Due to the vigorous development of neural networks in recent years, deep learning network has made progressive breakthrough. However, the most important thing about deep learning is enough datasets to achieve the best learning curve. But the process of data collecting often requires a lot of hard work in order to ensure that the data is correct enough to be used. Our study focuses on activity recognition, using modified conditional generative adversarial network, and augmented data with original data to build a dataset with similar features. Comparing with the data augmentation use of generative adversarial network and the original data, its generalization ability and accuracy have been similar with original dataset's result.