Human identification via unsupervised feature learning from UWB radar data

This paper presents an automated approach to automatically distinguish the identity of multiple residents in smart homes. Without using any intrusive video surveillance devices or wearable tags, we achieve the goal of human identification through properly processing and analyzing the received signals from the ultra-wideband (UWB) radar installed in indoor environments. Because the UWB signals are very noisy and unstable, we employ unsupervised feature learning techniques to automatically learn local, discriminative features that can incorporate intra-class variations of the same identity, and yet reflect differences in distinguishing different human identities. The learned features are then used to train an SVM classifier and recognize the identity of residents. We validate our proposed solution via extensive experiments using real data collected in real-life situations. Our findings show that feature learning based on K-means clustering, coupled with whitening and pooling, achieves the highest accuracy, when only limited training data is available. This shows that the proposed feature learning and classification framework combined with the UWB radar technology provides an effective solution to human identification in multi-residential smart homes.