Generative Adversarial Networks to Augment Micro-Doppler Signatures for the Classification of Human Activity

Collecting a large amount of data for radar requires a significant amount of time, labor, and money. In deep convolutional neural networks, a small dataset causes the problem of overfitting. We herein introduce the employment of data augmentation using generative adversarial networks (GANs) to solve the data deficiency problem. In this study, we tested the feasibility of using generative adversarial networks to generate micro-Doppler signatures for seven human activities. Moreover, we use produced fake images to train deep convolutional neural networks. We found that the use of augmented data improves classification accuracy. In addition, the quality of GAN output was evaluated in terms of classification accuracy.