Are you an EPFL student looking for a semester project?
Work with us on data science and visualisation projects, and deploy your project as an app on top of Graph Search.
Introduction: A reliable measurement of the physical activity in everyday life should allow a better assessment of the utility and the relevance of a number of medical treatments. Continuous 24-h recordings of posture and motion can be generally useful in behaviour assessment. Categorising body postures (sitting, standing and lying) and locomotion (walking) based on only one miniature kinematic sensor has been studied [1]. The purpose of this study is to improve the accuracy of that system and to provide other parameters of movement such as vertical body displacement and absolute trunk tilt. Materials/ Methods: 11 elderly Community-dwelling subjects (79±6 years) carrying a kinematic sensor on the chest performed 6 tests involving various postural transitions and dynamic activities. The kinematic sensor was composed of one miniature gyroscope measuring medial-lateral trunk angular velocity and two miniature accelerometers measuring longitudinal and anterior-posterior trunk acceleration. Signals were recorded by a portable datalogger. Trunk tilt was estimated from the gyroscope signal. Changes in trunk tilt during postural transition (sit-stand or stand-sit) were enhanced using wavelet transform based on Mallat algorithm[2] and the precise time and period of each postural transition were detected. Based on a kinematic model of body motion in the sagittal plane, vertical and frontal body accelerations (relative to the room reference) were estimated from the measured accelerations and the trunk tilt. Trunk vertical displacement was then estimated from the double integral of vertical acceleration. The patterns of vertical displacement during each transition allowed us to recognize the transition between sit-stand and stand-sit. The results of posture classification were further enhanced by walking detection and analysing body trunk tilt. In fact, during the sitting, walking state is impossible; more over for elderly persons during the standing posture the state of leaning backward is quite improbable. In addition, the vertical acceleration signal, enhanced by wavelet transform, was used to detect walking and lying states. In order to validate our method a 3D optical motion analysis system (ViconTM, OXFORD METRICS) was used as reference to capture the actual chest movements and positions. In addition the accuracy of the system was evaluated for two elderly individuals by monitoring during one hour their physical activities, involving lying, sitting, standing, walking and climbing. The results were compared to those obtained from an observer reporting all periods of activity. Results: Sensitivity of posture transition detection was 99%. Sensitivity and specificity were 93% and 80% in sit-stand, 80% and 93% in stand-sit respectively. However, in the tests involving walking period these figures were increased to 100% and 94% for sit- stand and 96% and 97% for stand-sit posture transitions. Results from the monitoring of the two elderly persons showed that errors in posture classification for sitting and standing were 6% and 8%, and the errors for walking and lying duration estimation were 3% and 1% respectively. Conclusion: The most important and original aspect of this study is categorizing body postures (sitting, standing and lying) and locomotion (walking, walking upstairs and downstairs) using only one kinematic sensor, hence not hindering the subject in long term monitoring. This new approach introduces a new tool to estimate absolute trunk tilt and displacement. The results show also that wavelet transform is a powerful technique to detect postural transition and enhancing the pattern recognition of activities of interest. This method offers a promising tool for long term monitoring of daily physical activities and can also be used in motion analysis during sit-stand and stand-sit transitions [1] ] B. Najafi, K. Aminian, F. Loew, Y. Blanc, Ph. Robert, ' An Ambulatory System for Physical Activity Monitoring in Elderly', IEEE EMBS, p. 562-566, Oct. 2000. [2] S.Mallat, ‘’A theory for multi-Resolution signal decomposition (The wavelet representation)’’, IEEE Trans., Pattern Anal. & Machine Intelligence, vol. 11(7), pp. 674-693, 1989.
Michael Herzog, Leila Drissi Daoudi - Kleinbauer, Lukas Vogelsang
Auke Ijspeert, Grégoire Courtine, Joachim von Zitzewitz, Miroslav Caban, Hari Prasanth, Urs Keller