Considering the incapacity of the traditional network structure to fully extract spatiotemporal information in data, a spatiotemporal pyramid pooling model is proposed. A three-dimensional, densely connected convolutional neural network based on dense computation and fusion of spatiotemporal features is designed. The combination of this model with non-local computation operation of features improves the effectiveness of spatiotemporal-feature extraction from videos. The algorithm is applied to the analysis of teachers' actions in classroom scenes. The experimental results show that the designed network structure produces high recognition accuracy on the teachers' actions dataset.