Enhancing room occupancy inference from sparse PIR sensor data using Transformer models

Abstract: The residential energy sector plays a vital role in global environmental impact. In this domain, Internet of Things (IoT) systems are widely used for energy conservation and management. A key aspect of improving energy efficiency is establishing room occupancy ground truth for further analysis and modeling. Passive infrared (PIR) sensors are favored methods to detect occupancy for their simplicity and non-invasiveness. However, they often exhibit limited accuracy. The lack of reliable occupancy ground truth complicates subsequent work in optimizing building management.

Hence, this research aims to develop a reliable method for mapping PIR sensor data to occupancy status. In this way, we addressed a key challenge by building a non-invasive occupancy detection solution. To this end, our project collected a data set in 10 houses over two months. We also used a public dataset in this paper for comparison and transfer learning. By developing models that combine PIR sensor data with environmental factors such as humidity, CO2 levels, and temperature, this research demonstrates the feasibility of accurately inferring room occupancy without direct observational data. Therefore, this paper contributes by developing methods to map PIR sensor data to occupancy reliably. Based on the modeling approach taken, transformer methods achieve better metrics than other methods and demonstrate transferability. Our code and dataset are available at GitHub Repository.

Recommended citation: Sheng, Y., Özbakır, A.D., Maathuis, C., Iren, D., & Bromuri, S. (2025). "Enhancing room occupancy inference from sparse PIR sensor data using Transformer models." In Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing (SAC '25), 1570-1579. https://doi.org/10.1145/3672608.3707755
Download Paper