Information Technology Reference
In-Depth Information
between two SPs (standing and sitting) which usually present high interclass error.
The second type occurs during PTs in between two SPs (second 4). This misclassi-
fication is generally characterized by incorrectly predicting PTs as AAs.
In more realistic circumstances, it would be desirable to acknowledge unusual
events (unknown-activity) such as PTswhen the systemdoes not match a newwindow
sample to any of its studied activities. The figure also depicts the expected correct
predictions assuming only the 6 BAs as its classification output and the case when
the unknown-activity class is also taken into account. Moreover, the exploration
of human habits in real life examples also suggests that some sequences of events
are very unlikely to happen, for instance, that a person walks upstairs right after
lying-down and just before standing.
We have developed a set of filters to heuristically improve the probabilistic output
of the SVM using temporal information from each prediction and its neighboring
samples. This process is divided in two parts: probability filtering which directly
handles the probability signals and discrete filtering that further filters the activity
output after the discretization of probabilities into activities.
7.3.3.1 Probability Filtering
The visualization of activity probability signals in time provided us with information
regarding the behavior of BAs. SPs are different from AAs with respect to the way
their associated activity probabilities manifest (the normalized SVM output). This
signal fluctuates more in AAs. As a result, the implemented filters are conditioned
with the type of activity, whether they are an SP or not. They are rule-based and use
the P matrix as input which is composed of the activity probability vectors of the
last s overlapped windows. This number of windows is selected based on the filter
requirements. In this application s
=
5 which is equivalent to a prediction delay
of 5.12 s.
The filters use probability thresholds to define, for instance, whether a class is
considered active (e.g. p c >
threshold ) or condition the filtering of an activity based
on the value of other classes. P = (
represents the application of the probability
filters over the activity sequence and they are described as follows:
P
)
Transition Filter This filter is aimed to remove peaks and transients of dynamic
activities when they appear amongst static ones. This is applied toAAs as they exhibit
a spiky behavior (e.g. during PTs). These usually take a short time (from 2 to 3 s),
therefore this filter measures the length of the activation of these dynamic signals for
a number of overlapping windows (maximum 3). Their filtering is also conditioned
with the intensity of the SPs probability signals in the selected windows. A high
probability in these indicates it is unlikely that an AA can appear simultaneously.
For the PTA-7A method, the AAs are also filtered when the PT output probability
surpasses a threshold. The transition filter is not applied over the PT class because
contrarily to AAs, its appearance is rather short and it is desired to be kept in such
way instead of removing it.
Search WWH ::




Custom Search