Visual Override for Occupancy Sensing with Indoor CO2 Concentration

Document Type

Conference Proceeding

Publication Date

1-1-2026

Abstract

Reliable occupancy information is crucial for improving HVAC efficiency, maintaining indoor air quality, and reducing energy consumption in smart buildings. However, unpredictable occupancy patterns create uncertainty in control algorithms, requiring robust detection methods. This study proposes a vision-based approach that integrates a physics-driven CO2 estimation model with a deep learning computer vision framework. A custom-trained YOLO model was employed to determine occupant count and location, attaining near-ideal accuracy with a mean Average Precision (mAP@0.5) of 0.995, a precision of 1.0 at a confidence threshold of 0.639, the recall remains 1.0 at a threshold of 0.60, and an F1 score of 1.0 at a threshold of 0.56. The predicted CO2 levels closely matched sensor measurements, with most differences within ±50 ppm. An override logic further enhanced robustness by resolving discrepancies during transitional occupancy periods. By combining vision-based and physics-based estimates via a confidence weighting factor, the framework demonstrated improved reliability and robustness for real-time DCV by reconciling sensing and model imperfections and is readily deployable with existing cameras and CO2 sensors.

Publication Source (Journal or Book title)

ASHRAE Transactions

First Page

248

Last Page

257

This document is currently not available here.

Share

COinS