CornerSense: Virtual Path-Augmented WiFi Sensing for Human Proximity Detection Around Corners on Mobile Robots

WiFi sensing has attracted significant attention over the past decade for its potential to enable ubiquitous human activity monitoring. While prior approaches primarily rely on stationary transceivers, this paper is the first to investigate the feasibility of WiFi sensing on mobile robots. We introduce CornerSense, a new framework that empowers mobile robots to leverage their WiFi interfaces for detecting human proximity around corners, effectively addressing the non-line-of-sight challenges commonly encountered by traditional robotic sensors such as cameras and LiDAR. Our analysis demonstrates that the design of CornerSense fundamentally differs from, and is more challenging than, conventional WiFi sensing with stationary transceivers. In particular, the intertwined movement of the robot and nearby humans complicates the isolation of human-induced signal variations, which is critical for accurate proximity detection. To address this challenge, CornerSense develops a virtual path-augmented, two-stage dominant path extraction approach based on principal component analysis (PCA). In the first stage, through strategically introducing a virtual path, CornerSense extracts a reference path that incorporates only the robot's motion by applying PCA on the power of the virtual path-augmented channel state information (aug-CSI). During the second stage, the reference path is first subtracted from the aug-CSI. This subtraction allows for a second application of PCA to the power of the residual aug-CSI, thereby enabling the extraction of another distinct dominant path that is reflected off the human body before arriving at the robot. This path is referred to as the dominant human-reflected path. Finally, the reference path extracted in the first stage is employed to compensate for the hardware-induced phase offset in the dominant human-reflected path, yielding cleaned CSI ready for accurate and robust detection of human proximity. Real-world experimental evaluations conducted across nine different corners in three groups and under four distinct human walking patterns reveal that CornerSense achieves an average true positive rate (TPR) of 96% while maintaining a low average false positive rate (FPR) of 3%. In contrast, a baseline system that directly applies an algorithm intended for stationary transceiver-based proximity sensing only reaches a TPR of 84% and suffers from a markedly higher FPR of 46%, which is over an order of magnitude higher than that of CornerSense.