Deep Learning-Assisted Sub-6 GHz Environment Sensing via Reconfigurable Intelligent Surfaces

This paper introduces an approach to recognizing human posture using reflected wireless signals and Deep Learning (DL). The system combines Frequency-Modulated Continuous Wave (FMCW) signals with Reconfigurable Intelligent Surfaces (RIS) to reflect the signals throughout the space to improve signal diversity. It uses two software-defined radios (USRPs), one as a transmitter and one as a receiver. This setup helps the system pick up useful reflections regardless the person's position in the space, whether they're closer to the transmitter, the receiver, or somewhere in between. At the receiver's end, we process the reflected signals by performing Fast Fourier Transformation (FFT) analysis to generate Power Spectral Density (PSD) plots, which we convert into image data. These are then fed into a DL model combining a Convolutional Neural Network (CNN) module which extracts spatial features from each frame and a Long-Short Term Memory (LSTM) module that captures temporal patterns across multiple frames. This combination allows the system to make accurate posture predictions in real time.