Omni-TransPose: Fusion of OmniPose and Transformer Architecture for Improving Action Detection

The field of computer vision research has been experiencing rapid and remarkable development in recent years, aiming to analyze image and video data through increasingly sophisticated machine learning models. In this research domain, capturing and extracting relevant features plays a crucial role in approaching the detailed content and semantics of image and video data. Among these, skeleton data, with the ability to represent the position and movements of human body parts, along with its simplicity and independence from external factors, has proven highly effective in solving human action recognition problems. Consequently, many researchers have shown interest and proposed various skeleton data extraction models following different approaches. In this study, we introduce the Omni-TransPose model for skeleton data extraction, constructed by combining the OmniPose model with the Transformer architecture. We conducted experiments on the MPII dataset, using the Percentage of Correct Key Points (PCK) metric to evaluate the effectiveness of the new model. The experimental results were compared with the original OmniPose model, demonstrating a significant improvement in skeleton extraction and recognition, thereby enhancing the capability of human action recognition. This work promises to provide an efficient and powerful method for human action recognition, with broad potential applications in practical scenarios.