AASH – Analysis of Ambient Sound in Home Environments
Sounds in the home capture the context of activities of daily living, as well as other events taking place within and around the home. This research project will explore the possibilities of analysing sounds in a home environment for home-based healthcare monitoring and assistance systems.
This project concerns the techniques and methods for the automatic detection of acoustic events and classification of acoustic scenes in home settings. An acoustic event, for example a footstep, is an event created by single sound source. Multiple acoustic events create acoustic scenes representing people, activity or location. The objective of acoustic event detection is to process an acoustic signal to label temporal regions within that signal, that is to find the start time and end time for a specific event type. This process typically requires a lot of labeled data, which in turn is typically manually annotated. In the context of smart homes, which is characterised by an inherent evolving diversity (that is the home layout itself, its occupants and their routines differ and change over time), manual data annotation is impractical and methods that can generalise across different homes and persons are difficult to accomplish.
As an approach to better understand and address these issues, preliminary studies and data collection will be done in Halmstad Intelligent Home (HINT). Data collection in actual in-use apartments (hired to tenants by Halmstads Fastighets AB, HFAB) are also planned. Besides microphones, and to enhance privacy, thermal cameras will be used for data annotation purposes. Collected data will be used to explore and build signal processing and machine learning models. In the home context, unsupervised or semi-supervised methods seem appropriate for the detection and classification of acoustic events and acoustic scenes.