Design and Development of Deep Learning Framework for Recognition of Calisthenics Movement

Calisthenics movement recognition has also received considerable interest in recent years given its applications to fitness monitoring, coaching, and rehabilitation. Classic systems to recognize exercises like push-ups, pull-ups, and squats typically employ isolated computer vision methods or shallow deep networks such as CNNs. Those current systems have significant disadvantages in representing fine-grained spatiotemporal patterns, leading to poor recognition performance and challenges to cope with variations in movement execution and style. These limitations reduce model generalization and lead to inconsistent recognition with approximate accuracy rates, precision and other performance metrics. To overcome these limitations, in this paper, research introduces a propoed deep learning architecture that combined model, a strong spatiotemporal feature learning. The CNN layers capture dense spatial representations of every video frame, whereas the GRU layer models temporal patterns throughout movement sequences. Comprehensive experiments on a large-scale as well as exercises filtered from the Kinetics-400 action recognition dataset self-collected and publicly available calisthenics dataset, as well as exercises filtered from the Kinetics-400 action recognition dataset, demonstrate that our proposed CNN+GRU architecture substantially enhances recognition performance. Our model recorded an accuracy of 95.8%, precision of 95.5%, recall of 94.9%, and F1-score of 95.2%, outperforming state-of-the-art methods by about 7-9% across all measures. This significant enhancement proves the efficacy of utilizing both convolutional and recurrent architectures for modeling spatiotemporal differences among exercises. In addition, the introduced framework is computationally lightweight and can be implemented for real-time feedback in training and fitness settings.