Human Action Recognition using BlazePose Skeleton on Spatial Temporal Graph Convolutional Neural Networks

The trend in multimedia transmission in social media has increased tremendously during the last decade and it is expected to continue growing during the next. Therefore, the need for new tools with the capacity of analyzing this kind of data grows accordingly. In this work, we implement the BlazePose skeleton topology into the ST-GCN model for action recognition. We test our experiments on the UCF-101 and HMDB-51 datasets. These are the first experiments of action recognition using the BlazePose skeleton upon these benchmarks. Moreover, we present an improved skeleton topology based on BlazePose that can enhance the performance achieved by its predecessor. By using the Enhanced-BlazePose topology presented in this study, we improved the results of the ST-GCN model on the UCF-101 benchmark more than 13% in accuracy performance. Finally, we have released the BlazePose skeleton data of the UCF-101 and HMDB-51 from our experiments to contribute future studies in the research community.