A Machine Learning Approach for Estimating Person Counts Using Anonymous WiFi Data in a University Library

Abstract

Accurately estimating indoor occupancy is essential for managing building spaces and infrastructure, with applications ranging from ensuring safe distancing and adequate ventilation during health crises to optimizing energy consumption and resource allocation. However, no existing technology simultaneously achieves accuracy, low-cost, and privacy preservation in indoor occupancy measurement. This study investigates the use of existing WiFi infrastructure as a non-intrusive sensing system, where access points operate as soft sensors that passively collect anonymized connection metadata serving as proxies for human presence. The proposed approach was validated in a university library over eight months, training supervised machine learning regression models on WiFi data and comparing predictions against computer-vision ground truth. The best-performing models (SVR, Ridge, and MLP) consistently achieved R2 ≈ 0.95, with mean absolute errors of about 8 persons and relative errors (SMAPE) below 10% at medium-to-high occupancies. Tree-based ensemble models, particularly XGBoost, exhibited weaker generalization at extreme capacity ranges, likely due to data sparsity and sensitivity to hyperparameters. Importantly, no temporal degradation was observed across the 8-month horizon, confirming the long-term stability of the method. Overall, the results demonstrate that WiFi-based occupancy estimation offers a robust, cost-effective, and privacy-preserving solution for real-world deployments.