Activity (ADL) Recognition in Smart-Home Environments

The content of this page is provided by Thomas Hauptvogel. All rights reserved.

Abstract

We live in an constantly aging society with a rising need for careworkers, which cannot be satisfied anymore. Assistance systems are a solution for this problem, especially robots, which can support careworkers with their work. Though robots are limited in their knowledge and ability to act by their hardware, which makes modular extensions via smart home technology a possible solution for this problem. A lot of people own sensors privately and they are relatively cheap and easy to get. In this work a system has been designed, which can detect activities of residents in their living environments by using sensors available on the market. Other assistance technology can communicate with the system and expand their knowledge about the household and its resident. A classification of recognized sensor activities is done by machine learning as well as by pre-defined rules, which classify the activities by the location of the sensors. With the help of a grafical user interface everyone in the local network can access the system and see their daily routine and insert activities by their own, which are used by the system to train the machine learning model. Before designing the concept the needs of the user groups were analyzed and written into requirements. An useable prototype has been devel- oped by using this concept. The prototype was tested by users from different groups, which could handle nearly all functionalities of the system, thought it was very transparent and had some ideas for extensions. Mainly the older users, who are also the target audience, could imagine using the system in their homes. At the end of this work there was an usable prototype which could classifiy activities of daily living with good accuracy and had a lot of potencial for more extensions.

Download Thesis

Language Notice
This document is available in German language only.

You can download the full submission as PDF file (2,2 MB) here.