Motion Capture and Recognition Based on Hybrid Deep Learning Model for Sports-Assisted Training

Sports motion capture and recognition is essential for enhancing athletic performance, injury prevention, automated sports analytics, and sports-assisted training. This paper presents a deep learning-based approach to accurately capture and classify sports movements using motion sensor data and video inputs. The key challenge addressed is the high variability in motion patterns across different sports and athletes, requiring robust feature extraction and temporal modelling. A hybrid deep learning architecture combining 3D Convolutional Neural Networks (3D-CNNs) and Bidirectional Long Short-Term Memory (Bi-LSTM) is proposed to effectively capture spatial and sequential dependencies in motion data. The model is trained and evaluated on the KTH Sports Action Dataset and UCF Sports Action Dataset, incorporating key pose estimation and optical flow features. Techniques such as data augmentation and attention mechanisms are applied to improve generalization. Experimental results demonstrate high recognition accuracy (96.5 %), with strong performance in precision, recall, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F} 1$</tex>-score, outperforming traditional motion recognition methods. The system is further integrated into a real-time sports analysis and assisted training framework, showcasing its potential for automated coaching and performance assessment, it can be used for daily sports-assisted training