Monitoring of indoors human activities using mobile phone audio recordings

In this paper we present a methodology for monitoring of human activities in home using audio recordings captured from mobile phone. Specifically, after estimating a large set of audio features, unsupervised clustering is performed in order to extract feature subspaces. Human activity sound models were trained using different combinations of these subspaces. The best performance 92.46% was achieved using a neural network classifier.

I. Introduction

Over the last decade the achievements in areas such as artificial intelligence, social robotics, sensors, statistical signal modeling and machine learning have resulted to the development of autonomous monitoring systems. Monitoring systems are mainly used in applications which promote private/social security, i.e. surveillance systems, or health and well-being. Monitoring systems can roughly be divided into outdoors and indoors, based on the area of application.