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

Area of Science:

• Computer Science
• Electrical Engineering
• Building Management Systems

Background:

• Accurate indoor occupancy estimation is crucial for building management, energy optimization, and public health.
• Existing technologies often lack accuracy, affordability, or privacy preservation.

Purpose of the Study:

• To investigate the use of existing WiFi infrastructure as a non-intrusive sensing system for indoor occupancy estimation.
• To develop and validate machine learning models for WiFi-based occupancy sensing.

Main Methods:

• Utilized WiFi access points as soft sensors to collect anonymized connection metadata.
• Trained supervised machine learning regression models (SVR, Ridge, MLP, XGBoost) on WiFi data.
• Validated predictions against computer-vision ground truth in a university library over eight months.

Main Results:

• Best-performing models (SVR, Ridge, MLP) achieved R² ≈ 0.95, with mean absolute errors of ~8 persons and SMAPE below 10% at medium-to-high occupancies.
• XGBoost showed weaker generalization at extreme capacities due to data sparsity and hyperparameter sensitivity.
• No temporal degradation was observed over the 8-month study, indicating long-term stability.

Conclusions:

• WiFi-based occupancy estimation is a robust, cost-effective, and privacy-preserving solution.
• This method offers a viable alternative for real-world building management applications.
• The system demonstrates long-term stability and high accuracy in diverse occupancy scenarios.