Transfer Learning from Audio Deep Learning Models for Micro-Doppler Activity Recognition

Abstract

This paper presents a mechanism to transform radio frequency micro-Doppler signatures into a pseudo-audio representation, which results in significant improvements in transfer learning from a deep learning model trained on audio. We also demonstrate that transfer learning from a deep learning model trained on audio is more effective than transfer learning from a model trained on images, which suggests machine learning methods used to analyse audio can be leveraged for micro-Doppler. Finally, we utilise an occlusion method to gain an insight into how the deep learning model interprets the micro-Doppler signatures and the subsequent pseudo-audio representations.