Revolutionizing Activity Recognition Through Walls with Deep Learning

This study introduces a novel approach to human activity recognition through walls, utilizing a non-invasive sensing system powered by advanced deep learning (DL) algorithms. Conventional activity detection techniques typically depend on cameras, which pose privacy challenges, are influenced by lighting conditions, and demand extensive training with long video sequences. Addressing these limitations, this research presents a privacy-focused activity recognition system that integrates cutting-edge UWB radar technology with advanced DL methodologies. Experiments were conducted involving a single subject performing various activities to assess the system's effectiveness. The system specifically classifies three core scenarios: standing, sitting, and an empty environment. By converting the acquired data into spectrograms and leveraging sophisticated DL models such as MobileNet, ResNetSO, VGGI6, and VGG19, the proposed method delivers highly accurate activity recognition, achieving an impressive peak accuracy of 100.0% across all models.