Attention-Augmented Convolutional Autoencoder for Radar-Based Human Activity Recognition

We propose an attention-augmented convolutional autoencoder for human activity recognition using radar micro-Doppler signatures. We use attention to overcome the limited receptive field of convolutional autoencoders (CAE), thereby enabling them to learn global information in addition to spatially localized features, while preserving their unsupervised pretraining characteristic. The augmentation is accomplished by concatenating convolutional local-feature maps with a set of attention feature maps that capture global dependencies. Using real data measurements of falls and activities of daily living, we demonstrate that the incorporation of the attention mechanism yields superior classification accuracy with respect to training sample size, compared to the conventional CAE.