Presentation 11/27: Robust Distributed Detection Using Low Power Acoustic Sensors
7 views
Skip to first unread message
Su Kim
Nov 27, 2007, 2:13:56 PM11/27/07
to ASU:CSE535 FALL 07 Mobile Computing
Authors: Brian P. Flanagan and Kenneth W. Parker,
Title: "Robust Distributed Detection Using Low Power Acoustic
Sensors,"
Published: The MITRE Corporation Technical Paper, 2005
Authors: Quach and K. Lo,
Title: "Automatic target detection using a ground-based passive
acoustic sensor,"
Published: Proceeings of Information, Decision and Control (IDC 99),
8-10 Feb. 1999 Page(s):187 - 192, 1999
Title: "Robust Distributed Detection Using Low Power Acoustic
Sensors,"
Published: The MITRE Corporation Technical Paper, 2005
Authors: Quach and K. Lo,
Title: "Automatic target detection using a ground-based passive
acoustic sensor,"
Published: Proceeings of Information, Decision and Control (IDC 99),
8-10 Feb. 1999 Page(s):187 - 192, 1999
swaroo...@asu.edu
Nov 27, 2007, 5:57:34 PM11/27/07
to ASU:CSE535 FALL 07 Mobile Computing
Answers to some of the questions:
1. In the Maximum Power Detector, how is the value of Alpha
determined?
---- The detection threshold is given by <alpha> times P, where P is
the sum of the spectral powers at 10 Hz, 50 Hz, 100 Hz and 200 Hz in
the noise spectrum. Since the power of the frequencies in the noise
spectrum is very low as compared to the power of the frequencies in
the data set, we use a scaling factor to increase the noise power
level and then compare these with the power levels in the data set.
This is done to reduce the possibility of a false alarm. My guess is
that the value of <alpha> would depend on the power levels of the
noise frequencies at 10 Hz, 50 Hz, 100 Hz and 200 Hz.
2. What would be the best window size for calculating the short time
Fourier transform?
--- This would depend on the frequency resolution you wish to attain
in the spectral graph. The more you decrease the window size, the less
chances of detecting low frequencies. This is because the signal with
low frequency may not complete its period within the time window, if
that time window is too small. Thus, there is generally a tradeoff
between frequency resolution and time resolution. So the window size
would depend on the what frequencies do you want to see in the
spectral graph.
3. Possibility of not collecting enough data for evaluation in the
distributed technique..
--- In the experiment scenario, they used 4 sensors, in a 10X10 area
in a parking lot and recorded results for a period of two hours. In
this window, they recorded the presence of 84 targets, and two out of
four nodes were required to flag a detection in order for the
distributed component to register the presence of a target. So, maybe
they did not collect enough data. But whats lacking here, and was also
pointed out in the class, is the performance of detection for
different number of nodes, and for varying ranges. So perhaps this was
not a holistic experiment. What was also lacking here, is the spectral
analysis of the target of interest. But again, since this experiment
is performed in the time domain, and not frequency domain, and hence,
the utility of spectral analysis is debatable.
In case I have missed out on some of the questions, please post them
here, and I would attempt to answer them to the best of my
understanding.
regards,
Swaroop Shere
1. In the Maximum Power Detector, how is the value of Alpha
determined?
---- The detection threshold is given by <alpha> times P, where P is
the sum of the spectral powers at 10 Hz, 50 Hz, 100 Hz and 200 Hz in
the noise spectrum. Since the power of the frequencies in the noise
spectrum is very low as compared to the power of the frequencies in
the data set, we use a scaling factor to increase the noise power
level and then compare these with the power levels in the data set.
This is done to reduce the possibility of a false alarm. My guess is
that the value of <alpha> would depend on the power levels of the
noise frequencies at 10 Hz, 50 Hz, 100 Hz and 200 Hz.
2. What would be the best window size for calculating the short time
Fourier transform?
--- This would depend on the frequency resolution you wish to attain
in the spectral graph. The more you decrease the window size, the less
chances of detecting low frequencies. This is because the signal with
low frequency may not complete its period within the time window, if
that time window is too small. Thus, there is generally a tradeoff
between frequency resolution and time resolution. So the window size
would depend on the what frequencies do you want to see in the
spectral graph.
3. Possibility of not collecting enough data for evaluation in the
distributed technique..
--- In the experiment scenario, they used 4 sensors, in a 10X10 area
in a parking lot and recorded results for a period of two hours. In
this window, they recorded the presence of 84 targets, and two out of
four nodes were required to flag a detection in order for the
distributed component to register the presence of a target. So, maybe
they did not collect enough data. But whats lacking here, and was also
pointed out in the class, is the performance of detection for
different number of nodes, and for varying ranges. So perhaps this was
not a holistic experiment. What was also lacking here, is the spectral
analysis of the target of interest. But again, since this experiment
is performed in the time domain, and not frequency domain, and hence,
the utility of spectral analysis is debatable.
In case I have missed out on some of the questions, please post them
here, and I would attempt to answer them to the best of my
understanding.
regards,
Swaroop Shere
jay.e...@asu.edu
Nov 30, 2007, 4:20:29 PM11/30/07
to ASU:CSE535 FALL 07 Mobile Computing
Here is my critique of these papers:
Papers:
* "Automatic Target Detection Using a Ground-based Passive Acoustic
Sensor" by Anthony Quach and Kam Lo
* "Robust distributed detection using low-power sensors" by Brian
Flanagan and Kenneth Parker
Critique of papers
Summary
These two papers present research in the area of using acoustics from
remote wireless sensors for the identification of vehicular traffic.
The studies dealt with slightly different target types (one study
focused on military targets such as tracked vehicles, heavy wheeled
vehicles and aircraft, the other study dealt with civilian wheeled
vehicles such as van and buses). Both studies attempted to deal target
identification in the face of background noise such as wind. One study
used Harmonic Signatures (HS), the other studied a detection method
that used multiple sensors.
Key Contributions
The mechanisms developed to deduce vehicle type from acoustic sensors
and experimental data that demonstrates that it is feasible to use
acoustic data for vehicle identification are key contributions from
these papers. The paper produced a variety of target detection
methods, and collected data to understand what circumstances each of
the methods work well in.
Analysis & Results
The earlier paper used harmonic signatures, which are robust (when
they work). The more recent paper noticed that while HS seem to work
for military targets, it does not work so well on certain types of
civilian vehicles. Frequency domain analysis distributed over multiple
sensors (using "binary fusion") seems to work better.
Critique
I would like to have seen some analysis that explains why HS works
well on military vehicles but not on civilian vehicles. Also, since
the second study proposed a different detection method, perhaps a
hybrid approach could be developed. This should have been in the
conclusion section of the second paper.
Conclusions
Several different methods to detect targets were studied. No single
method work best in all situations. HS seems to work for certain
classes of military targets (tracked vehicles, heavy trucks), while
frequency domain analysis from multiple sensors can be used to
identify certain types of civilian vehicles.
One study that could be conducted would be a hybrid approach - use HS
and binary fusion from multiple sensors.
Papers:
* "Automatic Target Detection Using a Ground-based Passive Acoustic
Sensor" by Anthony Quach and Kam Lo
* "Robust distributed detection using low-power sensors" by Brian
Flanagan and Kenneth Parker
Critique of papers
Summary
These two papers present research in the area of using acoustics from
remote wireless sensors for the identification of vehicular traffic.
The studies dealt with slightly different target types (one study
focused on military targets such as tracked vehicles, heavy wheeled
vehicles and aircraft, the other study dealt with civilian wheeled
vehicles such as van and buses). Both studies attempted to deal target
identification in the face of background noise such as wind. One study
used Harmonic Signatures (HS), the other studied a detection method
that used multiple sensors.
Key Contributions
The mechanisms developed to deduce vehicle type from acoustic sensors
and experimental data that demonstrates that it is feasible to use
acoustic data for vehicle identification are key contributions from
these papers. The paper produced a variety of target detection
methods, and collected data to understand what circumstances each of
the methods work well in.
Analysis & Results
The earlier paper used harmonic signatures, which are robust (when
they work). The more recent paper noticed that while HS seem to work
for military targets, it does not work so well on certain types of
civilian vehicles. Frequency domain analysis distributed over multiple
sensors (using "binary fusion") seems to work better.
Critique
I would like to have seen some analysis that explains why HS works
well on military vehicles but not on civilian vehicles. Also, since
the second study proposed a different detection method, perhaps a
hybrid approach could be developed. This should have been in the
conclusion section of the second paper.
Conclusions
Several different methods to detect targets were studied. No single
method work best in all situations. HS seems to work for certain
classes of military targets (tracked vehicles, heavy trucks), while
frequency domain analysis from multiple sensors can be used to
identify certain types of civilian vehicles.
One study that could be conducted would be a hybrid approach - use HS
and binary fusion from multiple sensors.
weiji...@asu.edu
Dec 5, 2007, 1:08:34 AM12/5/07
to ASU:CSE535 FALL 07 Mobile Computing
please find my summary below:
SUMMARY FOR :
Summary:
Both the papers introduced some techniques for using passive acoustic
sensor for target detection. This difference is, for the first paper,
which is robust distributed detection using low power acoustic
sensors, though three different types of detectors are examined,
namely, log sum, harmonic set and maximum power, only a single sensor
is taken into account each time. While for the second paper, that is
automatic target detection using a ground-based passive acoustic
sensor, a network of sensors is deployed and used for implementing the
target detection algorithm. Both the paper again analysis the
background noise and the target signatures in time domain and
frequency domain and try to use the information extracted to implement
or improve the algorithms.
Pros:
Both the paper give detailed analysis of back ground noise supported
with real data collection. And they are examined in both the time
domain and the frequency domain.
For the first paper, three types of sensors are tested with real
acoustic data recorded during the transit of a variety of wheeled
vehicles pass the sensor. Sensor wave forms, the probability of
detection as well as false alarm rates are given so that the argument
is highly supported by those data.
For the second paper, it employs more sensors and so it possesses a
better capacity of detecting targets. OS-CFAR algorithm is adopted so
that it will give good result even in the non homogeneous environment.
Cons:
For the first paper, the sensor is only capable of detecting targets
that emit strong acoustic signal that consisting of harmonically
related tones, thus the applications are limited. The second find this
problem in the first paper and tried to improve it by employing a
network of sensors. It achieves some kind of improvement for target
detecting, however, the probable applications are still limited. While
at the same time, the deployment cost increased dramatically,
resulting in a question that does those efforts worth it or not?
The back ground noise is analysis in details; however, when coming to
the experiments, the various environments are not taken into account.
At least, the comparison between strong wind an light wind or no wind
are not presented, degrading the significance for the former analysis.
The relationship between detection performance and false alarms are
still tricky. And some of the values used in the paper are less of
illustration or support.
SUMMARY FOR :
Robust Distributed Detection Using Low Power Acoustic Sensors
Automatic target detection using a ground-based passive acoustic
sensor
Weijia Che
sensor
Summary:
Both the papers introduced some techniques for using passive acoustic
sensor for target detection. This difference is, for the first paper,
which is robust distributed detection using low power acoustic
sensors, though three different types of detectors are examined,
namely, log sum, harmonic set and maximum power, only a single sensor
is taken into account each time. While for the second paper, that is
automatic target detection using a ground-based passive acoustic
sensor, a network of sensors is deployed and used for implementing the
target detection algorithm. Both the paper again analysis the
background noise and the target signatures in time domain and
frequency domain and try to use the information extracted to implement
or improve the algorithms.
Pros:
Both the paper give detailed analysis of back ground noise supported
with real data collection. And they are examined in both the time
domain and the frequency domain.
For the first paper, three types of sensors are tested with real
acoustic data recorded during the transit of a variety of wheeled
vehicles pass the sensor. Sensor wave forms, the probability of
detection as well as false alarm rates are given so that the argument
is highly supported by those data.
For the second paper, it employs more sensors and so it possesses a
better capacity of detecting targets. OS-CFAR algorithm is adopted so
that it will give good result even in the non homogeneous environment.
Cons:
For the first paper, the sensor is only capable of detecting targets
that emit strong acoustic signal that consisting of harmonically
related tones, thus the applications are limited. The second find this
problem in the first paper and tried to improve it by employing a
network of sensors. It achieves some kind of improvement for target
detecting, however, the probable applications are still limited. While
at the same time, the deployment cost increased dramatically,
resulting in a question that does those efforts worth it or not?
The back ground noise is analysis in details; however, when coming to
the experiments, the various environments are not taken into account.
At least, the comparison between strong wind an light wind or no wind
are not presented, degrading the significance for the former analysis.
The relationship between detection performance and false alarms are
still tricky. And some of the values used in the paper are less of
illustration or support.
Reply all
Reply to author
Forward
0 new messages