2. Worldwide KORUZA Experiment (1): Project Aim
• Low cost internet access is the primary goal of this time
• Prompt deployment of information network is the key to bridging the digital
divide
• “Last mile bottleneck/lag” due to high deployment cost, and RF spectrum
congestion disallows the use of wireless network in more urban areas
• Free space optical communication is extremely viable in such situation
• The aim of the project is develop low cost 3D printable open source FSO
system for use as a primary wireless network in the community
• The MEng project had actively contributed to the development of KORUZA
system
3. Worldwide KORUZA Experiment (2): KORUZA System
• Provides 1300nm/1500nm, 1Gbps communication at
up to 100m distance
• Provides bidirectional optical path
• Contains a RF management communication link used
for control electronics
• Wavelength choice of KORUZA system is optimal and
intersects with free space optical windows and the
fibre optical windows
• Currently in the testing phase of 5th generation
prototype
• Deployed in 10 locations as part of the worldwide
KORUZA experiment
4. Major contribution of the MEng project
• Development of low cost measurement system and devices
• Modification of the SFP module to increase the incident angle tolerance
• Replacement of the RF management communication link with auxiliary green
laser communication channel to make KORUZA an optical only
communication system
• Development and testing of efficient auto alignment algorithms
• Development of a novel data processing technique to remove the errors in the
sensors of KORUZA system
• Investigation of the data collected from KORUZA system to evaluate data
correlations and form hypothesis
• Development of a low cost fog detection system
5. Individual Contribution to the project
1. Random and Burst Error Removing Code
2. Analysis of Data Acquired from KORUZA Units
3. Fog Detection System via Image Based Visibility
Estimation using a Webcam
6. 1. Random and Burst Error Removing Code
Motivation: data acquired from KORUZA are used
to make scientific research and analysis. However,
data from KORUZA unit is very prone to noise and
errors. There are errors in visibility
measurements. So an error removing code was
imperative.
7. Types of error and their complications
• Burst Errors: in the context of the project we
define burst error as a sequence of values
which are significantly higher than
surrounding values (slightly different from
telecommunication definition).
• Burst errors can completely impair any useful
trend in the data sets, making it impossible
to do statistical analysis on data sets
• Random Errors: random noise which follows
Gaussian distribution with zero arithmetic mean,
where the values are dispersed randomly around
the true value.
• Minimised to a large extent by a simple moving
average filter.
8. Error correcting code
In the project the data are treated for errors via a
mixture of three processing
1.range capping
2.“bad data removal”
3.moving average filtering
And it is carried out in that sequence….
9. Range Capping
• Set a range of values which is allowed to pass
• While traversing the data if a value falls out of the range it will be substituted by the
previous value
• Chose not to substitute the value with the average of previous few values as averaging is
performed by the subsequent sections of the error correcting code
• Burst error is best removed by range capping
10. Bad Data Removal
• 2 windows having the same start point traverses the whole data set; they move side by side.
• Window-b is the sequence of values over which an average is calculated to form the criterion- if current value under
consideration deviates a certain percentage (5% is used in most case) from this average, then the value is error
corrected
• For error correction, the value under consideration is substituted with the average calculated over legitimate values
(meet the same criterion described above) within the sequence of data represented by the window-a
• When first few data are considered, the length of the windows are shortened and adjusted to accommodate the whole
sequence
• Original uncorrected data set is considered for both error correction and forming the ‘average criterion’
11. The reason for having 2 windows
Two separate windows are required because:
• On one hand we a need small window for error correction to capture the local effect
of a trend in data
• And on the other hand we require a bigger window to form the ‘average criterion’
• So size requirements of window are different for error correction and for forming the
‘average criterion’. This best explains the reason for using two different windows.
12. Result of ‘Bad Data Removal’
• removes the high frequency noise which are in large deviation from surrounding values and it also
smooth the data to a good extent
• does a mixture of both to a certain extent, it removes a lot of burst errors and it also smoothes the
data.
• Later it will be illustrated that this treatment is very good at eliminating burst errors.
13. Moving Average
• simple moving average described by a FIR filter
where all of the filter coefficients are equal
• remove short term or high frequency fluctuations
14. Result 1: Progressive effects of the error removing code on the visibility
dataset in a nutshell
15. Result 2: Progressive effects of the error removing code on a data set
acquired from KORUZA unit
Significance: difference of action brought about by ‘bad data removal’ and moving average.
• ‘Bad data removal’ might not pick up random errors which are not in large deviation from
average surrounding value
• they are best removed by moving average filter.
16. Result 3: Progressive effects of the error removing code on another data set
acquired from KORUZA unit
Significance: ‘bad data removal’ is very good at combating burst errors
17. 2. Analysis of Data Acquired from KORUZA Units
• KORUZA FSO unit acquires an enormous amount of real
time data
• Total of 19 sets of data acquired from various sensors
implanted in the KORUZA unit.
• The team analysed the processed data to find any
correlations between data sets.
• Strong correlation between received power of receiver
module and the temperature of transmitting module
18. Analysis of power and temperature data sets
• correlations between received power and
temperature of SFP module of opposing KORUZA
unit (four time series) are investigated
• Correlation of power (unit-A) vs temperature(unit-B)
= -0.6254
• Correlation of power (unit-B) vs temperature(unit-A)
= -0.8941
• received power and SFP module temperature are in
strong anticorrelation
• Whenever the temperature increases the received
power of opposing unit drops. This is due to
misalignment caused by thermal heating
• When the temperature decreases the received power
rises again, this is probably due to realignment of
KORUZA unit due to thermal contraction.
• opposite scenario is very much possible
19. Result and Conclusion
• Fine dependence between received power and
temperature change of opposing KORUZA units.
• Set up the motivation to plant a control unit in the
KORUZA which reacts to any temperature changes and
realigns itself. This can be incorporated as a separate
mechanism to that of the principal realignment algorithm
20. 3. Fog Detection System via Image Based Visibility
Estimation using a Webcam
Motivation:
• Under fog condition the signal is so heavily attenuated that there is no point in aligning
the receiving unit
• If the alignment algorithm persists then opposing FSO units would further misalign to a
point where the receiving unit cannot even point to any beam within its line of sight.
• We can know if a system failure is due to fog or anything else.
• According to ‘National Oceanic and Atmospheric Administration’ fog is airborne water
droplets or ice crystals which causes reduction of horizontal visibility to less than 1km
• Need to devise an approach to ‘roughly estimate’ visibility and form a visibility
threshold
• When visibility falls below a certain threshold the opposing FSO units alarmed about
fog condition and is triggered to stop alignment operation.
21. Methodology (1)
Dark Object Approach:
• In the dark object method the contrast of a dark
object against a bright background is measured.
• The object has to be dark enough for it to be
assumed as a black body.
• Visibility is the maximum distance an object can
be distinguished against an ideal bright
background.
22. Methodology (2)
• Pixel value (PV) represent the brightness of
an image.
• We characterise the relationship between
ratio of pixel values and ratio of exposure
through calibration of camera response
curve, which is a plot of log(exposure time)
against log(pixel value).
23. Methodology (3)
• do not need an accurate measurement of visibility
• Instead we need a threshold value which forms the criteria such that any other non-fog events must
not have values less than the threshold that will trigger the alarm.
• Most importantly the webcam we are using constantly auto-adjusts the different parameters like
gama, hue and this constantly changes the camera response curve anyway.
• Calibration plot in previous figure is a one to one mapping showing a single trend. Any non fog event
would never cause visibility the lower than fog event.
• We can remove the calibration of camera response curve.
• Pixel value is acquired from the grayscale image.
24. Behaviour of visibility function
• Visibility must be a positive
value; so any negative values
are invalid and will not be treated
for threshold criteria.
• NS (near site) PV must be
smaller than FS PV (far site) for
visibility to be valid
• Low visibility region is much
less sensitive to changes in the
pixel values.
• So the low visibility region where
the threshold lies is much more
tolerant to error.
25. Flow chart to acquire daytime visibility
Can measure visibility only
during daytime
26. Observation of salient features across daytime visibility plots
Medium visibility: mountain
vaguely appearing in the
horizon
Low visibility: mountain
completely vanishes from the
horizon
High visibility: mountain is
crystal clear in the distant
horizon
Observation: system reads visibility trend absolutely correctly
27. Fog Event
Small trees and building
that is usually on the
landscape completely
disappeared
Everything excepting the
mountain is visible
• Observation: System can read visibility around threshold accurately and hence can
act as a fog detection system.
• A reasonable threshold would be in the range of 900m-1300m.
• The system would trigger an alarm if the threshold continues for a period of 5 to 10
sample time.
28. Limitations (1)
• Sunlight saturating the far site PV close to the
background PV, burning a hole in the visibility,
creating a false impression of a fog event
• Shadowing of near site object can also burn a
hole, but to a much smaller extent. The system is
robust against this affect.
• Effect of sunlight falling directly on the camera
and saturating the near site PV. This causes NS
PV to rise above FS PV, and hence causes invalid
visibility values. Invalid values and low visibility
values are on opposite side of the asymptote, hence
these two types of values do not interfere. Thus the
system gains a inherent immunity against this affect.
29. Limitations (2)
• One way to overcome all these problems is by ideal positioning and orientation
of the camera.
• The camera should be positioned high above the ground level, facing down in
the north-south direction to avoid sunlight or reflectance of sunlight falling on the
aperture.
31. Possible improvements and modifications of the system
1. Image based visibility quantification using two
webcams
• This is basically the same effect, instead of using one
camera, we use two camera placed at two different
positions to measure the contrast of the same object.
• The measurement of the system is very accurate
• Method can read visibility during night time.
• This method can overcome all the problems mentioned
previously.
2. Low-light tone mapping described in
[Kirk, A. G., & O'Brien, J. F. (2011).
Perceptually based tone mapping for low-light
conditions. ACM Trans. Graph., 30(4), 42.] to
further enhance the night time visibility
reading.
3. Image registration to align the images
• The webcam suffer some positional shifts
due to various environmental effects. This
in turn will induce a shift in the image
position.
32. Conclusion to my part
1. Random and Burst Error Removing Code
• Data from KORUZA unit is very prone to noise and errors, especially the damaging burst errors; so
an effective error removing code is imperative.
• Devised mathematically simple but elegant and effective error removing code; completely met our
purpose
2. Analysis of Data Acquired from KORUZA Units
• Fine dependence between received power and change in temperature of opposing KORUZA units.
• Set up the motivation to plant a control unit in the KORUZA which reacts to any temperature changes
and realigns itself.
3. Fog Detection System via Image Based Visibility Estimation using a Webcam
• Fully analysed a working prototype of a fog detection system via image based visibility estimation
using a webcam.
• Carried out optimization to turn a system by leaving out the calibration of camera response curve
• made critical analysis on possible limitations of the system and suggested well researched solutions
• proposed some advantageous and worthwhile modifications and improvements