With the increasing emphasis on physical and mental health, aerobics and yoga have become two of the most popular forms of physical activity worldwide. However, traditional training methods rely heavily on professional instructors, which are limited by high costs, uneven teaching quality, and lack of real-time feedback. To address these issues, this study proposes a comprehensive framework for computer-based motion analysis and personalized training recommendation, integrating computer vision, deep learning, and biomechanical principles to achieve accurate motion evaluation and adaptive training guidance. First, a multi-modal motion capture system is designed to collect 2D/3D human skeleton data and surface electromyography (EMG) signals during aerobics and yoga exercises. Second, a novel cascade two-stream adaptive graph convolutional neural network (Cascade 2S-AGCN) is proposed for motion feature extraction and error recognition, which outperforms existing methods in terms of recognition accuracy and real-time performance. Third, a personalized training recommendation model based on reinforcement learning (RL) is established, considering user physical characteristics, training goals, and motion performance to generate adaptive training plans. Extensive experiments are conducted on a self-built multi-view aerobics and yoga dataset (MAY-Dataset) and public datasets (3D-Yoga, M3GYM), involving 80 participants with different fitness levels. The experimental results show that the proposed motion analysis model achieves an average recognition accuracy of 96.87% for aerobics movements and 97.53% for yoga poses, with a motion error detection rate of 95.21%. The personalized recommendation model significantly improves user training adherence (by 32.4%) and training effect (by 28.7%) compared with traditional uniform training plans. This study provides a scientific and intelligent solution for aerobics and yoga training, promotes the digital transformation of fitness services, and lays a foundation for the development of intelligent fitness systems.
Wen et al. (Fri,) studied this question.