This is published under the terms of CC BY-NC.

Downloaded: 465 times

Abstract

Behavior recognition using millimeter wave (mmWave) signals has become a hot topic in recent years. However, existing works are mainly based on the premise that training samples and test samples have the same distribution, which leads to weak robustness of the network model to the environment during actual deployment. In this paper, we propose a domain adaptation framework for action recognition based on mmWave radar signals. Specifically, we use a convolutional neural network to construct our encoder to extract behavioral features in RF signals, use a semi-supervised learning method to pre-train the network, and finally we design a pseudo-label-based fine-grained domain adversarial network to further train the encoder. We conduct extensive experiments on our own collected behavioral data and two publicly available datasets. Experimental results demonstrate the superiority of our method.