From Keypoints to Actions: Real-Time Motion Analysis Using MoveNet and Lightweight Conv1D Networks

Motion analysis provides insight into a subject’s physical condition or activity status and can help prevent hazardous movements, reduce the risk of accidents, or support sports performance evaluation. This study adopts a computer vision-based approach combined with deep neural networks for motion classification, avoiding the limitations of wearable sensors. The dataset consists of publicly available fitness instructional videos, from which 23 decomposed actions were manually annotated. The proposed two-stage model uses MoveNet to detect 17 human joint coordinates, which are then passed to a Conv1D neural network for classification. Triplet-Center Loss is applied to enhance class separation. Experimental results show that the MoveNet + Conv1D model achieved a training error of 0.2 with 100% accuracy and a test error of 0.62 with 92% accuracy. The use of Triplet-Center Loss resulted in well-separated class clusters in t-SNE visualizations. The proposed model demonstrates high precision and efficiency, making it suitable for future applications in sports analytics, action recognition, and workplace safety monitoring.