Multi-Person Activity Recognition in Continuously Monitored Smart Homes.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Designing activity recognizers for continuous, in-home settings is challenging, especially with multiple residents.
• Existing methods struggle to estimate resident numbers, split data streams, and identify individual activities.

Purpose of the Study:

• To develop a novel unsupervised method for activity recognition in multi-resident homes.
• To accurately estimate the number of residents and associate sensor data with each individual.
• To enable robust tracking of behavioral routines in complex domestic environments.

Main Methods:

• Introduction of Gamut, an unsupervised method utilizing a multi-target Gaussian mixture probability hypothesis density filter.
• Jointly estimating the number of residents and associating sensor readings with identified residents.
• Processing time-series sensor data streams to create separate substreams for each resident.

Main Results:

• Gamut demonstrates superior performance compared to previous unsupervised and supervised methods.
• The algorithm shows robustness in multi-resident home settings.
• Experiments were conducted using labeled data from 50 single-resident and 11 multi-resident homes.

Conclusions:

• Gamut offers a robust strategy for activity recognition in homes with two or more residents.
• The proposed method addresses key challenges in multi-resident activity monitoring.
• This work advances the field of activity recognition in complex, real-world environments.