Comparison of offline and real-time human activity recognition results using machine learning techniques

摘要

Today's human activity recognition is an important part of healthcare and ambient-assisted living where accelerometer and gyroscope sensors provide the raw data about physical activities and functional abilities of an observed person. Previous studies have shown that activity recognition can be seen as a machine learning chain with its particular data preprocessing technique. In recent past, several scientists measured rather high recognition accuracies on public databases or in laboratory environment but their solutions have not been tested in real environment. The goal of this paper is to examine the efficiency of previously used machine learning methods in real time by an Android-based, self-learning, activity recognition application which has been designed especially to this study according to the latest theoretical results (with the most relevant feature extraction and machine learning algorithms). Before real-time tests, we investigated the design considerations and application possibilities of different shallow and deep methods. The final outcome shows recognition rate difference between the "online" and "offline" cases. In the article we present some reasons for the difference and their possible solutions.