Unsupervised Domain Adaptation for Human Activity Recognition in Radar

The difficulty of obtaining large-scale labeled radar data hinders the generalization of supervised deep learning algorithms, which makes the performance of these approaches drop when applied to a new domain. In this paper, we propose an unsupervised adversarial domain adaptation method with radar micro- Doppler spectrograms for human activity recognition. An improved adversarial loss, which has the same mechanism as Generative Adversarial Network, is employed for optimization. Besides, a multi-layer domain discriminator is presented in this method. A simulated dataset (source domain) with labeled radar spectrograms and a measured dataset (target domain) with unlabeled spectrograms are utilized for experiment. Experimental results demonstrate that the proposed method has good performance (with an average accuracy of 81.6%) on recognizing human activities without labeled measured spectrograms in the target domain. And comparison experiments with several state-of-the-art domain adaptation methods also indicate the efficiency of the proposed model.