Through-the-wall Human Activity Recognition using Ultra-Wideband Radar and Deep Learning

This paper presents a novel system that leverages ultra-wideband (UWB) radar technology in conjunction with a deep neural network for through-the-wall human activity recognition (HAR). The human presence in a room in terms of four distinct classes (empty, standing, moving towards the wall, and moving away from the wall) is evaluated by observing the variations in the received signal from the UWB SLMX4 radar in the frequency range of 7.25-10.2 GHz. A dataset comprising 1200 images is produced through a multi-step preprocessing method, culminating in scalograms generated via the continuous wavelet transform (CWT). The scalogram encapsulates spatiotemporal features that allow for the differentiation of classes through a deep conventional neural network (CNN). The optimized 2DCNN model yields training and testing accuracies of 98.89% and 99.59%, respectively. With satisfactory outcomes, a graphical user interface (GUI) is devised with an intuitive structure that allows the visualization of real-time classification results.