Abstract
In view of the poor effect of current sports dance movement detection, this paper puts forward a sports dance movement detection method based on posture recognition. By using the principle of posture recognition to obtain the characteristic information of sports dance movement, construct the action characteristic evaluation system, optimize the action detection algorithm combined with posture recognition technology, and realize the classified detection of sports dance movements. Finally, experiments show that the sports dance movement detection based on pose recognition has high practicability and effectiveness in the process of practical application and fully meets the research requirements.
1. Introduction
Potential recognition is the main way to help learn and understand human actions and behaviors. Human motion analysis and action information preservation can be realized through human posture recognition. For example, in the process of dance movement teaching, students or coaches can standardize the movements through the results of human posture recognition. Human posture recognition can also gather and store crucial information about dance motions, reducing the chance of dance disappearing through the inheritance process [1]. To accomplish the goal of training, the current sports dance movement detection techniques are typically basic via repeated viewing of recordings or seeking the coach’s personal correction and advice. This not only takes a lot of time and effort but also presents learners with certain obstacles and hurdles. At present, the main process of human pose recognition includes three steps: data acquisition and preprocessing, human feature extraction and construction, and action recognition [2]. Among them, the extraction and construction of human feature is the key to human pose recognition, but the current feature extraction and construction methods usually have the problem of low accuracy, including bottom feature tracking methods and semantic-based methods. The traditional bottom-level feature-based tracking method can effectively detect and predict the abnormal events in the video by analyzing the composite motion features, which improves the accuracy of human posture recognition to a certain extent, but there is still the problem of low recognition accuracy. In order to solve this problem, a detection method of sports dance action based on pose recognition is proposed [3]. The dance action detection target of a complex dance scene is realized by selecting the characteristics of key points of human bones and through the residual block automatic action detection method. Combining dance training with posture recognition is what sports dance action detection based on pose recognition means. Posture recognition, bone tracking, and other technologies are used to capture the entire training action, dance action data are collected, the data are matched and compared with the standard action, and the gap between the user’s action and the standard action bone direction is analyzed, so that the user can more intuitively find the gap. This method allows the majority of dance instructors, dancers, and dance learners who have a high need for instruction and self-study to properly adapt their nonstandard dance posture in training [4]. At the same time, it is not constrained by time or geography. It may be used to train at any suitable time and location, expand teaching methods and means, and encourage the continued development of dance fundamental training and other dance learning via digitalization and informatization.
2. Research on the Sports Dance Movement Detection Method
2.1. Dance Movement Feature Extraction Method
Pose recognition, key point feature processing, and motion classification are the three aspects of the proposed dance motion detection system based on pose recognition. To begin with, the input picture is scaled down to 368 × 368 pixels, and the input posture recognition network is utilised to identify the human body’s main points [5]. The residual network then detects the human body area based on the contour value of the key spots of the human body. Finally, the classification of dance movements can be realized by integrating key point feature classification and image classification. The specific flow of dance motion detection based on pose recognition is shown in Figure 1. Its key node classification network is composed of three branches: key point feature extraction, image classification, and fusion.
After detecting and tracking the human body, we need to make the next pose. Human posture can be divided into direct method and indirect method. The direct method is to establish a human posture model, compare the collected human data with the model one by one, and find the one with the most similar appearance [6]. The indirect method is based on the prediction generation model. It predicts the feature contour according to the collected human body data and classifies the human posture according to the feature contour. Through human posture recognition, people can realize the posture communication with the computer. In recent years, gesture recognition has been popularized and applied in many industries [7]. In today’s highly developed modern competitive sports, computer technology has been paid much attention by sports workers. The “scientific” training of athletes has been verified and developed in all kinds of sports, forming a complete and systematic training system, as shown in Figure 2.
However, there are still many restrictions on the technical conditions of pose recognition. Various special interactive devices based on pose recognition are quite expensive. The existing interactive methods are not humanized, the operation is complex, and the real-time accuracy of the system is not high [8]. These constraints greatly limit the popularization and application of posture recognition technology in the field of practical physical training. With the continuous progress of pose recognition technology, more innovative and more applications can be expected in the near future.
2.2. Description of Movement Characteristics of Sports Dance
Posture recognition technology is to observe the motion law and characteristics of human body from the perspective of geometry, that is, to describe and study the law of human body position changing with time through physical quantities such as position, speed, and acceleration, without considering the reasons for the change of human body or instrument position and motion state [9]. In order to reduce the complexity of research, we usually use only the part of all motion parameters to describe the specific motion. Unmarked human posture recognition has been a research hotspot in order to expand the recognition range from gesture to the whole body. However, because to the intricacy of human mobility and the impact of occlusion light, recognising the unmarked human stance is very difficult [10]. Iterative prediction is performed using the posture recognition network. The mechanism of prediction calculation is as follows:Here, is the convolution operation corresponding to the branches. In order to avoid the gradient vanishing problem during the training of the network, the LSS loss function is usually added in the calculation process, as shown inHere, represent the confidence diagram of key points and the predicted value of PAFS, respectively; is the key point confidence map and the real value of . The maximum value of the personal confidence map s at the jth position of the kth person can be obtained by using the real confidence map predicted by the maximum value operation, as shown in
The theoretical basis of the adjacent frame difference method is the difference operation and , which judges the contour of the moving object by analyzing the difference between the two frames in the video stream [11]. This method can effectively judge moving objects and can be applied to object recognition of multiple objects and multiple videos at the same time.
In theory, this process can successfully extract cloud moving objects, but in practise, ensuring the quality of the video obtained is problematic. The surrounding surroundings and the camera itself will introduce a significant amount of noise into the gathered data [12]. As a consequence, the difference operation result must be processed to remove the impact of noise. Because noise has a Gaussian distribution, we may use a threshold to reduce the impact of noise on difference operations. For further information, see the following formula:
The motion retrieval can be divided into two parts: offline and real-time. After establishing the motion database, the dance designer can establish the directed graph offline part of the motion database from a large number of real motion clips [13]. The offline part mainly retrieves the “difficult movements” from the real motion database and calculates the time point of their emergence. Then, the dance designer subjectively judges the action segments that can be connected before and after the “difficult action,” also known as “connecting action.” The real-time retrieval part only needs to search in depth or breadth according to the connection relationship between points and lines on the established action “graph.” The real-time part is used to select a series of vertices and edges from the established directed graph according to the difficult actions selected by the dance designer and add these calibrated difficult actions and connecting actions to the directed graph structure [14]. A “difficult action” can only be automatically found from the “initial motion segment” once the similarity between motion frames has been determined. The approach based on quaternion is utilised to define the interframe distance in this research.Here, represents the important positive number of the first joint of the human body and is two quaternions. represents the distance. The distance between frames can be obtained through equations. The distance value indicates the similarity between the two frames. The smaller the value, the higher the similarity. By calculating the distance between the “difficult action” frame and the “Original Motion” frame, the position of the “difficult action” in the “Original Motion” can be located. If the video image is regarded as composed of moving target and background , then
There are many kinds of pose recognition technologies, involving a wide range of fields. We mainly discuss pose recognition operations including motion path editing, motion blending, and pose recognition constraints [15]. Motion blending technology can connect multiple motion segments into a new motion segment and use linear interpolation method to calculate the mixed motion of the overlapping part. Then,where represents the frame, is an integer of an adjusted frame sequence number, is a weighting coefficient, and ∈ (0, 1). The translation position of the human body root node in the second frame can be obtained by mixing the motion of one motion from frame a to frame a+ and another motion from frame B = to frame B; then,
Based on this, the quaternion can represent the rotation degree of human joint in frame during synthetic motion and each pixel information can be obtained by body component recognition reasoning. The density estimator of body components is defined as
Here, refers to three-dimensional space coordinates, refers to the number of pixels, refers to the pixel weight, Y refers to the projection of pixel x into world space, and B represents the width of each component [16]. Then, the pixel reasoning probability and spatial area probability are balanced.
This technique not only enhances joint prediction accuracy but also assures that density estimation is done at a consistent depth. The standard dance posture database is constructed as a template, and the standard dance posture data are gathered to the computer using posture recognition. The acquired data on trainer dance posture is compared and analysed against the template posture, and the trainer’s dance posture is intuitively assessed, ushering in a new era of digital dance instruction. Ballet is the most standardised and scientific form of dance [17]. It emphasizes movement correctness and employs high-standard motions to portray dancers’ feelings and ideas. A obligatory and important course of dance training is basic sports dance training. Its distinguishing feature is that it clarifies the movement’s uniformity, systematicity, and scientificity.
2.3. Implementation of Sports Dance Movement Detection
Sports dance can be divided into two groups: Latin dance and modern dance. Latin dance is divided into five kinds: Samba, cowboy bullfight, rumba, and cha cha. Sports dance is divided into five kinds: Tango, trot, Foxtrot, Waltz, and Vienna Waltz. They have their own technical characteristics and action skills. In a word, sports dance is a kind of sports with strong standardization and appreciation [18]. From the perspective of sports dancers themselves, they can be divided into individual dancers, Latin dancers, modern dancers, decathlons, etc. Through the analysis of the above different dances and their respective characteristics, it can be found that the overall characteristics of Latin dance are passionate, relaxed, and lively, giving people a romantic and free feeling, and its dance movements are mainly hip movement and foot movement, with many movement frequencies and fast speed. Therefore, Latin dance has certain requirements for the speed, explosive power, and control of dancers [19–22]. Sports dancers should have high strength and physical quality. They should carry out comprehensive training for all parts of the body, especially the upper and lower limbs, waist, abdomen and back, and ankle and knee joints, so as to ensure the perfect completion of technical movements at the same time of successfully completing the arrangement routine and accurately expressing music. A dance-aided training system based on posture recognition is developed, which combines dance training with posture recognition, human bone recognition, and tracking technology to recognize posture. Data processing and data analysis are the main components of this system [23–25]. The collected data include the movement data of dance coaches and the movement data of trainers. By collecting the movement information of dance coaches, a set of database of standard dance postures is established. The primary goal of data processing is to fix occluded joint points and restore human bone information. The data analysis section compares the trainer’s exercise data with the standard action’s exercise data and provides training suggestions based on the comparison findings, allowing the trainer’s level to develop fast. Figure 3 depicts the structure for dance movement detection training.
Several vertices on the motion graph can be obtained from several difficult actions specified by the dance designer. Starting from these vertices, the “edges” connecting the vertices can be obtained. In fact, these “edges” are the connecting actions used to connect the “difficult actions.” Since there may be more than one edge connecting the two vertices, then multiple connection actions can be provided to the dance designer as candidates for connection actions, as shown in Figure 4.
Listed in Figure 5 are 20 “difficult movements” selected from the complete set of unarmed movements required in college sports dance courses. The peripheral rectangular box represents the movement, nine “connecting movement” samples, and the elliptical box represents the connecting movement. The dance designer may further define a new motion route for the created motion by using the “motion path editing” module of the tool set layer. At this moment, the major focus is on collision detection and route avoidance between virtual persons in order to re-edit the path. The dance designer may examine the created motion for additional adjustment after finishing the abovementioned task. The dance designer may save the motion after finishing the motion design for each section and then create the motion for the next segment or quit the system. The system will automatically synthesize the transition motion between the motion segments matching to the nearby segments in order to achieve a smooth transition between the planned multisegment movements. These designated “connection actions” are fed back to the dance designer as “candidate connection actions,” and the dance designer will add the connection actions to the “candidate connection actions.” Then, use the “motion connection” of the tool layer to connect the action segments. At this time, the overall sports meeting appears very rough. Use the “motion mixing” to smooth the motion. Finally, use the “sports dance action synthesis” to add the equipment. The joint sequential use of the three modules can obtain the smooth continuous motion segments that meet the requirements of the dance designer.
Any two challenging actions may be linked using the elliptical box’s connecting action to create a new motion segment, segment the motion segment and categorise the emotion, and fetch the actions in the poser action library that match the emotion type of the sub segment. If there is more than one, find the action with the highest matching degree according to the matching value. Repeat this process and finally get the action sequence corresponding to the music/action script, as shown in Figure 6.
The reasons for the promotion of vestibular function by sports dance are not single. The possible stimulating factors are as follows: sports dance makes skeletal muscle play an important role in regulating the redistribution of blood and can promote the improvement of nerve regulation, blood circulation, and body fluid regulation. Other studies believe that the music of sports dance can make the resonance of various organs and tissues of the body, coordinate the functions of various organs, and concentrate on dancing with the music, which can play a role in strengthening the brain, which comes from people in the dance industry. From the analysis of these different angles, the conclusion is that sports dance can effectively improve vestibular function. Sports dance promotes the improvement of proprioception. The following are the reasons: sports dance teaching enables female college students to master certain motor skills, and the formation of skills is to establish a complex and chained proprioception reflex, so that the inhibition and excitation processes of the cerebral cortex can be strictly transformed, and so female college students’ perception ability to complete actions, such as strength, amplitude, space, and time, can be improved. This can improve the precision and coordination of the movement. This procedure must be repeated because of the learning properties of sports dancing talents. To complete the action, it is important to precisely determine the strength, force duration, spatial direction, and amplitude. We may continually modify the sensitivity and make the proprioceptive sensor provide a more suitable action reflection in learning practise by using proprioceptive feedback. With greater experience and deeper learning, the cerebral cortex’s excitation and inhibition processes may become more rhythmic and precise in time and space, resulting in more precise and coordinated behaviour. The “stereotype of motor power” tends to consolidate and form the automation of action, and a good proprioception is formed. Moreover, it is often practiced again and again, and the perception is constantly fed back. If the proprioceptive sensitivity is constantly revised, the proprioceptive sensitivity will be more accurate.
3. Analysis of Experimental Results
The experiment verifies the proposed algorithm on PyTorch open-source neural network. The network framework contains some commonly used data sets, which are convenient for learning NN module packaging loss function and torch optim module packaging optimization function. The experimental data set includes 4405 images extracted from a concert video and dance video. 4340 images in the data set are randomly selected as the training set, and the remaining 345 images are used as the test set. Under the same environment, the recognition of sports dance feature data set of the traditional method and this method is compared and analysed, as shown in Figure 7.
Based on the comparison and analysis of the above detection results, it is not difficult to find that this method has higher recognition accuracy of feature data compared with the traditional methods. Further comparative analysis is made on the fit between the trainer’s action recognition and the standard action joint coordinates in the process of dance movement, as shown in Table 1.
Collect the motion information of the three hand movements of the dance movement, and calculate and decompose the angle changes of each joint in the movement for detection and analysis, as shown in Table 2.
Based on the above experimental test results, it can be seen that there is a more or less gap between the trainer’s action and the standard action and the trainers at different levels can train according to their own level. In order to ensure that the whole experiment is not affected by other factors, 40 students were tested for functional movement and special skills before the experiment and the measured data were distributed in the “s” mode. There were 20 students in the experimental group and the control group, respectively. Table 3 shows the results of functional movement detection and special skills test of students in the experimental group and the control group.
There are 7 test items in the functional movement test (FMS). The full score of each action is 3 points, and the total score is 21 points. From Table 4, we can see that the data gap between the experimental group and the control group is not too large. Through the independent sample t-test, we can see that the scores of the 7 items and the value of the total score are >0.05, indicating that there is no difference between the experimental group and the control group in the 7 items of the functional test before the experiment. It shows that the test results are more accurate. In order to verify the superiority of this algorithm, this algorithm, the traditional identification algorithm residual network, four-channel algorithm network, and the calculation Hu moment algorithm are used for identification test on the test set. The results are shown in Table 5.
It can be seen from Table 5 that compared with the other algorithms, the accuracy of our algorithm is higher, reaching more than 92%, indicating that the recognition effect of our algorithm on dance movements is ideal and has a high accuracy. In comparison, the running time of this research algorithm is basically not greatly improved. Therefore, this algorithm has higher efficiency and better algorithm performance.
4. Conclusion
To sum up, the dance motion detection method based on pose recognition designed in this study can automatically detect the dance motion in complex scenes by combining the bone key point information and residual network and the recognition accuracy is significantly higher. Compared with the traditional dance motion recognition methods, the recognition accuracy of our algorithm is the highest and the recognition efficiency of our algorithm is almost not affected by the number of characters in the image. The recognition efficiency is faster, which can meet the needs of actual dance movement detection, and has a certain reference significance.
Data Availability
The data used to support the findings of this study are included within the article.
Conflicts of Interest
The authors declare that they have no conflicts of interest.