Semi-Supervised Convolutional Autoencoder With Attention Mechanism for Activity Recognition

In this paper, we consider human activity recognition with a semi-supervised convolutional autoencoder (CAE), augmented by an attention mechanism, using radar micro-Doppler signatures. The attention-augmented CAE (AA-CAE) learns both global information and spatially localized features, thus enabling the classifier to overcome the limited receptive field of a CAE. Considering training data comprising both labeled and unlabeled samples, a semi-supervised training regime is implemented through a joint loss function, with training of the encoder part performed in an unsupervised fashion using all training samples and the classifier and attention mechanism trained at the same time using only the labeled samples. Using real-data measurements of six different human activities, we demonstrate that the jointly trained AA-CAE classifier yields higher classification accuracy with fewer labeled data than the semi-supervised AA-CAE trained via a conventional two-step process.