Action recognition is an important research topic in computer vision. As hardware equipment becomes increasingly advanced, research on action recognition using deep learning methods has increased, but equipment may be affected for various reasons. Limitations and the efficiency of everything in our modern fast-paced society make how to produce better results with limited resources in a short period of time extremely important. This thesis selects a smaller-structure C3D convolutional network, uses Transfer Learning to shorten training times, and compares the optimizer, Activation Function, and weight initialization method to provide the best combination using migration learning theory and Momentum(use Nesterov) to optimize the model with Xavier weight initialization and Activation Function RELU, which is better with short times and equipment constraints.