Classification of basic daily movements using a triaxial accelerometer.

A generic framework for the automated classification of human movements using an accelerometry monitoring system is introduced. The framework was structured around a binary decision tree in which movements were divided into classes and subclasses at different hierarchical levels. General distinctions between movements were applied in the top levels, and successively more detailed subclassifications were made in the lower levels of the tree. The structure was modular and flexible: parts of the tree could be reordered, pruned or extended, without the remainder of the tree being affected. This framework was used to develop a classifier to identify basic movements from the signals obtained from a single, waist-mounted triaxial accelerometer. The movements were first divided into activity and rest. The activities were classified as falls, walking, transition between postural orientations, or other movement. The postural orientations during rest were classified as sitting, standing or lying. In controlled laboratory studies in which 26 normal, healthy subjects carried out a set of basic movements, the sensitivity of every classification exceeded 87%, and the specificity exceeded 94%; the overall accuracy of the system, measured as the number of correct classifications across all levels of the hierarchy, was a sensitivity of 97.7% and a specificity of 98.7% over a data set of 1309 movements.

Detection of daily physical activities using a triaxial accelerometer.

Triaxial accelerometers have been employed to monitor human movements in a variety of circumstances. The study considered the use of data from a single waist-mounted triaxial accelerometer to distinguish between activity states and rest A method using acceleration magnitude was applied to data collected from 26 normal subjects performing sit-to-stand and stand-to-sit transitions and walking. The effects of three parameters were investigated: the length n of a smoothing median filter, the width w of the averaging window used to process the signal and the value of the acceleration magnitude threshold th. These were found to be inter-related, and sets of parameters that resulted in accurate discrimination were determined by the relationship between th and the product of w and n, and by the relationship between n and w. The subjects were randomly divided into control (N = 13) and test (N = 13) groups. Optimum parameter sets were determined using the control group. Eleven sets of parameters yielded the same optimum results of a sensitivity of 1.0 and a specificity of 0.96 in the control group. Upon application to the test group, using these parameters, the system successfully distinguished between activity and rest, giving sensitivities greater than 0.98 and specificities between 0.88 and 0.94.