Vision-based occupancy detection in indoor environments: A comparison of standard RGB and thermal cameras

Buildings significantly contribute to emissions, primarily through energy consumption. Reducing these emissions requires optimizing building systems while accounting for occupancy levels, as occupant presence and behaviour directly impact heating, cooling, lighting, and ventilation demands. Accurate occupant detection is crucial for managing these factors, with vision-based methods emerging as a promising solution. Unlike traditional approaches like CO 2 sensors and motion detectors, vision-based detection provides richer contextual data, capturing occupant count, distribution, locations, activities, and postures. However, these methods often rely on standard (RGB) cameras, which are susceptible to visual distractions such as portraits or photographs and raise privacy concerns, limiting their broader adoption in buildings. This study explores the feasibility of using low-cost thermal cameras for occupancy detection, comparing their performance with standard cameras using deep learning-based object detection models, specifically YOLOv8 and v10. Field experiments were conducted in offices, meeting rooms, and classrooms, covering scenarios with varying occupant densities, environmental complexities, and potential sources of error, such as screens, posters and heated objects. The models were trained and tested under different experimental setups to assess their ability to generalize across locations and conditions. Results indicate that standard cameras generally outperform thermal cameras in controlled environments due to their higher resolution and visual detail. However, thermal cameras performed competitively with sufficient dataset, achieving up to 88% accuracy in real-world validation tests. Challenges such as overlapping occupants, heat signatures from non-occupant sources (monitors and vacated chairs), and dataset biases were analyzed to improve model robustness. The study also found that incorporating a diverse dataset improved detection performance, particularly in complex, crowded scenarios where occupants overlapped. These findings highlight the potential of thermal cameras as a viable alternative to standard cameras for occupancy detection, particularly in privacy-sensitive applications. • Thermal cameras preserved privacy but required tailored datasets to generalize. • YOLOv10 outperformed YOLOv8 in some cases but struggled with heat-emitting objects. • Overlapping cut accuracy by ∼12%, but dataset improvements reduced misdetections. • Investigates the impact of dataset complexity on model generalization performance. • Pretraining bias favored RGB; thermal models required domain-specific tuning.