Skeleton-Based Action Recognition and Evaluation Using Dynamic Time Warping Algorithm Enhanced by Spatial-Temporal Feature Engineering Techniques

Contemporary advancements in human action detection and recognition offer significant benefits by reducing professional workloads and empowering individuals with convenient and autonomous self-training opportunities. Our methodology revolves around extracting skeletal information from videos and extracting relative features. These features aim to capture not only the spatial relations among human skeleton joints but also valuable long-term temporal information. Furthermore, we utilize a dynamic time warping-based k-nearest neighbor algorithm to recognize and evaluate the similarity between individuals replicating exercises and the guidance provided by a reference instructor. This innovative approach enables effective home-based virtual coaching without the need for large-scale training data, thereby empowering users to reference and refine their movements independently.