Abstract
The article presents a challenge in automatically monitoring and evaluating physical exercises in real time. The goal of this study is a comparative analysis of the effectiveness of modern deep learning models for determining human posture (BlazePose, YOLOv8 - Pose, MoveNet and HRNet) and to identify the possibilities of their inclusion in the training quality assessment system. The system is built on an RGB video stream and consists of pre-processing, posture points determination, joint angles calculation and training classification (correct/incorrect) stages. From the experimental results, it can be found that HRNet model has the highest accuracy index, but as a result of its high calculation complexity, delay value increases. It had been shown that BlazePose provides high frame rates at low-resource devices. The YOLOv8-Pose model showed the best compromise between accuracy and performance and was suggested as an optimal option for Real-Time Fitness Systems. The method of biomechanical analysis based on calculating joint angles allows to quantify the quality of exercise performing has been proved. The systematic alignment score may be applied in the scope of fitness, sports training and medical rehabilitation and can further be enhanced by adding on 3D body pose models as well as spatio-temporal neural networks.