Robust and Practical WiFi Human Sensing Using On-device Learning with a Domain Adaptive Model

Abstract

The ubiquity of WiFi devices combined with the ability to cover large areas, pass through walls, and detect subtle motions makes WiFi signals an ideal medium for sensing occupancy. While extremely promising, existing WiFi sensing solutions have not been rigorously tested outside of lab environments and don't often consider real-world constraints associated with non-expert installers, cost-effective platforms and long-term changes in the environment. This paper presents M-WiFi, a user-in-the-loop self-tuning framework for WiFi-based human presence detection with on-device learning and domain adaption capabilities that operates entirely on an embedded platform. M-WiFi robustly detects human presence by separating human-specific disturbances on WiFi signals from those of static objects, moving furniture or even pets. The high-level features of human presence are captured in an initial generalized classification model which adapts over time to a new building by selectively asking users to annotate a small number of critical time periods. We evaluate M-WiFi in 7 different houses, for a total of 100 days, with a mixture of pets and including periods of sleep and stationary activities. We show that our domain adaptive model can detect the human presence with an average accuracy of 90% in a completely new house after only 3 days of self-tuning and rapidly reaches a steady-state performance of 98% in long-term operations.