The prevalence of radio frequency signals in indoor environments has in recent years given rise to new technologies across many domains such as robotics, healthcare, and surveillance. Radio frequency signals propagate in the wireless medium through multiple paths and carry useful environment-dependent information. Capturing and analyzing these signal patterns can offer new solutions for a number of applications relevant to ranging, tracking, perception and recognition. In this work we propose a novel architecture, separating physical, back-bone networks, and inference layers, towards fully ubiquitous passive recognition systems that scale with the number of environments and applications. We propose a back-bone architecture that utilizes a novel Cross Dual-Path Attention (CDPA) block to capture spatial and temporal correlations from Channel State Information (CSI) for device-free, multi-task applications. Subsequently, a distill and transfer algorithm is proposed to generalize the inference capabilities of CDPA over multiple target environments for scalable training and reduced computational costs. By sharing knowledge between models across a shared network, experimentation shows that edge devices can be deployed with improved performance while simultaneously meeting strict computation and memory requirements. Our distributed learning paradigm demonstrates that CDPA-based models are capable of using passive signals in a non-intrusive and privacy-protecting manner, in order to achieve ubiquitous recognition at scale in smart environments.
Shervin Mehryar (Tue,) studied this question.