Biomedical Engineering Reference
In-Depth Information
where
J
is a true or false value based on if a jump has occurred,
H
y
is the height of
the head position,
H
y
is the calibrated normal height of the head joint and
C
is some
constant.
C
would then be set to a height that a person would only get to by jumping
from the ground. Such recognition is very specialized but simple, explainable and
can determine in an instant whether a jump has occurred.
Selecting between algorithms for your recognition will be highly dependent on
your devices and the needs of your application. When performed properly, these
choices can result in recognition with high accuracy [
6
,
27
,
28
]. However, errors and
uncertainty are always going to remain, even with your best efforts. No translation
from raw data to real world actions is going to be perfect but you can minimize the
impact of these issues, as discussed in the next sub-section.
16.3.3 Modifying the Models to Address Error and Uncertainty
With any device or solution to a difficult problem, there are chances for error and
uncertainty. From the data itself, to the results from the algorithms, or even the user
themselves, there are several possibilities and variables to consider. While all error
and uncertainty is not likely to be completely removed, by properly addressing them,
they can be minimized. In this section we start with issues that arise at the device
level and present solutions following the process all the way to the end user.
Errors from the Device
. Most raw sensor data, such as the accelerometers and
gyroscopes, have the potential to fluctuate slightly as a result of their mechanical
design. This can be observed often while the device is just sitting steady on the table
(in fact this is a good technique to measure the error ranges), and in some cases it
is exasperated as the device is moved around. While these fluctuations are small,
over time they can begin to add up creating monumental drift in a matter of seconds
from the user's perspective. In some cases the algorithms themselves have built in
options to reduce these errors. For the Kalman Filter, error matrices are present
which can tell the system how much to trust raw data, or for a linear classifier as
long as the error variance is present in the training data the algorithm is relatively
unaffected in its predictions, though confidence values might be slightly lower. Thus
it is important to look into the details of the hardware being used to gather exact
variance numbers, even if supplied by the manufacturer. Perform experiments with
the device under known conditions to see what levels of drift are present and map
these into the algorithms that you decide to use or apply filters to the data before it
enters the algorithm to minimize variances.
Errors in the Algorithms
. Most of the algorithms we discussed above are designed
for general purpose and need to be modified to match a specific project. Even so, they
may encounter situations that were unplanned and will produce incorrect data. In the
case of the linear classifier, often better training data will resolve many issues. For
theKalman Filter andHiddenMarkovModels, the initial states can greatly change the
overall performance and prediction capability of the model. To reduce these errors,
