Room-Level Occupancy Estimation via Wi-Fi Channel State Information on ESP32 Nodes: A Multi-Zone Experimental Study with Formal CSI–Occupancy Modeling

Room-level occupancy information is essential for demand-driven HVAC control and space utilization analytics, yet conventional solutions rely on PIR or camera systems that require dense instrumentation, line-of-sight, and frequent maintenance. Wi-Fi Channel State Information (CSI) offers a passive alternative: human presence perturbs the multipath structure of wireless channels in ways that can be detected without additional hardware. This paper develops a formal CSI–occupancy modeling framework and validates it experimentally using low-cost ESP32 nodes that expose CSI in the 2.4 GHz band. We model the complex CSI of each link as a superposition of environment-dependent and occupancy-dependent components and cast room-level occupancy estimation as a multi-class classification problem on CSI-derived features. A mathematically defined feature set based on per-subcarrier amplitude statistics, Doppler-band energy, and inter-subcarrier correlation is combined with three lightweight classifiers (logistic regression, gradient boosting, and a compact GRU network). A multi-zone experimental campaign in three residential rooms and one small office zone yields more than 250 hours of labeled data, with room occupancy (0, 1, ≥2 persons) obtained from door sensors and schedule logs. The gradient-boosted model achieves zone-averaged F1-scores of 0.96 for binary occupancy and 0.89 for three-class occupancy, with median inference latency of 38 ms on a Raspberry Pi edge node. Cross-zone transfer experiments and a simple calibration scheme are analyzed, and a theoretical upper bound on achievable accuracy under label noise is derived. The results demonstrate that commodity ESP32-based Wi-Fi infrastructure can support mathematically grounded, room-level occupancy sensing suitable for embedded smart building deployments.