In this paper, we proposed an action recognition algorithm based on the adaptive fusion of RGB and skeleton features to efficiently improve the accuracy of action recognition. However, the conventional action recognition algorithms based on RGB and skeleton features generally suffer from the inability to effectively utilize the complementarity of the two features and thus fail to focus on important frames in the video. Considering this, we first used the bidirectional long short-term memory network (Bi-LSTM) combined with a self-attention mechanism to extract spatial-temporal features of RGB and skeleton images. Next, we constructed an adaptive weight computing network (AWCN) and computed these adaptive weights based on the spatial features of two types of images. Finally, the extracted spatial-temporal features were fused by the adaptive weights to implement action recognition. Using Penn Action, JHMDB, and NTU RGB-D dataset, the experimental results show that our proposed algorithm effectively improves the accuracy of action recognition compared with existing methods.