<p>Many motion sensors can directly acquire human skeletal data, and then extract features on the skeletal data through GCNs (graph convolutional networks) to perform action recognition. However, almost all state-of-the-art (SOTA) methods are offline methods, cannot perform online inference, wasting computational resources. The existing approach to transforming offline action recognition into online action recognition is to reconstruct the network structure of the offline method. This requires developers to have a deep understanding of the algorithm’s network structure and make extensive modifications, which results in slow development. To address the above issue, this paper points out that to convert offline methods to online ones, the key is removing outdated frame features and fusing new frame features. Furthermore, we propose a general and simple model called Encode One Frame (EOF), which achieves feature removal and fusion by a correlation matrix and the guidance of a teacher model. The EOF model has online inference capabilities, requiring only the input of the new frame of the current sample and the features encoded from the old sample. Based on the EOF model, we further propose the You Only Encode One Frame (YOEOF) algorithm to correct the cumulative errors generated during EOF model online inference. By coupling these proposals, YOEOF achieves online inference and outperforms some SOTA methods on public datasets. The deployment at the application level indicates that our method meets the requirements of high accuracy and real-time performance for dangerous action recognition.</p>