Occupancy Prediction Using Low-Cost and Low-Resolution Heat Sensors for Smart Offices.

Area of Science:

• Smart Buildings and Sustainable Architecture
• Sensor Technology and Data Analysis
• Human-Computer Interaction

Background:

• Occupancy prediction is vital for optimizing office energy use, lighting, and HVAC systems.
• Low-cost, low-resolution thermal sensors offer a privacy-preserving solution for occupancy monitoring.
• Existing methods struggle with noise artifacts from thermal sensor data.

Purpose of the Study:

• To develop and compare novel workflows for accurate occupancy prediction using low-resolution thermal sensors.
• To address and compensate for noise artifacts affecting thermal sensor data.
• To analyze the influence of algorithm parameters and image properties on prediction performance.

Main Methods:

• Utilized a low-resolution (8x8) non-intrusive thermal sensor in a meeting room.
• Proposed two distinct workflows: one based on computer vision and another on machine learning (feature classification).
• Employed state-of-the-art explainability methods for detailed analysis of algorithms and image properties.

Main Results:

• The feature classification method achieves high accuracy with noise-free data.
• The computer vision method demonstrates robustness by compensating for noise artifacts.
• The choice between methods depends on data quality (noise presence) and availability of empty room recordings.

Conclusions:

• Novel computer vision and feature classification workflows effectively predict occupancy using low-resolution thermal sensors.
• Understanding and mitigating noise artifacts is crucial for reliable thermal sensor-based occupancy prediction.
• These methods have broad applications beyond smart offices, including elderly care.