ClassAct: accelerometer-based real-time activity classifier
Ramona Rednic, Elena I. Gaura, James Brusey · Coventry University Open Collections (Coventry university) · 2009
In enclosed bomb disposal suits, posture affects the air flow and is thus a key indicator for predicting the onset of Uncompensable Heat Stress (UHS). In order to allow the exploration of this effect, a system was developed to monitor the posture of human subjects during bomb disposal missions using only low cost accelerometers. Decision trees are used to identify in real-time, within the suit, eight mission-like postures: standing, kneeling, sitting, crawling, walking and lying on front, back, and one side. A variety of time domain features were explored to aid differentiation between the static and dynamic postures. An average classification accuracy of 97.2% over the nine postures are obtained when using the windowed variance and nine accelerometers. Similar performance was obtained with as little as two accelerometers, whilst a single hip accelerometer was shown to classify standing, walking and sitting with an average accuracy of 96.4%. Overall the instrument exhibits a suitable level of performance for the application at hand, in terms of wearability, accuracy, timeliness and data yield. The classification technique developed could be extended to the classification of other task oriented activities.