Leveraging Language Models for Human Activity Recognition for Intelligent Lighting

Recognizing human activities is necessary to achieve fully context-aware lighting in a smart home. Ambient sensors suit this task as they are ubiquitous, non-intrusive, and nonobstructive. However, each home has a unique floor plan and a unique number and placement of these sensors. This leads to heterogeneity of input data spaces for machine learning models, which hinders their use in different homes. We propose that converting the data into the semantic input space of a language model can not only unify the input space across different houses but also improve the prediction performance thanks to the model’s understanding of semantics. We compared a pre-trained BERT model with one trained from scratch to test the proposed semantic conversion on three datasets from CASAS. The results show that the pre-trained model performs better on accuracy and F1-score.