Abstract: Aiming at the task of video action recognition, we proposed an action recognition method based on two-stream network. Firstly,
 a sparse sampling strategy was adopted to avoid the redundant information of adjacent frames from affecting the recognition effect. Secondly, the convolutional neural network was used to predict the optical flow map,  improve the acquisition efficiency of  the optical flow map and reduce the amount of calculation. Finally, the residual network was used to extract the completed video information and simplify the training process of neural networks simultaneously. In order to verify the effectiveness of the two-stream action recognition network, we carried out comparative experiments on two classical data sets. The experimental results show that the proposed two-stream action recognition network has good recognition effect and  can be applied to intelligent video surveillance, human-computer interaction, public security and other fields.

Key words: action recognition, convolutional neural network, two-stream network, sparse sampling