Aiming at the problems that traditional human action recognition algorithms cannot effectively suppress spatial background information, the lack of information interaction between networks, and the inability to model global temporal correlation, a human action recognition algorithm of feature fusion Bi-LSTM based on segmentation attention is proposed. First, 30 frames of images are sampled at a certain sampling rate, extract the depth features of the images by split-attention network, and introduce a feature fusion mechanism to enhance the information interaction between different convolutional layers. Then input the depth features into the Bi-LSTM network to model the long-term information of human actions, and finally use the Softmax classifier to classify the recognition results. Compared with the traditional two-stream convolutional network, the accuracy of this algorithm on the UCF101 and HMDB51 datasets is increased by 6. 6% and 10. 2%, respectively, which effectively improves the recognition accuracy.