Occupancy detection, Machine learning, Indoor environment, Temporal cross-validation, Cross-room transfer, Smart buildings
BIMA JOURNAL OF SCIENCE AND TECHNOLOGY GOMBE · 30.07.2026 · 10 (3) · 201-212
DOI: 10.64290/bima.v10i3.1680 ↗Kurz gefasst
Fachartikel in der Zeitschrift BIMA JOURNAL OF SCIENCE AND TECHNOLOGY GOMBE, online erschienen am 30.07.2026. Verfasst von Ahmed Mohammed und 4 weiteren Autor:innen.
Abstract (Originalsprache)
Occupancy-driven control of lighting and air-conditioning can cut building energy use, but the passive infrared (PIR) sensors normally used miss stationary occupants, and camera-based alternatives raise privacy concerns. Ambient environmental sensors offer a privacy-preserving complement, yet published accuracies are often obtained with randomly shuffled cross-validation and sensor-derived labels, which may overstate performance. This study evaluated device-free occupancy detection from temperature, humidity, illuminance and time features in five naturally ventilated offices in Gombe, Nigeria (56,049 one-minute samples over 8.28 days), and tested how validation design, label choice and cross-room transfer affect the reported accuracy. Six classifiers were compared under shuffled five-fold and leave-one-day-out validation, against a firmware-derived and an independent PIR-derived label, and transferred between offices. Under shuffled validation a random forest reached an F1-score of 0.90 and an area under the ROC curve (AUC) of 0.95. Under leave-one-day-out validation the best model, gradient boosting, reached only F1 = 0.76, AUC = 0.71 and Matthews correlation coefficient = 0.33, below the F1 of an always-occupied baseline (0.78); against the independent PIR label AUC fell to 0.55–0.64. Leave-one-office-out transfer gave a mean AUC of 0.71, rising only to 0.73 with per-office standardisation and 24 h of target-office data. Without CO2 sensing, these ambient variables therefore carry a weak occupancy signal that does not generalise reliably to unseen days; the high shuffled-validation accuracy reflects temporal leakage and label circularity. Device-free occupancy studies should report day-blocked validation with independent labels, and practical deployments should add CO2 sensing or fuse ambient data with PIR motion.
Zitieren
APA
Mohammed, A., Ismail, M., Nafada, A. I., Dawaki, M., & Ajuji, M. (2026). Occupancy detection, Machine learning, Indoor environment, Temporal cross-validation, Cross-room transfer, Smart buildings. BIMA JOURNAL OF SCIENCE AND TECHNOLOGY GOMBE, 10(3), 201-212. https://doi.org/10.64290/bima.v10i3.1680
Vancouver
Mohammed A, Ismail M, Nafada AI, Dawaki M, Ajuji M. Occupancy detection, Machine learning, Indoor environment, Temporal cross-validation, Cross-room transfer, Smart buildings. BIMA JOURNAL OF SCIENCE AND TECHNOLOGY GOMBE. 2026;10(3):201-212. doi:10.64290/bima.v10i3.1680
BibTeX
@article{Mohammed2026,
author = {Mohammed, Ahmed and Ismail, Mustapha and Nafada, Abdulhamid Ibrahim and Dawaki, Muhammad and Ajuji, Mohammed},
title = {Occupancy detection, Machine learning, Indoor environment, Temporal cross-validation, Cross-room transfer, Smart buildings},
journal = {BIMA JOURNAL OF SCIENCE AND TECHNOLOGY GOMBE},
year = {2026},
volume = {10},
number = {3},
pages = {201--212},
doi = {10.64290/bima.v10i3.1680}
}
EUVIMED