Improved Presence Detection for Occupancy Control in Multisensory Environments

Presence detection is used in occupancy control to dynamically adjust energy-related appliances in smart building applications. Yet, practical applications typically suffer from high sensor unreliability. We propose a computationally efficient approach, based on Hidden Markov Models, to fuse sensor observations from multiple sensors to better estimate user state (presence/absence). Our model considers a realistic scenario, where sensor communication may be limited or unreliable, thus some sensor observations data may be missing for some intervals. Compared to state of art classifiers (Logistic Regression, Naive Bayes, SVM), our approach achieves improved results while maintaining low computational and memory requirements or even relaxing these. Judging from our experiments, the algorithm appears to work well also in real-world test set-up where user presence and sensors error may not exactly follow our idealized model assumptions.