Human posture estimation and action recognition on fitness behavior and fitness

Abstract

In daily fitness activities, posture recognition and motion capture are extremely important. This study aims to explore skeleton-based action recognition technology and compare the effects of using only skeleton information and fused image features. We introduce an improved spatiotemporal pyramid graph convolutional network, which enhances the performance of the model by introducing edge importance scores and multi-level feature representation when dealing with specific action recognition tasks. The proposed model is tested on the UW-IOM and TUM kitchen datasets to verify its applicability in real-world scenarios, the mAP of UW-IOM reached 85.89, achieving the best result and proving the effectiveness of the model. Through experiments, we confirm that in certain cases, the accuracy of skeleton information is the key to improving action recognition performance, while in scenarios where image information is not rich or skeleton information is inaccurate, the fusion of image features can provide performance supplements. In addition, we demonstrate the advantages of our method by comparing with existing methods.