WiFi Sensing on the Edge: Signal Processing Techniques and Challenges for Real-World Systems

In this work, we evaluate the feasibility of deploying ubiquitous WiFi sensing systems at the edge and consider the applicability of existing techniques on constrained edge devices and what challenges still exist for deploying WiFi sensing devices outside of laboratory environments. Through an extensive survey of existing literature in the area of WiFi sensing, we discover common signal processing techniques and evaluate the applicability of these techniques for online edge systems. Based on these techniques, we develop a topology of components required for a low-cost WiFi sensing system and develop a low-cost WiFi sensing system using ESP32 IoT microcontroller edge devices. We perform numerous real world WiFi sensing experiments to thoroughly evaluate machine learning prediction accuracy by performing Tree-structured Parzen Estimator (TPE) hyperparameter optimization to independently identify optimal hyperparameters for each method. Additionally, we evaluate our system directly on-board the ESP32 with respect to computation time per method and overall sample throughput rate. Through this evaluation, we demonstrate how an edge WiFi sensing system enables online machine learning through the use of on-device inference and thus can be used for ubiquitous WiFi sensing system deployments.