Abstract: Existing skeleton-based action recognition methods require large computation, which makes them unsuitable for online applications. Aiming at this problem, this paper proposes an online skeleton-based action recognition method with multi-feature early fusion. The algorithm integrated different types of input feature through the early embedding layer and combined the max pooling and hierarchical pooling to extract multi-semantic spatial information. The selection strategy of skeleton sequences was designed based on the characteristics of daily actions. A new 3D skeleton dataset, NTU-GAST Skeleton, was made for online action recognition. Experiments on NTU60 and 120 RGB+D dataset indicate that the proposed method achieves higher recognition accuracy with less computational complexity.