Fitness with AI-trainer: Automatic Gym Exercise Monitoring

In recent years, the popularity of sports, especially strength sports and yoga, has increased worldwide. In these sports, technique plays an important role. Improper technique can make the training not only ineffective but also dangerous. In this paper, a study and comparison of models for Pose Estimation task has been done. The models considered were YOLOv8n-Pose and BlazePose. BlazePose was used to develop the solution as it is the best in terms of quality to speed ratio, which is important for using the model locally on mobile devices. During the research, a system was developed that corrects the technique of some exercises and also collects statistics about the workout, such as the number of repetitions or the time spent in an exercise with the correct technique. Several exercise classification models have also been trained to preserve the training history. The paper compares algorithms based on convolutional neural networks (CNNs) and algorithms based on decision trees. The algorithms described above are planned to be used in the development of a self-paced strength sports and yoga application. The top-level architecture of the application is also described in the work, but has not yet been implemented at this stage of the project. As a result of the work, the first part of the application was developed, for which machine learning is responsible: key point detection and exercise classification. In addition, the first part of the application was developed.