Capsule Network With Multiscale Feature Fusion for Hidden Human Activity Classification

Abstract

This article considers the problem of human activity classification behind the walls using ultrawideband (UWB) radar. The complex-valued multiscale feature fusion capsule network (CV-MCNet) is proposed, which consists of a feature extractor, a multiscale feature fusion (MFF) block, and a capsule block. Specifically, the feature extractor with two complex-valued convolutional layers is designed to extract the deep features from the range profiles. Then, the MFF block is developed to enrich the feature representation of the activity. Finally, a capsule block is applied to implicitly encode the spatial relationship among the features in vector form and aggregate the vectors to get accurate classification results. The proposed CV-MCNet is evaluated by real data, and the results show that it achieves better classification performance compared with the deep convolutional neural network (DCNN), convolutional autoencoder (CAE), and complex-valued convolutional neural network (CV-CNN).