quality of production and service. It is used in the environments where manual operation cannot meet the requirements. It can monitor the product’s quality and greatly improves the production efficiency in large-scale mechanical production [4–6].
Zhou et al. compared the efficiency of support vector machine (SVM), logistic regression (LR), and artificial neural network (ANN) for shoulder motion pattern recognition of surface electromyogram (EMG). They studied the effect of sliding time window epochs on the recognition accuracy and verified the accuracy of the LR recognition algorithm [7]. Storey et al. proposed an end-to-end framework named 3D-PalsyNet. The framework is used for mouth motion recognition and facial paralysis scoring. 3D-PalsyNet utilizes a 3D-convolutional neural network (CNN) architecture with a ResNet backbone to predict these dynamic tasks. A pretrained 3D-CNN on the kinetic dataset is used for transfer learning for general action recognition. The model is modified to combine center and softmax for supervised learning [8]. Zhang et al. proposed a solution to identify table tennis motion using commercial smart watches and developed a data acquisition system based on IoT architecture. The system is used to obtain data on the acceleration, angular velocity, magnetic induction, etc. of the watch. Based on the features of the extracted data, the main machine learning (ML) classification algorithms such as k-nearest neighbors, SVM, naive Bayes model (NBM), LR, decision tree, and random forest (RF) are used for verification experiments [9]. Most of the previous studies have used ML techniques such as supervised learning. The innovation lies in the model constructed by using Haar-like features combined with Adaptive Boosting (AdaBoost). The AdaBoost algorithm makes good use of weak classifiers for cascading and can use different classification algorithms as weak classifiers with high accuracy. Relative to the RF algorithm, AdaBoost fully considers the weight of each classifier.
In the information age, there appear various emerging computer technologies, and intelligent image recognition technology applied to badminton teaching is included. In the use of intelligent technology, the teaching action images are collected by cameras, and Haar-like is used for feature extraction and image denoising. The action images are preprocessed and the training set and test set are constructed. The model is implemented by AdaBoost, and the performance of the constructed recognition algorithm is tested. According to computer vision technology, the collected images are preprocessed and the implemented model is trained, achieving intelligent image processing and recognition timely, efficiently, and accurately.
2. Methods
2.1. Machine Vision Sensing
Vision is originally a way for organisms to obtain external information. Now, it is one of the core components in promoting biological intelligence. It is known that 80% of information is obtained by vision. Figure 1 shows some ways to obtain information through the vision. Inspired by this, researchers install “eyes” on machines, so that machines can get necessary external information through “seeing” like humans, which is called computer vision. Machine vision is a comprehensive technology, and it is mainly applied to establish an image capture system, a light source system, an image digitization module, and so on. After that, researchers make a machine vision system by analyzing biological vision systems. The key to vision sensor technology is image processing; that is, an image is processed by intercepting the signal on the object surface [10, 11].
The vision sensor has a huge number of pixels, which can capture the light of an image. And the number of pixels determines the clarity and fineness of an image. After an image is captured, the vision sensor compares it with the given image for analysis. Visual sensing technology includes 3D visual sensors, which is applied in many fields, such as multimedia mobile phones, network cameras, digital cameras, robot visual navigation, automobile safety systems, biomedical pixel analysis, man-machine interfaces, virtual reality (VR), monitoring, industrial detection, wireless remote sensing, microscope technology, astronomical observation, marine autonomous navigation, and scientific instrument tests. In particular, it can be used in industrial control and automobile autonomous navigation. A complete machine vision system contains the following five parts: lighting systems, lens, high-speed cameras, image acquisition cards, and vision processors [12, 13]. The requirements of each part are shown in Table 1.
Intelligent visual sensing technology is also a visual sensing technology. It is also an intelligent camera with a small machine vision system, which can collect and process images and transmit the relevant information. It assembles the image sensors, digital processors, communication modules, and other peripherals into a camera, simplifying the system and improving its reliability. This widens the application range of vision technology. It can help to build a reliable detection system because it is easy to learn, use, and install. Its image collection unit comprises a charge-coupled device (CCD)/a complementary metal-oxide-semiconductor (CMOS), an optical system, a lighting system, and an image acquisition card. The optical image is converted into the digital image and transmitted to the image processing unit, as shown in Figure 2.
The machine vision system can improve flexibility and automation in the production process. In some dangerous environments, it can replace artificial vision. In mass repetitive production, it can detect and improve working efficiency.
CCD has the functions of photoelectric conversion, information storage, delay, and sequential transmission of electrical signals and has high integration and low power consumption, so it has developed rapidly. CCD is an indispensable key device for image acquisition and digital processing. It is widely used in scientific, educational, medical, commercial, industrial, military, and consumer fields. CCD system mainly includes optical system (microlens), CCD, and image processing module, and some also include color filter. A CCD image sensor is an array of capacitors arranged according to certain rules. A very thin layer of silicon dioxide (SiO2 about 120 nm) is formed on the silicon substrate, and then metal or doped polysilicon electrodes (gates) are sequentially deposited on the SiO2 thin layer to form a regular capacitor array, thus forming a CCD chip. The working process of CCD is as follows: (a) the generation of signal charge. The CCD can convert the incident light signal into a charge output. The principle is the photoelectric effect (photovoltaic effect) in the semiconductor. The metal-oxide-semiconductor capacitor is the most basic unit that constitutes a CCD. (b) Storage of signal charge: it is the process of collecting the charges excited by incident photons and converting them into signal charge packets. (c) Transmission and transfer of signal charges: it is the process of transferring the collected charge packets from one pixel to the next until all the charge packets are output. (d) Detection and output of signal charge: it is the process of converting the charge transferred to the output stage into current or voltage. There are three main output types: current output, floating gate amplifier output, and floating diffusion amplifier output.
2.2. ML
2.2.1. Deep Learning
In 2006, deep learning (DL) appears and it is a branch of ML. It is studied by academia and gradually applied in the industry. In 2012, the Stanford University uses 16000 CPU core parallel computing platforms to expound deep neural networks (DNN), which have an advantage in speech and image recognition. In 2016, “go” is developed from DL and it helps defeat Li Shishi, the world’s top master, in the competition. After that, well-known high-tech companies around the world begin to pay more attention to DL and establish research institutes for it, expanding the size of the research team of DL [14].
ML studies how computers simulate or realize the learning behavior of animals and use new knowledge or skills to rewrite the existing data structure, improving the program performance. According to statistics, it can estimate data distribution, learn a data model, and predict new data through this model. ML uses algorithms to analyze data, learn from it, and make decisions. This means that it can teach computers to develop an algorithm and complete assignments instead of compiling programs to perform certain tasks. There are three main types of ML: supervised learning, unsupervised learning, and reinforcement learning. They have specific advantages and disadvantages. Supervised learning targets labeled data. In the learning process, the computer identifies new samples by the specific patterns, and classification and regression are achieved. In terms of classification, the machine is trained and data are divided into specific classes. This process is like the spam filter on your email account. The filter analyzes the previous email messages and compares them with new ones. If the given ratio is met, the messages are marked and sent to the corresponding folder, and the rest are sent to the destination mailbox. As for regression, the machine can predict future ones according to the labeled data. This is the same as a weather broadcast. With historical data (average temperature, humidity, and precipitation), the APP on mobile phones can predict the weather conditions in the future. Unsupervised learning is for unlabeled data, and it can perform clustering and dimension reduction. Clustering is completed according to the attributes and behaviors. It divides a group into different subgroups (according to age and marital status) and applies them to marketing schemes. Dimensions are reduced under common ground.
Reinforcement learning is implemented on the personal experience of the machine. It is like playing games. It focuses on performance. For example, you play chess on a computer and cannot move the king into the space that the opponent’s chess pieces can enter. The experience of playing chess is inferred until the machine beats (and eventually defeat) the top players.
It makes the test data and training data identically distributed. It tries to imitate the transmitting and processing mode of brain neurons. It is mostly applied to computer vision and natural language processing (NLP). DL relies on neural networks in ML. Therefore, DL is called the improved neural network [15, 16]. Figure 3 shows the structure of the convolutional neural network (CNN).
2.2.2. Artificial Neurons
The artificial neuron is a mathematical model created by imitating the basic operation function of biological neurons. The artificial neuron receives the given signal from the front neuron, and each given signal is attached with a weight. Under the joint action of all weights, this neuron shows a corresponding activation state [17], and its principle is shown in Figure 4.
The principle of artificial neurons is expressed by
In equation (1), represents the final output state, is the input signal, and is its weight. There are groups in total.
When it receives an input signal, a neuron gives a certain output, and each neuron has a corresponding threshold. If the sum of the inputs received by this neuron is greater than the threshold, its state will change to an active state. When it is less than the threshold, it will show an inhibitory state. The transfer functions of artificial neurons are as follows [18, 19]. (a)The linear function is calculated by (b)The slope function is calculated by
The transition function is calculated by
Sigmoid is calculated by
The transfer function needs to be selected according to the specific range. The linear function amplifies the output signal. The nonlinear slope function prevents the degradation of network performance. The S-type function sets parameters , , , , , and in the hidden layer. In equation (5), and represent the input and output values, respectively.
Neurons can be used for calculating weights and summations. is the input value of the -th neuron, is the weight between the -th and -th neuron, is a threshold, and is the transfer function.
Net output value of neuron is calculated in equation (6). When the threshold is 0, is 0.
2.3. Image Recognition and Processing
2.3.1. Image Recognition
Image recognition is shown in Figure 5.
2.3.2. Feature Extraction
Nowadays, the features of image recognition fall into global and local features. The global features are extracted by the full graph search window, which reflects the general feature. The local feature is the detailed features of the image in the sliding window. Global feature extraction method can make principal component analysis (PCA) [20], gray-level gradient cooccurrence matrix (GLCM) [21], and frequency domain. Local features include scale invariant feature transform (SIFT) [22] and Haar-like features [23].
Haar-like features are extracted for recognition. They have four basic structures and are viewed as windows. This window slides in the image with a step of 1, and a complete image is finally extracted. The length and width of the sliding window are increased. The process is repeated until they are enlarged to the largest.
2.3.3. Classification and Recognition Method
Image recognition needs to use the classifier and a classifier should be designed. And feature extraction is realized by supervised and unsupervised learning. Supervised methods need a huge number of training sample sets and a classifier, including neural networks, SVM [24], and AdaBoost [25]. Unsupervised methods classify data according to their similarity and characteristics. The mainstream method is K-mean [26]. Canadian EC650C camera and 5G are used for data wireless network transmission by AdaBoost.
2.4. Recognition Algorithms
The mainstream recognition algorithm is local vision recognition, but there is less research on image recognition using omnidirectional vision. Therefore, the action recognition model based on DL is used to improve local vision recognition by Haar-like and AdaBoost.
First, the video data are collected and the sample image is preprocessed. Then, the collected images are used as the training set, and the classifier is constructed. After that, the algorithm is used to obtain the frames of images. If tracking fails, badminton players’ actions will be retracted.
2.4.1. Sample Set Construction
A sample set should be constructed for feature training and algorithm recognition. It consists of a training set and a test set. Because there will be a lot of training, a large number of images are needed. The training set should include geographical features and light conditions. Finally, 263 players’ images and 557 scene images are collected.
2.4.2. Image Preprocessing
The scale and size of the sample set should be adjusted for Haar-like features’ calculation. Figure 6 shows the data pre-processing process as follows.
2.4.3. Classifier Algorithm
The AdaBoost algorithm is used for establishing the algorithm of the classifier [27, 28], and the equation is as follows: (a)Training set is constructed(b)Initial parameters and their weights are determined by (c)Weighted mean square deviation is estimated by (d)The classifier is updated by (e)The sample weight is updated and normalized by (f)Iterate 3-5 steps, and the times are (g)The result of the classifier is obtained by
2.4.4. Data Enhancement
is the data coordinate information of the image, and is the original labeled sample of the actual value of the human action, denoted as . The sample after adjustment in the expression frame is denoted as .
Data enhancement requires a variety of operations on the original image. In order to describe the rotated coordinates, set the top left corner of the image as point O, and set its coordinates as (0,0). Starting from (0,0), the coordinates from top to bottom are set as the -axis, and the coordinates from left to right are set as the -axis. The picture size is , the rotation angle is , and the clockwise direction is the positive direction. is the coordinate of the rotated human body joint , and is the rotated image data. (a)Rotation is shown in
The rotation angle is set to be random to obtain many training samples. (b)Translation: different points are selected on the image to translate the human body. According to equation (14), the minimum clipping region is calculated.
is the coordinate of the pixel at the top left of the cropped area. is the coordinate of the bottom right pixel. (c)Zoom(d)Horizontal flip: equation (15) is the coordinates after horizontal flipping.
represents the abscissa after flipping. represents the ordinate after flipping. represents image data after flipping. represents the pixel coordinate value after flipping. is the result after the original joint is flipped.
Figure 7 shows the training and recognition process.
2.5. Testing Environments
The recognition rate of Haar-like is tested. 300 teaching and 500 nonplayer images are extracted, including 600 training images and 200 test images. They are used to identify the athletes’ serve, positive rubbing, reverse rubbing, positive flutter, reverse flutter, positive push, reverse push, positive pick, reverse pick, and high and far actions. Among them, each action has a training set of 40 pictures and a test set of 10 pictures. The contents of the three experiments are shown in Table 2.
3. Result
3.1. Influence of Different Sliding Windows on Segmentation
The comparison of recognition rates of three different segmentation methods is shown in Figure 8.
Figure 8 shows that method A based on the hitting time window has a higher recognition rate than B and C. Its recognition rate in each event is greater than 90%, which is much higher than the effect of the other two. Therefore, the method based on the hitting time window is used to extract the players’ hitting actions. For each hitting point, the peak value of hitting time is detected, and it is superior.
3.2. Recognition Rates of Different Algorithms
The comparison of recognition rates of three different training and recognition models is shown in Figure 9.
Figure 9 shows that the recognition rate of AdaBoost for each action is greater than 90%, and the average recognition accuracy is greater than 95% compared with SVM and HMM. The average action recognition rate of SVM is only 78%, and that of HMM is between 90% and 95%. This shows that the effect of AdaBoost is much higher than that of traditional algorithms.
3.3. The Image Recognition Rate of the Same and Different Players
The comparison of the recognition rates of the same and different players is shown in Figure 10.
Figure 10 shows that AdaBoost’s recognition rate of each action of the players is greater than 90%, and it has a more accurate recognition rate for the actions of one player. Its recognition rate for the same player is more than 99%, and its lowest rate is 92%. For the action recognition rate of different players, its recognition rate is also up to 99%, and the lowest is 90%. Therefore, in badminton teaching action decomposition, it is best to use the action image of the same player.
4. Conclusion
Based on the intelligent image recognition of football robots, a vision sensing system suitable for badminton teaching action decomposition based on DL and machine vision is proposed. The image data are collected, and Haar-like is used for feature extraction. The data of badminton players’ actions is preprocessed, and the dataset is constructed. A new model is implemented and trained by Haar-like and AdaBoost, and the performance of the constructed recognition algorithm is tested and analyzed. The experiment shows that the badminton teaching action recognition technology based on Haar-like and AdaBoost can successfully capture and classify images intelligently, improving the quality of badminton teaching.
The study achieves the expected research results and draw valuable conclusions, but there are still deficiencies: (1) the algorithm may have some difficulty in recognizing players’ actions in professional competitions, and more action data in formal competitions will be collected in the future; (2) the action recognition rate of different players is not ideal, and more action data of different players will be collected for training afterward.
Data Availability
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
Conflicts of Interest
The authors report no declarations of interest.
Acknowledgments
This work was supported by the Key Project of Quality Engineering Project Teaching and Research in Anhui Province, Research on the “Trinity” Health Literacy Cultivation Model of College Students—A Case Study of Bozhou University (Project No. 2019jyxm0538).