Acoustic monitoring of the patterns of activity in the office and the garden

A. Harma, J. Skowronek and M. F. McKinney

Philips Research, Eindhoven, The Netherlands

Sound is one of the most useful modalities for monitoring activity of humans, animals, and machines in our environment. Sound propagates through light walls and diffracts around obstacles and thus is not as limited as visibility from the point of monitoring. In addition, a large number of activities and behavior produce recognizable audible events. In this paper, we describe an experiment based on the automatic classification and recognition of short isolated acoustic events in our natural environment. The system comprising of a computer and a microphone was originally designed for the purpose of long-term automatic bird monitoring. However, in the current article, the system was used in three quite different environments. The first location was a typical office room where all interesting acoustic events were recorded during a continuous monitoring session of more than two months. The second environment was a side street in a closed industrial area with several weeks of continuous data. Finally, the same system was used at the backyard of a typical Dutch town house over a period of 20 days. At all sites, the automatic system collected hundreds of thousands of sound events, and computed low-level parametric representations of those sounds. The parametric representations were then used to train a separate classifier for each site. Finally, each classifier was used to construct different types of logs and statistical representations reflecting various types of behavior at the three sites. For example, the office data reveals interesting details of the behavior of the people working at the site, and the backyard data draws a detailed picture of the daily activities of various bird species, and dogs, in the neighborhood. In the final article, we give a brief description of the basic methodology. Then we compare the performance of generic classification methods for the data at different sites and show some specific cases of regular patterns of activity. Finally, we try to characterize usability of the proposed method in a generic case of measuring patterns of activity in an arbitrary acoustic environment.

Paper presented at Measuring Behavior 2005 , 5^th International Conference on Methods and Techniques in Behavioral Research, 30 August - 2 September 2005, Wageningen, The Netherlands.

© 2005 Noldus Information Technology bv