Overview
In order to enable smart homes to offer their occupants the greatest possible added value, the installed sensors must be able to detect user activities in good time. For example, if a smart home detects that the last person is leaving the house, the user can be reminded to activate the alarm.
Currently, however, this is not the case because the number, type, placement, and quality of installed sensors vary widely, making it very difficult to solve the problem because separate training would be required for each smart home, which would require an enormous amount of effort.
Overcoming these challenges holds tremendous market potential for providing analytics and recommendations to smart home manufacturers, who in turn can sell premium services to end users.
In this project, we will explore how to combine weakly-supervised machine learning techniques (i.e., when activity tags are inaccurate) with other activity recognition methods, such as ontologies or rule-based approaches, to obtain a system capable of recognizing user activity from sparse and heterogeneous sensor configurations.
The ultimate goal is to provide activity recognition capabilities without having to perform specific training for each smart home. Or just a minimal configuration.
We will also develop a unique agent-based simulator that generates realistic, labeled smart home data that can be used both to benchmark the activity detection system and to generate synthetic training data.
Novelty and uniqueness: Uniquely combines different machine learning approaches to provide reliable behavioral detection of occupants in smart home environments that differ in size, number of occupants, or installed smart home components with minimal configuration and training effort.
Novelty and uniqueness: Uniquely combines different machine learning approaches to provide reliable behavioral detection of occupants in smart home environments that differ in size, number of occupants, or installed smart home components with minimal configuration and training effort.
