Abstract:Passive radio frequency (RF) sensing and monitoring of human daily activities in elderly care homes has recently become an emerging topic due to the demand with ageing population. Micro-Doppler radars are an appealing solution considering their non-intrusiveness, deep penetration, and high-distance range. This study presents an unsupervised framework for human activity monitoring using Doppler streams. Two unsupervised feature extraction strategies based on convolutional filtering and texture analysis of Doppler images are considered. For the former, encoded features using Convolutional Variational Autoencoder (CVAE) are compared with Convolutional Autoencoder (CAE) features. For the latter, Grey-Level Co-occurrence Matrix (GLCM) is used. These methods are further compared with unsupervised linear feature extraction based on Principal Component Analysis (PCA) and Singular Value Decomposition (SVD). Using these features, unsupervised samples clustering is performed using K-Means and K-Medoids. Actual labels are solely used for evaluation and visualisation. The results showcase 82.5% and 84% average testing accuracies for CVAE features and 77.5% and 72.5% average testing accuracy using texture features based on GLCM using K-Means and K-Medoids respectively. The results show superiority of CVAE and GLCM features compared to PCA, SVD, and CAE with more than 20% average accuracy. Furthermore, for high-dimensional data visualisation, three manifold learning techniques are considered including t-Distributed Stochastic Neighbour Embedding (t-SNE), Multidimensional Scaling (MDS), and Locally Linear Embedding (LLE). The visualisation methods are compared for projection of raw data as well as the encoded features using CVAE. All three methods show an improved visualisation ability when applied on the transformed CVAE data.