With the infiltration of digital computer technology into the animation industry, the traditional animation production process began to transition to computer animation, the most representative feature of which is the use of 3D and CG technology. However, in the process of transformation from classical two-dimensional planar animation to three-dimensional animation, there are also many problems, such as the audience does not adapt to the viewing habit, the low production cost leads to the poor rendering effect of three-dimensional objects, and the low frame number of three-dimensional animation leads to the incoordination of motion. Drawing 3D models with real action recognition technology is an excellent way to make 3D models more realistic and smooth. At present, the main algorithm of action recognition is convolutional neural network algorithm. However, there are still some problems such as large number of convolution parameters, long operation time and difficult parameter adjustment in 3D network model. In order to solve these problems, the original 3D convolutional neural network model is improved in this paper. The model divides the traditional 3D CNN network structure into spatial flow and time flow for data calculation, and uses the design idea of ResNet network for reference to reduce parameter Settings and avoid gradient disappearance. The experimental results show that the training speed of the model is greatly improved under the condition that the recognition accuracy is guaranteed.