On time series sensor data segmentation for fall and activity classification

Abstract

The vast amount of literature on human ambulation and Activities of Daily Living (ADL) events classification has highlighted significant details on most aspects of the research area including: monitoring techniques, Wearable Sensor-based Monitoring Device (WSMD) placement on human body parts, and ambulation and ADL data collection methods, among others. However literature has failed to highlight meaningful details on one of the most important aspects of such studies, sensor data segmentation for feature extraction. The choice of segmentation techniques is in general very important, because inappropriate segmentation will most likely result in features without discriminant power. No classifier of whatever sophistication will give meaningful results with features that have no discriminant power. The optimal segmentation technique has been empirically investigated using sensor data from a bi-axial accelerometer. Results of the empirical investigation are presented.