On the market, there are more and more di®erent systems for monitoring
elderly people. Most of the systems are expensive and inaccessible since they
use very sophisticated monitoring equipment. Basic functions of monitoring
could be done by reasonably priced sensors such as accelerometers.
The object of this work was to develop the system for monitoring changes
in human gait. We used data from accelerometers attached to the left and right
ankle. Accelerations are represented in three-dimensional coordinate system.
We realized the best solution to our problem was to divide data set into ¯xed
time intervals. Without this assumption it would have been di±cult to deter-
mine the right attributes for machine learning.
To better understand the data in one interval, we divided walking into its
primitive elements - steps. By using some gait characteristics we were able
to conceive two algorithms. Together they enabled us to divide data into
single steps. Eight di®erent attributes distinctive for walking were chosen.
The next thing was an attempt to create universal training data set that
could recognize the di®erence between normal walking and limping. Testing
was done on di®erent time intervals. It proved successful to divide data into
longer time intervals. By doing so we got better information about gait in one
time interval. SVM proved to be the optimal machine learning model for the
presented classi¯cation problem.
|
