Tracking daily routines of elderly users through acoustic sensing: An unsupervised learning approach

Assistive technologies that can passively track people’s daily activities with dementia can deliver significant benefits for the patients themselves and their carers. This work investigates the feasibility of developing a system for the unsupervised tracking of daily activities at home through acoustic sensing. Motivated by the wide adoption of intelligent voice assistant devices in home environments, we developed a prototype algorithm to identify diversions from typical activities using the captured sounds, without the need for activity labeling. The system relies on sound embeddings through a pre-trained model, a novel dimensionality reduction algorithm, and the application of dynamic time warping for pattern matching. Our evaluation through synthetic activity sequences using data from our data collection in addition to public datasets shows very good performance (precision 0.99, recall 0.95).