Radar Dataset Synthesis Approach for Gesture Recognition

In recent years, gesture recognition using radar has attracted a lot of attention. Machine learning is often used as an approach to solving gesture recognition problems. Its main disadvantage is that it requires a large amount of data to train the machine-learning models. Collecting this data using radar can be a painstaking and lengthy process. Therefore, in this paper, a generative adversarial network-based (GAN) methodology for generating hand gesture synthetic data sets is proposed. A GAN model consists of a discriminator and a generator, which compete with each other. To avoid mode collapse of the GAN, synthetic gesture samples are generated while training. Experimental results indicate that the synthetic dataset and the real dataset have a high degree of similarity. Through the training of the neural network with the synthetic data and the testing of the model with real data, the recognition accuracy reaches 95.31%. Moreover, the synthetic data effectively augments the original data. Recognition accuracies are better than those obtained using original data only.