Toward High-level Activity Recognition from Accelerometers on Mobile Phones
Sozo Inoue, Yuichi Hattori,
Proc. IEEE International Conference on Cyber, Physical and Social Computing (CPSCom 2011)
In this paper, we propose an unsupervised method for multi-level segmentation, which could be used for a pre- process of non-sequential activity recognition, and could con- struct a high-level activity recognition using accelerometers on mobile phones. We extend single-level segmentation to multi- level by sweeping the temporal parameter. To confirm the valid- ity of our approach. we pursued the experiment of gathering accelerometer data of real nursing in a hospital. After the experiment and multi-level segmentation, we confirmed several phenomena to imply the validity of multi-level segmentation such that sequence seems to be properly segmented fitting to the annotations transcribed from the voice, that there are peaks of lower-level segment boundaries without higher-level boundaries, and that higher-level boundaries are not lower- level boundaries.