Abstract
With the continuous emergence and innovation of computer technology, mobile robots are a relatively hot topic in the field of artificial intelligence. It is an important research area of more and more scholars. The core of mobile robots is to be able to realize real-time perception of the surrounding environment and self-positioning and to conduct self-navigation through this information. It is the key to the robot’s autonomous movement and has strategic research significance. Among them, the goal recognition ability of the soccer robot vision system is the basis of robot path planning, motion control, and collaborative task completion. The main recognition task in the vision system is the omnidirectional vision system. Therefore, how to improve the accuracy of target recognition and the light adaptive ability of the robot omnidirectional vision system is the key issue of this paper. Completed the system construction and program debugging of the omnidirectional mobile robot platform, and tested its omnidirectional mobile function, positioning and map construction capabilities in the corridor and indoor environment, global navigation function in the indoor environment, and local obstacle avoidance function. How to use the local visual information of the robot more perfectly to obtain more available information, so that the “eyes” of the robot can be greatly improved by relying on image recognition technology, so that the robot can obtain more accurate environmental information by itself has always been domestic and foreign one of the goals of the joint efforts of scholars. Research shows that the standard error of the experimental group’s shooting and dribbling test scores before and the experimental group’s shooting and dribbling test results after the standard error level is 0.004, which is less than 0.05, which proves the use of soccer-assisted robot-assisted training. On the one hand, we tested the positioning and navigation functions of the omnidirectional mobile robot, and on the other hand, we verified the feasibility of positioning and navigation algorithms and multisensor fusion algorithms.
1. Introduction
In order to further improve the robot’s load capacity, movement flexibility, and for adaptability to small spaces, various types of omnidirectional mobile robots have emerged. The most important is the omnidirectional vision system, which collects images through hardware devices such as omnidirectional cameras and image capture cards to establish a model of the real environment, thereby controlling the robot to recognize the ball, goal, and other robots and providing information for the decision-making system. Therefore, the research on omnidirectional vision has become the focus and difficulty of the research on the vision system of the medium-sized soccer robot. Target recognition has long occupied a very important position in the field of pattern recognition. It is one of the most advanced research directions in computer vision. Through target recognition, computers can have perception and recognition capabilities similar to those of humans. Therefore, more and more scholars are concerned in related fields.
The world’s first autonomously mobile robot, Shakey, was successfully developed. The robot was developed by the Stanford International Research Institute in the United States and possesses a certain degree of intelligence. For the first time, human beings successfully applied artificial intelligence under the conditions of the development to a mobile robot entity. Computer technology has been developing rapidly, and at the same time, microelectromechanical technology and sensor technology have also been greatly improved. Related researchers have begun to gradually apply multisensor technology to the research of mobile robots. Cho et al. applies the extended Kalman filter to the SLAM problem. The algorithm breaks through the linear assumption of the system and extends to the nonlinear system, and it is still widely used [1]. Wang et al. proposed a nonlinear filtering algorithm based on SIS and really began to study particle filtering algorithms. This algorithm is based on the idea of Monte Carlo and expresses the probability of each group of particles, which is applicable to all nonlinear systems [2]. Wang et al. proposed the FastSLAM algorithm based on particle filtering, which can establish a more accurate map in a large-scale environment with few landmarks [3].
In recent years, our country has increased its investment in the research of mobile robots and made great progress. Although color contains more and richer information than grayscale images, it also brings new difficulties. An important aspect is that color images are easily affected by ambient lighting conditions. The study of the relationship between color and lighting, the disclosure of its inherent stable relationship, and the development of powerful algorithms and technologies in lighting changes are among the most urgent research areas today. Jia et al. proposed the idea of using evolutionary strategy in the study of SLAM algorithm based on monocular vision to improve the RBPF-SLAM algorithm and verified the reliability of the algorithm [4]. Yang et al. improved the RBPF-SLAM algorithm and then proposed a data association technology based on global constraints, which robustly reduces the probability of mismatch [5]. Ammar et al. introduced the fixed-level model of route planning and solved it using fast forward method and gave the result of route model calculation and visual interface [6].
The purpose of the image enhancement recognition algorithm in this article is to remove the unfriendly effects of uneven lighting on the color image, enhance the contrast and clarity of the image, and highlight the color features of the target object in the image, so as to better further feature extraction and target recognition. According to the experimental results, the average score and the highest score of the experimental group were higher than those of the control group, and the data were better than those of the control group. If the learning time is the same, the learning effect of the experimental group will definitely be much better than that of the control group, which also proves from the side football assist robot has a broader development prospect.
2. Soccer-Assisted Training Robot Based on Image Recognition Omnidirectional Movement
2.1. Deep Learning
In deep learning, by studying the inherent laws and representation levels of massive data samples, the simple features at the bottom are combined to form more abstract and advanced composite features, and the distributed representation of data features is realized. The dominant neural network is a typical representative of the principle of deep artificial intelligence in neuroscience. As a powerful deep learning technology tool, it has become an important part of the deep learning algorithm driven by the wave of artificial intelligence [7, 8]. Deep learning is the process of learning features. The principle is to describe the characteristics of the model by learning the inherent characteristics and abstraction levels of the original data and then fine-tune the test data.
2.1.1. Deep Learning Framework
With the rapid development of deep learning technology, researchers are no longer satisfied with the basic theoretical analysis, and the requirements for practical applications are getting higher and higher, and the idea of creating a large-scale deep learning framework is being proposed [9]. The in-depth learning framework can contain basic algorithm modules and provide a solid foundation for the subsequent rapid construction of required models or training, coordination, testing, and development of existing models.
2.1.2. Deep Learning Model
After the concept of deep learning was proposed, more and more models appeared, such as deep belief network, convolutional neural network, and recurrent neural network.
(1) Deep Belief Neural Network. Deep belief network is a probabilistic generative model. Unlike traditional neural network models, training data is generated with the greatest probability by observing the joint distribution of data and labels. We can use DBN not only for feature selection and classification but also for data generation [10, 11]. The more common DBN neural network structure is shown in Figure 1.
In Figure 1, there are connections between neurons in the upper and lower layers of the DBN neural network structure, but there is no connection between neurons in the same layer. The neural unit in the optical layer is mainly used to receive input, and the neural unit in the hidden layer is mainly used to capture the high-order data correlation displayed in the optical layer.
(2) Convolutional Neural Network. Convolutional neural network is a feedforward neural network that uses convolution kernel operations instead of general matrix multiplication operations. It has the characteristics of sparse interaction and parameter sharing. In traditional neural networks, any pair of input and output neurons will interact to form a densely connected structure, while in convolutional neural networks, each output neuron will only exist with neurons in a specific area of the previous layer connect weights to achieve the characteristics of sparse interaction [12, 13].
Because of the progressive characteristics of its network structure, the extraction of image features also gradually changes from low-level simple texture features to high-level complex structure features, and the final combination is a feature map [14]. Image features extracted based on this characterization learning method often have good stability, so convolutional neural networks have huge advantages in processing two-dimensional image data.
2.2. Image Recognition Classifier Algorithm
Sparse representation classifier is a very important feature encoding classifier and also a nondictionary learning classifier model. The SRC model is widely used in computer vision fields such as face recognition, object recognition, and scene understanding. The mathematical representation of the SRC model is given below in detail. We assume that there are a total of C categories in the image data set. The dimension of each image sample is , and there are a total of training samples for the -th category. For a test image sample from the -th category, intuitionally, can be approximated by a linear combination of samples in , namely
is the dictionary in the model. It can be seen that the original sample is directly used as a dictionary in the SRC model, instead of retraining a dictionary. If we use to represent the test sample , then can be expressed as
In the actual SRC model, and are known, and the coding coefficient is unknown. We can solve the sparsest solution by imposing the -norm minimization restriction on the linear representation system, so formula (2) can turn into the following optimization problem:
Further, if only column vectors are selected from to represent , then it will be equivalent to the following optimization problem:
Because of the presence of noise, the sparse solutions of problems (3) and (4) can be approximated by solving the following optimization problems:
Furthermore, according to the Lagrangian multiplier theory, the above two optimization problems are equivalent to the following unconstrained minimization problems:
It can be seen from the above algorithm (6) that when solving the sparse coding coefficient , the dictionary based on is , that is, the original training image samples, and these initial samples are not used to train a new criterion based on a new criterion dictionary. In addition, the SRC algorithm uses the size of the reconstruction residual of each category to determine the label of the test sample, and most feature coding classifiers also use the size of the reconstruction error to predict the category, so most feature coding classifiers are based on reconstruction methods.
2.3. Image Enhancement Algorithm for Omnidirectional Vision System
There are many factors that make the image not clear enough. The images collected by the camera are converted from analog to digital, and line transmission will cause noise pollution. Uneven changes in indoor and outdoor lighting will cause excessive concentration of grayscale images, which affects image quality. The shallowest degree indicates that there is noise in the image, which will affect the representation of image details; the more serious degree indicates that the image is blurred, and even the rough outline of the object is difficult to see [15, 16]. Therefore, before processing and analyzing the collected images, we can improve the image quality in a targeted manner, that is, image improvement. Image enhancement processing is to highlight the interesting part of the image, reduce or delete the uninteresting part, thereby improving useful information, obtaining an image that is more in line with human visual acceptance or transforming it into a more suitable machine for processing and analysis Image. There are many image enhancement methods, and they have a wide range of applications.
2.3.1. Selection of Color Space for Image Enhancement
Color space, also known as color coordinate system, is a method of expressing colors with numbers. The color space is usually a three-dimensional model, and colors can be defined by hue, saturation, and lightness. It is the basis of image processing, analysis, and understanding. The purpose of image enhancement is to remove the effect of changing or uneven lighting on the image color. RGB color space is closely related to hardware devices. It is not completely equivalent to the difference in color perception. It is rarely used to process pictures under variable lighting conditions. Therefore, people often convert it into other color spaces that are more consistent with visual characteristics and more stable [17, 18]. Compared with the normalized YUV color space, the HSI color space has independent brightness components, but the HSI color space is the space most suitable for human visual characteristics and is more suitable for images based on the color perception characteristics of the human visual system processing, and the effect of changing lighting response on the color space is mainly concentrated in the brightness of the brightness component.
2.3.2. Selection of Color Space for Target Recognition
The task of target recognition is to separate the orange ball from other colors, such as green, white, and black, and to improve the accuracy of target recognition. Because the RGB color space is not suitable for images with changing illumination, the feature quantity that is more suitable for identifying the target orange ball is mainly selected from the two color spaces of HSI and YUV. HSI color space and YUV color space are two color spaces that are often used in football robot vision systems, and neither of them is specifically for identifying a certain color or a few colors. The two color spaces are merged, and several components that have a better aggregation effect on the orange sphere are extracted from the six components and combined together. Since this paper inputs the color feature quantity into the support vector machine for target recognition, in order to reduce the complexity of the algorithm, only two components are selected from these six components to form a new color model [19, 20]. The separability criterion of the distance between classes within a class is widely used because of its intuitiveness, clear concept, and convenient calculation. However, it is calculated directly based on the distance between various types of samples, without considering their probability distributions, and cannot clearly calculate various overlapping problems, which leads to the final selection of the optimal feature subset when the classification decision is made. Not necessarily optimal. Therefore, the intraclass divergence matrix is introduced as the criterion.
2.4. Robot Rapid Target Acquisition and Distance Measurement
In the autonomous robot soccer game, the robot is a highly intelligent body that collects information from sensors, performs data fusion, and then makes behavior decisions, and then controls the robot to move according to the planned behavior strategy. Many robots complete competition tasks through mutual cooperation. It can be said that it is a microcosm of robot society. As we all know, about 80% of the external world information that people perceive is obtained through visual means [21]. Similarly, in order to make intelligent robots reachable, providing robots with human vision functions is extremely important for the development of intelligent robots. Especially for the quadruped robot football game, vision is its only source of information. Unlike the human eye, the quadruped robot has a very narrow viewing angle, and the amount of information that can be received at the same time is very small.
From the human point of view, it is very simple to understand the surrounding scene correctly, but it is a great challenge for the machine. Therefore, although computer vision has made significant progress in recent years, a large number of technologies and algorithms have been studied and widely used in various fields. However, machine vision technology is still at a very immature stage, and there are still many problems, especially in the following aspects: the 3D scene is projected into a 2D image, and the depth and invisible parts of the information are lost, so the images displayed in different 3D shape image planes. The object will produce the same image. At the same time, the images of the same object taken from different angles will also be very different [22]. Many factors are in the real scene, such as shape, color, lighting conditions, camera, and noise. The image of the object will affect the image and bring difficulties and errors to the recognition of the target object. The amount of information in images is very high, especially color images, which have become a major obstacle to fast image processing. At the same time, the huge amount of data requires huge storage space and will increase the system load.
2.5. Image Color Model
In order to use colors correctly, a color model is also needed. A color can be described by a basic quantity, so to establish a color model is to establish a coordinate system, in which each spatial point represents a certain color.
The most commonly used color model for hardware devices is the RGB model, and the most commonly used color model for color processing is the HSI model, where H represents hue, S represents saturation, and I represents density, corresponding to imaging brightness and image grayscale [23]. These two color models are also the most common models in image technology [24].
Since people’s sensitivity to brightness is obviously stronger than that of color shades, in order to better color processing and recognition, robot vision systems often use HSI color space, which is more suitable for the characteristics of the human eye than RGB color space. A large number of algorithms in image processing and computer vision can be conveniently used in the HSI color space. They can be processed separately and are independent of each other.
The formula for transforming from space to space widely used in the visual processing of soccer robots is as follows:
Among them, the value range of is . When ,
The formula for converting from HSI to RGB has different formulas according to the different sectors of the color circle:
When ,
When ,
When ,
3. Experimental Football Aided Training Robot Based on Image Recognition and Omnidirectional Movement
3.1. Test Subject
In this study, we used an omnidirectional football-assisted training robot. This experiment selects two novice football players with similar technical level from a college from our province Sports Institute, and they are divided into experimental group and control group. The football players in the experimental group use football-assisted robot training when training, and the football players in the control group use the method of training. A commonly used training method, the experiment duration is 3 months, and every member of the team is tested for shooting and dribbling every half month. A total of 22 football players participated in this experiment. The tests of these football players are the main source of data.
3.2. Mean Shift Segmentation Based on Confidence
In low-level image visual processing, combining image segmentation algorithms with edge detection algorithms can improve the quality of segmented images. Combine the edge confidence map into a color image segmenter based on the Mean Shift program. This method can recover areas with weak signals but strong boundaries, so that it can provide more accurate input data for high-level interpretation.
Using the chart in the confidence edge detection, the weight assigned to each pixel (, ) can be defined as
is any value in [0,1]; this parameter is used to control the mixed information of and , so this parameter is called a mixed parameter. When , the weight is 0. When the pixel is closer to the edge, these weights are smaller, which further strengthens the Mean Shift filtering effect.
3.3. Material Handling
Production materials include five categories: text, graphics, audio, animation, and video. Text material is the simplest material. When using text in multimedia, we should focus on the accuracy, simplicity, and functionality of the content. The sound material is generally selected from the existing sound material library and collected from the microphone through the sound card in the computer. The more commonly used method of image material collection is the use of digital camera collection. Flash animation material is currently the most popular two-dimensional animation technology. Video material is a combination of one or more of text, graphic images, sound, and animation.
3.4. Experiment Procedure
Because of its unique characteristics, surveillance cameras make it possible to extract some basic attribute information of the surrounding environment from remote sensing images. These basic information include the location of obstacles, the height of obstacles, and the area of obstacles. The shadow has obvious spectral characteristics in the surveillance camera and has a lower gray value, and the gray value has a strong consistency.
The 22 novice football players we selected were randomly divided into two teams with 11 people in each team. In them, the football players in the experimental group were trained with a football-assisted robot-assisted training method, and the football players in the control group were trained by using common training methods. Before the start of the experiment, we conducted shooting and dribbling tests and scores on these novice football players. After one and a half months, we conducted the second shooting and dribble test and scored. At the end of the experiment, we conducted the shooting and dribble tests and conducted the score again. Score. And analyze and get a conclusion.
3.5. Gather Data
In order to obtain accurate data to compare and analyze the feasibility and effectiveness of this experiment, this paper uses the Cora dataset and the IMB dataset. The statistical data used in this article has a different unit dimension for each index data. After calculating the data in the previous steps, we can get the similarity between users and select several users closest to user interests and preferences to form set . Then, calculate the user score on according to the user’s score on the unrated item in the set . The prediction formula is shown in formula (8)
where is the predicted score of user for the unrated item. In the recommendation system, users scoring preferences are sometimes different. For example, some users are accustomed to giving higher ratings to items, while some are accustomed to giving lower ratings. In order to reduce the difference between users scoring preferences and improve the accuracy of scoring predictions, the average rating of the user is introduced, and the specific form is shown in formula (9)
4. Soccer-Assisted Training Robot Based on Image Recognition Omnidirectional Movement
4.1. System Function Test Analysis
The function test of the system mainly includes main functions such as forward, backward, left turn, right turn, automatically bypass obstacles, and recognize football. The results are shown in Table 1.
In the process of testing the monitoring effect, technical tests such as background sampling and encoding are carried out using specific video formats, and the server and other hardware are repeatedly tested until satisfactory test results are obtained. It can be seen from the table that the system is basically designed and tested successfully after analyzing the required functions.
4.2. Shot and Dribble Test Data Analysis
4.2.1. Pretest Score Data Analysis
The data obtained through shooting and dribbling tests and scoring can calculate the mean, standard deviation, and standard error of the pretest scores of the experimental group and the control group. At the same time, independent samples drawn at the same time, the individual and overall variance scores are not equal waiting for data is also an important condition for the beginning of the experiment, and the results are shown in Table 2 and Figure 2.
The control group’s shooting and dribble pretest scores are tested by independent samples from the experimental group’s shooting and dribble pretest scores. The standard error of the mean is , and the mean and standard deviation of the scores are similar; that is, there is no significant difference in results, which shows that there is no significant difference between the shooting and dribbling skills of the experimental class and the control class before the start of the experiment, which meets the preconditions for the start of the experiment.
4.2.2. Data Analysis of Pretest Scores and Posttest Scores of the Control Group
In order to have a deeper and more accurate understanding of the improvement of football training brought by the omnidirectional football-assisted training robot, we analyzed the data of the pretest and posttest results of the control group and drawn a line chart, as shown in Table 3 and Figure 3. At the same time, the pretest scores of the control class and the posttest scores of the control class are tested in pairs, and data such as the mean, standard deviation, and standard error of the mean are obtained.
It can be seen from Table 3 and Figure 3 that there is not much change between the pretest and posttest results of the shooting and dribble tests of the control group. At the same time, the pretest scores of the shooting and dribbling tests of the control group are tested in pairs with the posttest scores of the shooting and dribbling tests of the control group. The standard error value of the mean is 0.262. The value is greater than 0.05, indicating the control group shooting and dribbling that there is no significant difference between the pretest results of the test and the posttest results of the control group’s shooting and dribbling tests, which also shows that the traditional football training model has little effect on improving the training effect of athletes.
4.2.3. Data Analysis of Pretest Scores and Posttest Scores of the Experimental Group
We analyzed the data of the pretest scores and posttest scores of the experimental group and drew an area chart, as shown in Table 4 and Figure 4. At the same time, the pretest scores of the control class and the posttest scores of the control class are tested in pairs, and data such as the mean, standard deviation, and standard error of the mean are obtained.
It can be seen from Table 4 and Figure 4 that the pretest and posttest scores of the experimental group’s shooting and dribble tests have significantly increased. At the same time, the pretest scores of the shooting and dribble tests of the experimental group are tested in pairs with the posttest scores of the shooting and dribbling tests of the experimental group. The standard error value of the mean is 0.004, which is less than 0.05, indicating that the experimental group’s shooting and dribbling the pretest scores of the test are significantly different from the posttest scores of the experimental group’s shooting and dribbling tests, which also proves the effectiveness of the football-assisted robot-assisted training method.
4.3. Training Effect of Football Match
We analyze the data of all the athletes after training to test and score the football match and draw an area chart, as shown in Table 5 and Figure 5. Reduce the difference between the teacher’s scoring preference, improve the accuracy of scoring prediction, and compare the average score of the experimental group and the control group to get a conclusion.
It can be clearly seen in Table 5 and Figure 5 that the experimental group is relative to the control group. The average and highest scores of the experimental group are higher than those of the control group, and all data are better than those of the control group. If the learning efficiency is the same, the learning time of the experimental group must be longer than the control group, which also proves the use of football assistance robot-assisted training method makes the athletes’ training effect better. It also proves from the side that football-assisted robots have broader development prospects.
4.4. Soccer Robot Target Image Recognition Algorithm
Simulate the construction of the football robot game environment in the laboratory, and change the lighting situation in the laboratory by adjusting the curtains and light tubes in the laboratory, so as to collect the field images under different lighting required for algorithm verification. At the same time, in order to verify the effectiveness of the proposed algorithm, a dynamic test of 80 frames was carried out on the algorithm. The improved RHT algorithm was compared with traditional threshold method, edge detection algorithm, and improved BP neural network algorithm. The experimental results are shown in Table 6 and Figure 6.
From the data trend in the chart, it can be seen that the traditional threshold method and edge detection method have reached more than 90% in recognition accuracy, but the recognition speed is slow; based on the method of Gabor filter + SVM, the recognition accuracy rate reaches 93.4%. However, the same recognition speed is slow and cannot meet the real-time requirements of soccer robot matches, and the improved BP neural network algorithm mentioned above has achieved certain improvements in both the real-time performance of the algorithm and the recognition accuracy. The frame takes 32.1 ms, and the recognition accuracy is more than 96%. The improved RHT algorithm greatly improves the recognition speed while ensuring that the recognition accuracy rate is not reduced. The average time per frame is 5.6 ms, which meets the real-time requirements of the football robot game.
To further illustrate that the improved RHT algorithm has improved recognition speed and accuracy, a detailed comparison is made with the unimproved RHT algorithm. Comparative experiments were carried out at the 10th, 20th, 30th, 40th, 50th, and 60th frames of the collected video images. The experimental results are shown in Table 7 and Figure 7.
According to the data analysis in the chart, it can be seen that the improved RHT algorithm has higher recognition accuracy than the unimproved RHT algorithm, and at the same time, the running time of the algorithm is shorter. This is because the improved RHT algorithm improves the performance of RHT by limiting the radius range, reducing the effective calculation area of the picture, and calculating using the image gradient, which greatly improves the real-time performance of the algorithm. Therefore, the improved RHT algorithm is more suitable and applied in soccer robot games with high real-time requirements.
5. Conclusions
In recent years, robot vision systems have always been a hot and difficult point in computer vision research and digital image processing. The medium-sized football game provides a typical test environment for the research of machine vision-related technology. Expand the application scope of image processing and recognition technology, and enrich the theory of image processing and recognition technology. This dissertation conducts an in-depth study on how to improve the adaptability and target recognition rate of the robot’s all-view system under changing lighting conditions. Recreate the filtered wave factor, and then perform an antagonist transformation to finally obtain an improved image. This method has strong adaptability and high stability and is suitable for image preprocessing in competitions. This paper studies the adaptive target recognition problem of the midsize vision system of soccer robots and has achieved some results. However, due to the complexity of robot vision problems and the limited research time, many of the above issues should be further studied.
For a long time, people have been doing many explorations to understand the world through automatic perception similar to human vision. These explorations are very active in the research fields of computer vision, mechanical imaging, and robot vision. As automation and the latest equipment, intelligent robots can enter the network world and play more and more roles. This has important practical significance for mankind to open up new industries and improve production and living standards. With the continuous development and progress of the system, the structure of the system, the search, and tracking algorithms for target recognition are becoming more and more perfect.
Correspondingly, the hardware and software systems of the omnidirectional mobile robot are designed, the working principles of the main hardware units used on the platform are introduced, and the communication methods between the modules are designed; the software in the program of this document has also been analyzed and tested. After simulation analysis and experimental testing, the module driver in the system is normal, and the communication between the modules is normal and stable, indicating that the system module driver and the communication between them are complete and feasible. The omnidirectional robot positioning and navigation system designed and researched in this field is appropriately designed, has integrated and feasible functions, and can achieve the original design goal.
Data Availability
No data were used to support this study.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work was supported by the Hunan Provincial Philosophy and Social Science Project “Research on the Model and Mechanism of Hunan Foreign Sports Aid under the Background of ‘The Belt and Road’” Project No. 18YBA208.