Human Activity Recognition Based on WRGAN-GP-Synthesized Micro-Doppler Spectrograms

In recent years, deep convolutional neural networks (DCNNs) have demonstrated the prominent performance in the radar-based human activity recognition. However, collecting and labeling the radar data are usually expensive and time-consuming. It is very challenging to obtain sufficient training data in practice. The DCNNs often suffer from the overfitting problem due to the lack of data and cannot effectively exert their performance. A large number of synthetic micro-Doppler spectrograms similar to the real micro-Doppler spectrograms can be produced by generation adversarial networks (GANs) to increase the training dataset. As a fact of matter, the quality and diversity of synthetic spectrograms are particularly important for the training process of DCNNs. In this paper, we propose a more stable and effective Wasserstein refined generative adversarial network with gradient penalty (WRGAN-GP) to generate the spectrograms for expanding the training dataset. The DCNN is trained to recognize human activities through the new training data composed of real spectrograms and the augmented synthetic spectrograms. The experimental results show that compared with existing GANs-based spectrogram augmentation methods, the proposed WRGAN-GP method provides higher stability and accuracy in generating the augmented spectrograms. Accordingly, the classification accuracy of human activities can also be greatly increased.