Assessment of Structural Damage Following Tornadoes using Point Cloud
This video describes a poster that assess the structural damage done following tornadoes in a point cloud.
Toggle between list and paragraph view.
- [00:00:00.000]Hello, my name is Mario
Esquivel and here's my poster.
- [00:00:04.156]The title of my summer research is the
- [00:00:07.243]assessment of structural damage following
- [00:00:09.296]tornadoes using point cloud.
- [00:00:11.166]I am an undergraduate research assistant
- [00:00:14.017]studying construction management.
- [00:00:16.363]The coauthors are Dr. Mohammadi and
- [00:00:18.683]my faculty mentor is Professor Wood.
- [00:00:21.367]There are also two other
research assistants working
- [00:00:24.497]with us and their names
- [00:00:26.447]are Pooja and Awang.
- [00:00:28.648]Oftentimes, tornadoes severely damaged
- [00:00:32.648]leaving a path of their
- [00:00:34.721]behind. To address these disasters
- [00:00:37.551]my team has made it their objective to
develop a reliable post disaster damage
- [00:00:41.551]identification, predictive
model for damage assessment.
- [00:00:45.551]This will be useful in emergency
response and recovery operations
- [00:00:49.551]following catastrophic events.
To complete this objective,
- [00:00:53.551]we have to follow these methods.
- [00:00:56.872]First data is collected using a UAS with
a camera. From the images taken
- [00:01:01.744]we then process them using Pix4D
software to create a point cloud.
- [00:01:05.767]This procedure is demonstrated in figure
1. After the point cloud is created,
- [00:01:10.894]the utilization of a software called
CloudCompare is necessary to segment the
- [00:01:15.274]cloud. The cloud is segmented
into 14 different categories.
- [00:01:20.731]To assess the structural damage,
- [00:01:24.011]we looked at the percentage of
roof failure and wall structure.
- [00:01:28.036]Field assessments done during the same
time are also indicators of what level of
- [00:01:32.127]damage the structure has endured. Below
- [00:01:36.127]there is a 3D convolutional neural
network with six convolutional layers and
- [00:01:40.757]four fully connected layers.
- [00:01:43.029]The segmented and labeled point cloud
data within this work will be used to
- [00:01:47.929]develop the 3D convolutional neural
network models for damage assessment
- [00:01:55.242]There's also an example of the DJI Mavic
Pro used to capture the images as well
- [00:02:01.762]as the cloud compare software used
to segment the point clouds. Up next,
- [00:02:07.657]we have 14 different classifications
for the segmented objects.
- [00:02:12.156]9 of these labels are more general
terms, while the other 5 are our main
- [00:02:16.156]focus of work. The general
labels are boats, vehicles,
- [00:02:20.156]debris, fallen trees,
trees, water bodies,
- [00:02:24.370]roads, terrain, and poles. The five
structural levels are destroyed,
- [00:02:29.460]which can be classified by having
more than 15% roof structure failure,
- [00:02:33.928]failure of wall structure, or
more than 25% roof deck loss.
- [00:02:39.528]The next is severe,
- [00:02:41.895]which has more than 50% roof cover or
wall cladding damage, between 5 to 25%
- [00:02:48.826]roof sheathing loss, or less than 15% roof
structure damage. Moderate, has between
- [00:02:55.296]15 to 50% roof cover, wall cladding
damage, or less than 5% roof substrate
- [00:03:03.220]Minor damage is less than 15% roof
cover or wall cladding damage.
- [00:03:07.859]The last label is no damage, which
has no visible exterior damage.
- [00:03:13.048]So far, during our research time,
we have worked on two datasets.
- [00:03:17.428]The first was a Mount Juliet
housing triple crown neighborhood
- [00:03:21.432]and the second was the West Lebanon
housing Stonebridge neighborhood.
- [00:03:25.452]These data sets are then split up into
4 segments for us to work on together.
- [00:03:29.452]For each of these sets,
- [00:03:30.892]a UAS derived point cloud was collected
from a tornado near Nashville,
- [00:03:35.734]Tennessee, which occurred
on March 3rd, 2020.
- [00:03:39.734]Field assessments are assessments done
by the StEER team following the event.
- [00:03:45.024]These are useful because images with
the UAS are not able to display what
- [00:03:49.885]happened to these structures.
- [00:03:51.879]They also tend to group similarly
damaged structures in one area.
- [00:03:58.448]For my labeling results,
- [00:04:00.208]we have been able to complete the Triple
Crown neighborhood as of now. Figure
- [00:04:04.068]8 demonstrates a red,
- [00:04:05.378]green and blue image of the point cloud
while figure 9 demonstrates the same
- [00:04:10.858]image except the objects have been
labeled and colored they're represented
- [00:04:14.598]label. Analyzing my results,
- [00:04:17.780]I found there was a massive debris field
toward the North region due to roof
- [00:04:22.600]tile and structure failures. Meanwhile,
- [00:04:25.220]a majority of the undamaged structures
are located in the South of the region
- [00:04:29.220]because they happen to be located
further away from the path of the tornado.
- [00:04:33.220]Generally, most of the structure labels
are represented in the same area.
- [00:04:38.723]In conclusion,
- [00:04:40.204]this research has allowed me to utilize
a software to classify point clouds for
- [00:04:44.742]different levels of damaged structures.
- [00:04:47.252]The field assessments helped create a
general region of damaged structures and
- [00:04:51.584]this labeled data is used for deep
learning models that can identify damage
- [00:04:56.560]distribution using the point clouds
of damaged areas in the aftermath of
- [00:05:01.280]windstorms. Lastly,
- [00:05:03.055]future work within this research
includes completing the labeling for
- [00:05:07.252]Stonebridge, verification of the labels,
and processing the tagged point clouds
- [00:05:12.642]into instances with multiple labels.
- [00:05:14.992]These are the acknowledgements I would
like to make because they have allowed me
- [00:05:19.382]to be a part of this great research.
- [00:05:22.773]Funds by the NsF StEER used
by Richard Wood and his team.
- [00:05:27.053]Holland Computing Center of
the University of Nebraska,
- [00:05:29.743]which receives support from the Nebraska
Research Initiative and the College of
- [00:05:34.835]Engineering Graduate Program at the
University of Nebraska-Lincoln through the
- [00:05:39.545]Summer Nebraska Engineering
Research Program (SNERP).
- [00:05:43.545]And that concludes the presentation
of my poster. Thank you for your time.
Log in to post comments