Information Technology Reference
In-Depth Information
polynomial curves. The method was implemented as described in the paper,
taking care to use the same 8-bit range for the chroma planes. The results were
also aggregated into 4x4 blocks using a majority voting rule, to make comparison
with our method as fair as possible. We consider this method to be the state-
of-the-art single-pixel fire color filter against which to judge the benefits of our
contextual modeling.
The first part of the test set consists of 30,000 frames depicting fire, to obtain
the detection rate. These video frames show a number of controlled fires in an
outdoor firemen training complex built to resemble an industrial site. The fires
include a burning petroleum tank, a ruptured gas pipe, a round tank engulfed
in flames and a fire in a maintenance trench. The fires were monitored by six
cameras of different types, placed on different elevation levels and angles. The
fire is considered detected as soon as at least one of its pixel blocks is detected as
a fire block. In the interest of fairness, we should note that the training images
for our method were captured on the same site, albeit at a different time with
different sunlight levels.
The second part of the test set contains over 18,000 video frames captured
from a moving vehicle in an urban environment. This set is representative of the
occurrence of fire-colored objects to be expected in the busiest environments,
e.g. red and yellow clothing, vehicles or advertising. An image is counted as a
false positive when one or more blocks in the image are classified as fire.
The results obtained on this data set are shown in Table 1. The reference method
by Celik
et al.
scored a detection rate of 99.95% on the fire frames, while generat-
ing over 50% false positives. This illustrates the high occurence of fire-colored ob-
jects in the second dataset: over half of the frames contain at least one fire-colored
4x4 block. In comparison, the detection rate of our proposed method after just
the first stage was 99.98%, with a false detection rate of 42.82%. This shows that
even after just the least discriminative of the two stages, there is an improvement
over the reference method. After both stages of the method, the detection rate
drops only slightly to 99.57%, while false positives are much reduced to 21.80%.
These statistics prove the adequacy of the system as standalone fire detector. The
cases in which the fire was not detected are mostly transition phases, either just
after the fire was started or when it was nearly extinguished. One can reasonably
assume that any spreading fire will be detected. The false negatives can thus be
considered rare and temporary manifestations of fire in which the spectral texture
is coincidentally and atypically low.
5Conluon
We have designed an MRF-based visual fire detection system which is easy to
train, and requires optimization of just one critical parameter (the lower classi-
fier energy threshold) rather than setting multiple fixed color rules. Furthermore,
after training on basic, generic ground-truth data the method is proven to yield
very good detection rates in a variety of circumstances, while at the same time
significantly reducing false positives over standard color-based methods. More-
over, it does not rely on any temporal information, and can therefore be applied
Search WWH ::
Custom Search