Feature extraction from sensor data streams for real-time human behaviour recognition

In this paper we illustrate the potential of motion behaviour analysis in assessing the wellbeing of unsupervised, vulnerable individuals. By learning the routine motion behaviour of the subject (i.e. places visited, routes taken between places) we show it is possible to detect unusual behaviours while they are happening. This requires the processing of continuous sensor data streams, and real-time recognition of the subject's behaviour. To address privacy concerns, analysis will be performed locally to the subject on a small computing device. Current data mining techniques were not developed for restricted computing environments, nor for the demands of real-time behaviour recognition. In this paper we present a novel, online technique for discretizing a sensor data stream that supports both unsupervised learning of human motion behaviours and real-time recognition. We performed experiments using GPS data and compared the results of Dynamic Time Warping.