Data Augmentation in Time and Doppler Frequency Domain for Radar-based Gesture Recognition

This paper presents novel data augmentation techniques for gesture recognition based on multistatic Doppler spectrograms. Contrary to intricate approaches such as generative models or simulations, the presented methods generate additional samples by manipulating data in time and Doppler frequency domain. Thus, they resemble well-known techniques from computer vision tasks, but produce physically plausible data. Moreover, based on a comprehensive dataset a thorough investigation of the impact of training data extent on cross-subject validation accuracy is conducted. It is demonstrated that accuracy tends to saturate with growing training sets, and that the techniques presented in this work can help to reach the accuracy at saturation with only half the number of measurements.