PIPE: Teaching WiFi Sensing to Ignore Position

Abstract

While WiFi signals are primarily intended for data transmission, their interaction with the environment causes subtle changes in the received signal. Receivers can measure these changes and use them to infer events occurring within the environment. For example, they can detect whether a subject is jumping or standing still. However, the observed signal depends strongly on geometry and the subject’s position. As a result, a model trained on one set of positions can fail entirely when the subject moves elsewhere in the same room.

Recent research has used contrastive learning to address this problem, with some success. Current solutions in this domain rely on distant proxies of the subject’s pose, such as activity labels, which may yield suboptimal results. Using a more direct proxy, such as a subject’s pose captured by a motion capture system, could improve generalization. In this paper, we introduce PIPE, a contrastive learning approach to WiFi sensing that defines positive pairs using motion-capture-derived pose similarity.

PIPE achieves performance comparable to SHARP [16] on human activity recognition (HAR), slightly outperforming it on unseen positions with an unseen subject. However, pose- and label-derived supervision for PIPE do not differ substantially in HAR performance. We attribute this result to the close alignment between activity labels and the objective of the task.