Comparison of Deep Learning Techniques for RF-Based Human Posture Detection Systems

This paper focuses on techniques for a human posture classification framework that implements radio frequency (RF) active systems. In the first step, we describe the general approach considered for human posture classification. To this effect, we propose four different solutions: one based on traditional signal processing (SP) techniques, where the detection is centred around a correlation of prior classification masks; a second based on a hybrid SP and deep learning (DL) technique, where the DL model is trained with supervised data gathered at a single distance to the target; a third based on a hybrid SP and DL technique trained with data gathered at multiple distances to the target; and a fourth that uses variational auto-encoder (VAE) for feature generation. Their performance is then compared on the basis of classification accuracy and computation time. We show that although the SP-based solution presents high accuracy, the hybrid SP/DL solutions are advantageous in terms of classification accuracy and robustness at multiple distances, albeit requiring higher computation time. We further show the slight edge that VAE-based solutions have over plain DL solutions in terms of accuracy.