Frugal learning is a topic that is crucial for mobile and wearable devices. Hardware that is presented in these devices has lower computational power compared to desktop computers and servers. Therefore, not every algorithm can be effectively used on a mobile phone or a smartwatch. The main goal of this project is identification of solutions that can show best results on wearable devices for Human Activity Recognition (HAR) problem. Analysis will be conducted on real data collected from open repositiories and partners.
When the set of algorithms appropriate for HAR is selected, the next step is to develop an application for a wearable device. This application should be able to recognize activity in real time with a standalone mode while exchanging very limited amount of data with a cloud. This practical part can reveal not only the CPU usage and performance in activity recognition for each selected algorithm, but how much memory and battery each of these algorithms consumes in a real world scenario.