Cross Frequency Adaptation for Radar-Based Human Activity Recognition Using Few-Shot Learning

Human activity recognition (HAR) using radar has been realized commonly using deep neural networks (DNNs). A change in radar operating frequency significantly changes the spectrogram characteristics compared to any other radar parameter. In this work, we consider three different approaches, viz., transfer learning, metric learning, and meta-learning (Reptile algorithm) for adapting the network developed for one radar operating frequency (source domain) to another operating frequency (target domain). Results and discussions on the performance of these algorithms on an openly available dataset are presented.