Unsupervised Adversarial Domain Adaptation for Micro-Doppler Based Human Activity Classification

Abstract

The fundamental difficulties in the supervised deep learning algorithm are obtaining large-scale labeled data and generalizing the trained model to a new environment. In this letter, we propose an unsupervised domain adaption method for human activity classification using micro-Doppler signatures. We study on how to classify micro-Doppler signatures in a new domain using only labeled samples from a different domain, mainly focus on simulation-to-real-world deep domain adaptation. First, we use motion capture (MOCAP) database to generate simulated micro-Doppler data to train the convolutional neural network (CNN). Then, considering the difference between simulation and real-world domain distributions, we introduce a domain discriminator to pit against the feature extractor part of the CNN. Through this adversarial process, like the generative adversarial network, the CNN trained on the simulation domain is able to generalize to the real-world domain. Experiment results show that the proposed method achieves over 84.02% accuracy in real-world micro-Doppler classification, which outperforms nearly 16% in CNN trained on the annotated simulation without domain adaptation and performs better than the existing domain adaptation methods.