Semi-supervised attention-augmented convolutional autoencoder for radar-based human activity recognition

In this paper, we consider semi-supervised training of an attention-augmented convolutional autoencoder (AACAE) for human activity recognition using radar micro-Doppler signatures. The AA-CAE learns global information in addition to spatially localized features, thus enabling the classifier to overcome the limited receptive field of a conventional convolutional autoencoder (CAE). The design also permits the possibility of semi-supervised training of the AA-CAE using training data comprising unlabeled and labeled sets. More specifically, the semisupervised training regime is implemented by first pre-training the AA-CAE via unsupervised training of the attention-augmented autoencoder with the unlabeled portion of the training data. This is followed by fine-tuning of the AA-CAE for classification using the labeled portion. Using real-data measurements of six different human activities, we demonstrate that the semi-supervised AA-CAE yields higher classification accuracy with much less labeled data than a fully-supervised conventional CAE.