Background-Calibrated Wi-Fi Sensing for Semantic Space-State Separation

Abstract

Wi-Fi channel state information (CSI) is promising for device-free sensing, but deployment is difficult when task-specific labeled calibration and high-dimensional processing are required. This letter presents a background-calibrated, label-free framework requiring only a short empty-room CSI interval for initialization. A dynamic cue from short-term variation and a static cue from deviation from the empty-space profile are jointly interpreted: low/low, low/high, high/high, and high/low cue combinations correspond to empty, static occupied, active occupied, and inconsistent states, respectively. Experiments with ESP32-based 2.4-GHz, 20-MHz Wi-Fi nodes in two meeting rooms achieved macro F1-scores of 0.933 and 0.943. A motion-only ablation cannot represent static occupancy (F1 = 0 in both rooms), whereas the proposed formulation achieved static-state F1-scores of 0.894 and 0.903. In addition, the semantic state was preserved down to latent dimension 8 (7.75× compression) with fixed pretrained autoencoders, while performance degraded sharply at dimension 4 and below. These results support interpretable three-state sensing using background-only calibration without task-state labels.