United States
Environmental Protection
Agency
Atmospheric Research and
Exposure Assessment Laboratory
Research Triangle Park, NC Z77\ \
Research and Development
EPA/600/S3-90/027 April 1990
SEPA Project Summary
Evaluation of the Wake
Effects on Plume Dispersion
Using Video Image Analysis
Sarah A. Rajala and David S. Trotter
The full report summarizes results
of a cooperative agreement. Video
images of smoke flow in the wake of
a model building were further
analyzed in this research. Three
projects were conducted. The first
evaluated existing image analysis/
processing techniques to determine
the contents of the data, develop a
scheme for separating the smoke
from the background, and determine
the potential for analyzing the motion
characteristics of the smoke flow.
The second used the theory of
fractals to extract information from
the smoke images. These two
projects identified a number of
difficulties in characterizing smoke
images. In the third project, a new
technique for video imaging using
laser sheet lighting was developed
and tested. The resulting smoke
images were more distinct and the
noise levels were lower.
This Project Summary was
developed by EPA's Atmospheric
Research and Exposure Assessment
Laboratory, Research Triangle Park,
NC, to announce key findings of the
research project that is fully
documented in a separate report of
the same title (see Project Report
ordering information at back).
Introduction and Procedure
The Clean Air Act calls for regulations
that specify the use of dispersion models
in evaluating compliance with air quality
standards. As a result, a special need
exists for improved dispersion models to
predict concentrations near sources that
exhibit frequent building downwash
problems. In addition, regulations on the
design of "good engineering practice"
stack heights require a better
understanding of the extent of building
wake effects. Until recently, researchers
have obtained meteorological
measurements and concentrations of air
pollutants by sampling several fixed
points over the spatial region and
temporal period of interest. The recorded
data have then been combined with
available mathematical models to infer
the overall pattern of pollutant
concentrations for risk assessment
Unfortunately, these limited
measurements do not always provide all
the information needed to adequately
understand atmospheric processes under
complex conditions, e.g., near buildings,
mountainous terrain and/or for situations
with^cnernically reacting pollutants. ^
One approach to improving dispersion
models is to combine the traditionally
obtained measurements with images of
smoke to provide a visualization of the
pattern of atmospheric transport. In fact,
exploratory studies at the EPA
Meteorology Division Fluid Modeling
Facility in cooperation with North Carolina
State University demonstrated the
feasibility of using video sequence image
analysis techniques for studying wake
effects. From the results of this initial
study, it was clear that further research
was warranted to develop and evaluate
techniques for recording, processing, and
analyzing video image sequences of
wake effects and to develop
methodologies for utilizing this
information to improve dispersion
models. As ~a result, a cooperative
-------
research agreement (CR-813368) was
established to continue this research.
To achieve the goals of this research,
three topics were addressed. In the first
two, we evaluated different approaches to
analyzing the smoke flow data. Project 1
evaluated standard image processing
techniques for analyzing the video image
sequences. Project 2 explored the
feasibility of using recent work on fractal
theory to model and analyze the video
image sequences. The results of the first
two projects indicated a need for cleaner,
more detailed imagery. We therefore
decided that in Project 3, a laser-sheet
lighting experiment for recording new
video image sequence data would be
designed and implemented. Projects 1
and 2 utilized video image sequence data
collected at the EPA Meteorology
Division Fluid Modeling Facility prior to
this cooperative agreement. In the full
report, we present the results from this
agreement.
Results and Discussion
One characteristic of the image data
we were interested in quantifying was the
motion of the smoke, i.e., the time-
varying changes in the image intensity
values. Because of the nature of the
smoke flow data, it was difficult or
impossible to use existing motion
analysis techniques. As a result, we
approached the problem using two very
different points of view. In the first case,
we applied many of the traditional image
processing/analysis techniques to assess
the contents of the data, to develop a
scheme for separating the smoke from
the background, and ultimately to
determine the potential for analyzing the
motion characteristics of the smoke flow.
The first set of image processing
techniques we applied to the data were
for determining its signaLand noise,
characteristics. First, we displayed the
individual bit planes of the digital image
data in order to see if the least significant
bits contained signal information or noise.
Visually it appeared that the least
significant bit plane contained only a
random distribution of ones and zeros. To
confirm this observation, we estimated
the autocorrelation for the three least
significant bit planes. The estimate of the
autocorrelation for the least significant bit
plane is approximately a unit impulse
function. For bit planes 1 and 2 the
autocorrelations become less like a unit
impulse; there is correlation between
adjacent pixel values. Thus, what we
conclude is that a 7 bit binary
representation is adequate to represent
the signal information contained in this
image data. Alternatively, we can say
that for each 8 bit pixel intensity there is
approximately 1 bit of random error.
Another interpretation is that these data
have a signal-to-noise ratio no greater
than 42 dB. This is based on the rule-of-
thumb that there is approximately a 6 dB
increase in signal-to-noise ratio for each
additional bit of information. This rule-of-
thumb is based on the performance of a
pulse code modulation (PCM) system
with uniform quantization.
In order assess the spatial bandwidth
of the image data, to determine whether
the image data is likely to have aliasing
due to spatial undersampling, and to
visualize the noise, a discrete Fourier
transform (DFT) of the image data was
obtained next. The magnitude of the DFT
of each image frame-was-displayed as-an
image, with the dynamic range of the
data rescaled so that the largest value is
assigned to 255 (white) and the smallest
value to 0 (black). It is clear from these
images that the image data has a low
spatial bandwidth; that is, the data has
been spatially oversampled (the highest
frequency of the data is much less than
the Nyquist frequency), so no aliasing is
present. In addition, the intraframe noise
was not a dominant factor even at the
high frequencies.
Next, several linear and nonlinear
image domain filtering techniques were
applied to the image data to determine
the feasibility of reducing what little noise
was present. For the set of data we have
processed, the spatial median filter
provided the best technique for reducing
unwanted intraframe noise and
background artifacts without sacrificing
the edge information in the images. In the
temporal direction, a temporal low-pass
filter yielded more reduction in interframe
noise than did a temporal median filter.
However,.. Jiejihgr_seemed warranted for
this data.
The next set of image-processing
techniques considered were the edge
enhancement techniques. However, for
the low contrast images in this data set,
edge enhancement is not Oextremely
beneficial, because the edge
enhancement process also enhances the
background. It should also be noted that
the histograms of the edge enhanced
data are different from those of the
original data. If one chooses a histogram-
based segmentation technique (see
1 Section 3), it is useful to know if edge
enhancement helps or hinders the
segmentation process. In this case, it
hinders more than it helps. We were
interested in segmenting the image data
into two pieces, smoke and background.
To do this, it is useful to have bimodal
histograms.
Edge enhancement on this data tends
to make the histograms trimodal. This
leads one to believe that the data should
be separated into three classes. The
question is what three classes? If it were
smoke, dark background, and grid work,
this might be useful. However, because
the range of intensity values for the grid
work overlaps with the range of values for
the smoke, it did not appear that this
would be effective and was not tried. It
should be noted, however, that the
segmentation approach based on optimal
thresholding described in section 3 could
be generalized for trimodal histograms.
Gradient-based edge extraction
techniques were evaluated next. Several
-gradient operators<-were-applied to the
smoke image data. The 3x3 gradient
operators provided edges that were less
sensitive to random fluctuations in the
image data. Although it is difficult to
extract any quantitative information about
the motion of the smoke from these
images, they may yield useful qualitative
information about the smoke flow when
viewed as a temporal sequence.
Since one of the objectives of this
research was to find a way to extract
motion characteristics from the data, we
evaluated methods for segmenting the
image data next. Two approaches were
considered, global thresholding and
optimal thresholding. From our
experiments, we concluded that it is
difficult to heuristically find a single
global threshold that would segment this
image data. However, we found that
optimal thresholding provided us with a
robust means for segmenting the image
data into two regions. We calculated
optimal thresholds for nine subimages for
each frame of image data and found
consistency from frame to frame. The
segmented image data can provide
useful qualitative insight about the outer
boundary of the smoke dispersion when
viewed as a sequence and has proven
useful for estimating motion parameters.
A very simplistic approach to motion
detection involves the subtraction of
adjacent frames in a sequence of image
data. The idea is that stationary objects
will subtract out and leave pixel values of
zero, and in areas where motion has
occurred, there will be nonzero pixel
values. A fundamental assumption of this
method is that all the- images are spatially
registered from frame to frame. We
performed first, second, and third-order
differencing on the image data. We
believe there is potential for obtaining
qualitative insight when viewing the first-
-------
order difference images as a video
sequence. Little insight was gained for
this data from second and third-order
differencing because of the amplification
of interframe noise. Perhaps if a better
means for suppressing the interframe
noise were found, the second and third-
order differencing would be more useful.
Alternatively, if cleaner image data are
recorded, all three difference sequences
may prove to be useful both for
qualitative and further quantitative
assessment of the motion characteristics.
Assuming that the boundaries of the
regions in the segmented image data
reasonably describe regions containing
the smoke, they can then be treated as
an object with mass and area. The
integrated optical density (IOD) is a
measure of the mass contained in each
thresholded image of smoke. This
information was then used to calculate
the center of mass for each of the
thresholded images. This allowed us to
track and estimate the motion of each
segment from frame to frame. Changes
in area and integrated optical density
generally reveal the stages of smoke
accumulation. An increase in area and
optical density would imply the end of the
cycle of smoke accumulation, and a
decrease in area would be expected to
follow. An increase in optical density and
a decrease in area might imply that the
smoke is beginning its cycle of
accumulating.
The last image processing operation
performed was an attempt to visualize the
boundary deformation of the segmented
image data from frame to frame. Using
the thresholded images, the boundary of
each image was extracted in a binary
image. Only the points on the boundary
contour have a nonzero pixel value. By
overlaying these binary images, a
^comparison ._oJL the^de.fprmation of .the._
boundaries over time may be made. At
this time only a qualitative comparison
has been made, but future work could
consider the boundary deformation more
quantitatively. Attributes of shape and
size could be computed and compared.
This same process could be performed
inside the smoke to characterize the
boundary between diffuse and more
concentrated smoke. Further work in this
area would possibly derive significant
information about changes in
concentration of the smoke.
In the second project, we used the
theory of fractals for extracting
information from the smoke dispersion. A
new model based on fractal concepts
was developed and applied to a
sequence of video images of smoke
plume dispersion. The new model pro-
vides some qualitative and quantitative
interpretations of transient flows by
utilizing several fractal parameters to
determine the flow activity. The
parameters used in this study were the
horizontal fractal parameter, the vertical
fractal parameter, and two higher-order
fractal dimensions, the correlation and
information dimensions. The results were
related to the motion characteristics and
plume concentration. The computed
fractal parameters were tested for various
sized windows and provided us with
information about the spatial-temporal
behavior of the smoke flow. The initial
results appear to be promising, but
further research is needed to collect and
analyze a set of data that includes video
imagery as well as output from other
measurement systems, e.g., hot-wire
anemometry or hydrocarbon tracer. As a
part of this research project, a review of
dimensionality measurements for chaotic
dynamical systems was prepared.
In the third project, we designed and
implemented an alternative approach for
capturing the smoke flow image data.
The image data recorded previously were
captured under flood lighting conditions.
As a result, the two-dimensional image
data was a projection of the three-
dimensional volume of smoke flow onto a
two-dimensional image plane. In the new
system, laser-sheet lighting was used to
generate a plane of illumination. Hence,
we could selectively illuminate planar
surfaces in the three-dimensional volume
of smoke flow. Data were collected for
various planar sheets of light generated
by both a rotating polygonal mirror and a
fixed lens configuration. Because these
data were collected in the last six months
of the research agreement, only
preliminary evaluations on their usability
_have been performed. Jn_.the preliminary
analysis, we found that the images gave
clearer, more detailed data for analysis.
The next step should be to apply
techniques tested in Projects 1 and 2
using the new data.
Summary and Conclusions
We estimated basic motion
characteristics of the centroid of the
segmented image data. The approaches
examined were simple and no constraints
had to be employed regarding the
linearity of the motion. The image data
files analyzed in the major portion of this
research were collected using flood
illumination. Hence, the motion
information extracted was for 3-D smoke
flow data projected onto a 2-D image
plane. Information that cannot be
recovered is lost in this projection. Future
work should include extending the motion
estimation technique to include higher-
order moments of the segmented image
data and to applying this approach to the
image data recently collected using the
laser sheet lighting. In addition, other
features of the image, such as
correlations of the original image data,
boundaries of segmented data, and
intensity contours should be analyzed as
a function of time.
U.S. GOVERNMENT PRINTING OFFICE: 1990/748-012/20076
-------
Sarah A. Rajala and David S. Trotter are with the Department of Electrical and
Computer Engineering, North Carolina State University, Raleigh, NC 27695-7911.
Alan H. Huber is the EPA Project Officer (see below).
The complete report, entitled "Evaluation of the Wake Effects on Plume
Dispersion Using Video Image Analysis," (Order No. PB 90-186 354/AS;
Cost: $11.00, subject to change) will be available only from:
National Technical Information Service
5285 Port Royal Road
Springfield, VA 22161
Telephone: 703-487-4650
The EPA Project Officer can be contacted at:
Atmospheric Research and Exposure Assessment Laboratory
U.S. Environmental Protection Agency
Research Triangle Park, NC 27711
United States
Environmental Protection
Agency
Center for Environmental Research
Information
Cincinnati OH 45268
Official Business
Penalty for Private Use $300
EPA/600/S3-90/027
------- |