随着信息技术的飞速发展，人工智能的理论和技术发展日趋成熟，智能化产 品在人们生活中的各个领域都得到了应用和普及。计算机视觉识别技术作为人工 智能和计算机科学的重要分支，在近年也取得了长足发展，受到社会广泛关注， 并被应用到车牌识别、人脸识别等与人们日常生活息息相关的领域。随着计算机 视觉理论与技术的不断发展，计算机模拟人的视觉机理来获取和处理信息的能力 越来越强，为应用在其他需要图像和动作识别的领域，尤其是体育运动与赛事， 提供了更多技术实现的可能性。计算机视觉识别技术在体育领域中已经有了许多 有益尝试和探索，主要被应用在虚拟现实人机交互、田径比赛运动员终点裁判， 篮球比赛数据记录和艺术体操运动员运动轨迹跟踪等方面。但运用计算机视觉识 别技术对运动员进行姿态识别却少有人研究和应用。 在艺术体操领域，现有教学和比赛的打分主要依靠教练员和裁判员的的人工 观察和个人经验来进行，这种方式使得评判工作不可避免地带有一定的主观性， 且对裁判员的专业要求极高。另外，艺术体操赛事时间较长，评委在评判时常常 伴随着工作量大、精力高度集中等特点，大量的评判工作难免导致或多或少的误 差出现，影响评判的效率和公平性。因此，寻找一种更加有效的方法来辅助艺术 体操的评判工作意义重大。 对艺术体操动作的准确识别也是提升艺术体操比赛、教学和训练质量的关 键，而计算机视觉识别技术可以更好地实现对运动员动作与姿态的精准识别与分 类。将这一科学技术应用于艺术体操体育教学的训练和比赛中，不仅可以辅助教 练员及体育教师的教学工作，提高其教学与评判水平，还将改进艺术体操的传统 评判方法，帮助裁判进行更加客观公正的评判，增进了艺术体操赛事的公平、公 正性。 因此，本文将聚焦艺术体操领域，基于计算机视觉识别技术，在 TensorFow 框架下，利用已开源的代码条件下，把 CNN（卷积神经网络）、Body-pix 技术、 LSTM（长短时记忆网络）运用到本研究，建立艺术体操动作识别系统，对艺术 体操中反跨跳和俯平衡动作的识别进行研究与探索。借助深度学习技术实现艺术 II 体操的反跨跳和俯平衡动作的自动识别，主要研究内容包括以下方面： （1）通过文献资料法梳理计算机视觉识别、艺术体操等领域的主要理论与 国内外研究进展，为后续建立素材库，进行系统实验,寻找相关理论依据与基础； （2）建立俯平衡和反跨跳的评判方法；（3）对图片进行分类整理，建立艺术体 操动作素材库；（4）根据动作的不同及评分标准的不同对图片进行标注，并对 艺术体操的动作进行打分；（5）构建艺术体操视觉训练模型；（6）利用 Body-pix 对人物进行提取；（6）利用 LSTM 分对 Body-pix 提取出的热图进行识别训练； （7）将未曾训练过的图片输入该系统进行识别测试，并不断调整识别参数，提 高识别的准确度。 经测试，本研究构建的艺术体操动作识别系统可以有效识别艺术体操反跨跳 和俯平衡动作，识别准确率达到 90%以上，该系统可以作为重要的教学和竞赛辅 助系统应用于艺术体操教学和竞赛中。

With the rapid development of information technology, the development of artificial intelligence theory and technology has become increasingly mature, and intelligent products have been applied and popularized in various fields of people's lives. As an important branch of artificial intelligence and computer science, computer vision recognition technology has also made great progress in recent years, and has been widely concerned by the society, and has been applied to areas such as license plate recognition and face recognition that are closely related to people's daily lives. With the continuous development of computer vision theory and technology, the ability of computers to simulate human visual mechanisms to acquire and process information is becoming more and more powerful, and is provided for applications in other fields that require image and motion recognition, especially sports and events. More possibilities for technology implementation. Computer visual recognition technology has many useful attempts and explorations in the field of sports. It is mainly used in virtual reality human-computer interaction, track and field athletes' final referee, basketball game data record and rhythmic gymnastics athlete's trajectory tracking. However, the use of computer vision recognition technology for gesture recognition of athletes has rarely been studied and applied. In the field of rhythmic gymnastics, the scoring of existing teaching and IV competition mainly depends on the manual observation and personal experience of the coaches and referees. This way, the judging work is inevitably subjective and subjective. The requirements are extremely high. In addition, the rhythmic gymnastics events take a long time. The judges are often accompanied by the characteristics of large workload and high concentration of energy. A large number of evaluations will inevitably lead to more or less errors, which will affect the efficiency and fairness of the judgment. Therefore, it is of great significance to find a more effective way to assist the judging of rhythmic gymnastics. Accurate recognition of rhythmic gymnastics movements is the key to improving the quality of rhythmic gymnastics competition, teaching and training, and computer vision recognition technology can better realize the accurate recognition and classification of athletes' movements and postures. Applying this science and technology to the training and competition of rhythmic gymnastics physical education can not only assist the teaching work of coaches and physical education teachers, improve the level of teaching and evaluation, but also improve the traditional evaluation methods of rhythmic gymnastics and help the referees to do more. Objective and fair judgments have improved the fairness and impartiality of artistic gymnastics events. Therefore, this article will focus on the field of rhythmic gymnastics, based on computer vision recognition technology, under the Tensor flow framework, using CNN (convolution neural network), Body-pix technology, LSTM (long and short time memory network) under the open source code conditions In this study, the rhythmic gymnastics movement recognition system was established to study and explore the recognition of anti-jumping and declining movements in rhythmic gymnastics. The deep learning technology is used to realize the automatic recognition of the anti-jump and the balance movement of the rhythmic gymnastics. The main research contents include the following aspects: (1) Through the literature method to sort out the main theories in computer visual recognition, rhythmic gymnastics and other fields and the research progress at home and abroad, for the subsequent establishment of the material library, systematic experiments, to find relevant theoretical basis and basis; (2) establish the balance and V anti-balance (3) classify the pictures, establish a rhythmic gymnastics action material library; (4) mark the pictures according to different movements and different scoring standards, and score the movements of the artistic gymnastics; (5) Build a visual training model of rhythmic gymnastics; (6) use Body-pix to extract characters; (6) use LSTM to identify the heat map extracted by Body-pix; (7) input untrained images into the The system performs the identification test and continuously adjusts the identification parameters to improve the accuracy of the recognition. After testing, the rhythmic gymnastics motion recognition system constructed in this study can effectively identify the rhythmic gymnastics anti-cross-skip and down-balance movements, and the recognition accuracy rate is over 90%. The system can be used as an important teaching and competition auxiliary system in rhythmic gymnastics teaching and In the competition.