Uav Ugv 2

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-2/W13, 2019
ISPRS Geospatial Week 2019, 10–14 June 2019, Enschede, The Netherlands
THE JOINT RESEARCH PROJECT ANKommEn – EXPLORATION USING

AUTOMATED UAV AND UGV
A. Alamouri1, , M. Gerke1, S. Batzdorfer2, M. Becker2, U. Bestmann2, M. Bobbe2, Y. Khedar2, T. Blume3, J. Schattenberg3,

J. Schmiemann3
1
Technische Universitaet Braunschweig, Institute for Geodesy and Photogrammetry, Braunschweig, Germany
{a.alamouri, m.gerke}@tu-braunschweig.de
2
Technische Universitaet Braunschweig, Institute of Flight Guidance, Braunschweig, Germany
{s.batzdorfer, m.becker, u.bestmann, m.bobbe, y.khedar}@tu-braunschweig.de
3
Technische Universitaet Braunschweig, Institute of Mobile Machines and Commercial Vehicles, Braunschweig, Germany
{t.blume, j.schattenberg, j.schiemann}@tu-braunschweig.de
KEY WORDS: Emergency, Exploration, UAV/UGV, Database management, Web application
ABSTRACT
Disasters such as floods, large fires, landslides, avalanches, or forest fires are often inevitable and cannot be fully prevented, but their
impact can be minimized with sound disaster management strategies aided by the latest technological advancements. A key factor
affecting these strategies is the time, where any delay can result in dramatic consequences and potentially human losses. Therefore, a
quick situation report of the disaster is highly demanded, but still not an easy task because - in most cases - a priori known spatial
information like map data or geo-databases, are outdated. In addition, visual and geometric information on the current situation is
needed to help rescue teams and first responders. From this point of view, we came up to the main idea of the joint research project
ANKommEn and its extension ANKommEn 2 (german acronym for Automated Navigation and Communication for Exploration).
The project idea embodies an exploratory investigation to be smart in providing correct and timely geodata that can help in
emergency cases; especially in support decision making in emergency risk management. For this purpose, automated unmanned
systems, both ground (UGV) and airborne (UAV), are being developed to provide up-to-date information of rescue scenarios.
1. ANKOMMEN - SYSTEM DESIGN ANKommEn-system using different data links, “Nav-Com-

Stack” shown in Figure 3. Using a proprietary connector at the
The overall system for data recording consists of Unmanned bottom of the UAV different modular payloads carrying visual
Aerial Vehicle (UAV), Unmanned Ground Vehicle (UGV) and sensors for exploration and mapping tasks are mounted at the
a central ground and control station. The ground station serves UAV. The GNSS antenna is mounted on the top right, antennas
as a central human-machine interface to monitor and manage for data exchange using WLAN and 4G mobile data link are
operation of the UAV and UGV by an operator. The system mounted at the landing gear and bottom side of the upper
enables an exploration process of an area of interest based on payload compartment.
initial geodata available like a priori known map. Exploration
areas and tasks are updated with current information of the
visual sensors such as combined RGB and IR camera within
one UAV-payload and LiDAR while performing a mission.
This is done by a data streaming to the ground station via data
link functionalities. For data capture and preprocessing where
applicable, a processing unit is integrated onboard the UAV/
UGV. Georeferencing of the exploration data is done using a
custom built-up GNSS/IMU system.
Project partners of the joint research projects ANKommEn and
ANKommEn 2 are Institute of Flight Guidance (IFF), Institute
of Mobile Machines and Commercial Vehicles (IMN) and
Institute for Geodesy and Photogrammetry (IGP), all
Technische Universitaet Braunschweig, and AirRobot GmbH &
Co. KG. For supporting the development and field tests,
associated project partners are included in the project
consortium. These are the Fire Brigade of Braunschweig and the Figure 1: UAV AR200 (customized for ANKommEn)
Lower Saxony Water Management, Coastal Defence and Nature
Conservation Agency (NLWKN). For ground based exploration tasks an UGV of type Robotnik
The UAV and UGV used for carrying the built up sensor sets, Summit XL is used for carrying the NAV-Com-Stack (mounted
data acquisition and processing hardware are depicted in Figure inside the chassis), GNSS and data link antennas as well as
1 and Figure 2. As UAV a AR200 hexacopter manufactured by visual sensors for exploration tasks, Figure 2.
AirRobot has been customized concerning the upper payload
compartment (black box on the front in Figure 1) for integration
the basic hardware setup for positioning (GNSS/INS), data
acquisition, processing and data exchange within the
This contribution has been peer-reviewed.

https://doi.org/10.5194/isprs-archives-XLII-2-W13-165-2019 | © Authors 2019. CC BY 4.0 License. 165
On the left the combined RGB-IR camera payload is depicted,

carrying and stabilizing a RGB camera (Allied Vison Manta G-
917C) as well as a thermal camera (Flir A65) for simultaneous
exploration within visual and IR spectrum.
For mission and task definition, controlling the UAV/UGV

operations as well as exploration data processing and
visualization a central ground and control station (GCS) has
been designed and built up, shown in Figure 5.
Figure 2: UGV Robotnik Summit XL carrying ANKommEn

payload
The configuration shown in Figure 2 is equipped with a LiDAR

(Velodyne VLP-16), a thermal camera (Flir A65) and a Pan-
Tilt-Zoom (PTZ) camera integrated in the front of the chassis.
The basic NAV-COM-Stack is depicted in Figure 3.
Figure 5: Central Ground and Control Station (GCS)
Therefore an industrial computer, GNSS and a communication

board are integrated within the GCS setup serving as HMI for
the overall system. Using the proprietary control software,
different missions and tasks can be defined by an operator and
assigned to different vehicles. In order to get exploration data
and status information of the UAV or UGV a second mobile
GCS has been built up for supporting in-site operators, shown in
Figure 6 using the same hardware setup for communication and
GNSS based positioning.
Figure 3: Nav-Com-Stack
The lower tier of the stack consists of a breakout board carrying

an IMU (Inertial measurement Unit, type ADIS 16488) and
GNSS receiver of type ublox M8T, built up by the IFF. These
are used for gathering position and attitude information of the
vehicles as well as provide timing information for sensor data
acquisition and georeferencing. As middle tier a processing
board of type Intel NUC7i5 is used for sensor data acquisition
and (pre-)processing, where applicable. The upper tier carries
the communication board of type Ventana GW5520 with
integrated modules for WLAN and mobile 4G data links to
ensure data exchange with other UAV/UGV operating
simultaneous and the ground and control stations. The built-up Figure 6: Mobile Ground and Control Station
visual sensor payload modules carried by the UAV are shown in
Figure 4.
2. SYSTEM COMMUNICATION
For fast and reliable data exchange, an intelligent

communication structure has been developed. The system’s
structure is illustrated in Figure 7.
Agents within range to each other are able to communicate
directly using ad-hoc WIFI, which offers high throughput (e.g.
the UGV and UAV on the bottom right of Figure 7). In case of
long distance communication, it’s even possible to
communicate without a direct link, utilizing data hopping over
arbitrary agents, e.g. the ground station communicates with the
UAV, utilizing the UGV as a range extender (bottom of Figure
7). In addition to WIFI, all agents are equipped with 3G/4G
modems, which enable communication via internet.
Figure 4: Combined RGB-IR-Payload (left) and LiDAR-
Communication via internet is encrypted and secured using a
payload (right)
central VPN server.

relief displacement induced by the terrain and perspective

model is corrected.
 In the last stage all incremental images are finally integrated
into a global mosaic and blended to correct for any
roughness.
Figure 8 shows the architecture of the library. As described

above, the different steps represent the various stages which
make use of the underlying library layer for all its function. The
transport layer is implemented in the ROS framework. More
details about the project can be found from the Github page
(Kern, 2018).
Figure 7: Flexible communication structure between arbitrary

agents using decentral ad-hoc networking and internet
The system adaptively aggregates multiple physical

communication channels between agents, when available.
Furthermore different compression algorithms were
implemented, to reduce bandwidth of data streams, which
allows delivering more data. Additionally data channels can be
prioritized by the user. This enhances flexibility and ensures Figure 8: The architecture of the OpenREALM
that data is delivered w.r.t. the user’s preferences e.g. delivering
live video feed, which is crucial for controlling the agent, in real
time with superior priority, while less sensitive data with The described SLAM-based solution relies on constant features
inferior priority is buffered where required and delivered when which might not be present in some scenarios (i.e. water, sand,
free capacity is available within the network (Quality of snow). Therefore, a fallback solution using only GNSS location
Service). and heading information was also implemented. This allows the
For convenient data access and storage, a central database system to generate a map even if there are not enough features.
server is integrated within the VPN, which delivers all gathered Even though the consistency and accuracy of the map obtained
data to arbitrary clients. with the fallback solution is affected badly, it still provides a
reasonable view of the mapped area. More can be seen in the
Figure 9, where the various options from this solution and the
3. MAPPING BASED ON PHOTOGRAMMETRY AND conventional photogrammetry are compared. Depending on the
SLAM resources available on the onboard PC i.e. CUDA compatible
GPU, the OpenREALM can be executed with or without 3D
Mapping of the desired area is done with two different methods reconstruction.
having complementary strengths. A SLAM based approach is
used to provide a real time ortho map of the scene to the
operators even during the flight while conventional
photogrammetry is used to generate a high-quality ortho map of
the area after the mission is finished. In the following first the
SLAM-based solution is described. The algorithm aims at
generating a high quality orthophoto and textured 3D meshes in
real-time. The proposed solution is published as an open source
library called OpenREALM (Kern, 2018) on GitHub. The
basic flow of the library works as follows:
 Geotagged images are first passed to a pose estimation

stage. There, an interface class allows integrating state of
the art visual SLAM frameworks for motion recovery
(currently only ORB_SLAM2 (Mur-Artal, 2015) is used but
can be extended easily with other methods).
 Extracted camera poses and the sparse cloud in the local
coordinate system are subsequently transformed into a
Cartesian frame using a best-fit Umeyama algorithm
(Umeyama, 1991).
 Densification of the tie points is applied next. External 3D
reconstruction frameworks can be integrated into the
mapping process to provide high quality, dense surface
information i.e. Plane Sweep Library (Haene, 2014).
Figure 9: A comparison of results achieved using visual
 Afterwards the rectification of the image data is carried out.
mapping approaches
By using the linear pinhole model, the prior extracted
elevation grid map is re-projected into the camera so the

The other solution uses an offline photogrammetry pipeline 4. POINT CLOUD PROCESSING AND
based on Agisoft Photoscan's Python API (Agisoft, 2019) was CLASSIFICATION
implemented to enable high quality results in variety of formats.
The process is automatically triggered after the last image of a Point cloud data is collected using Velodyne VLP 16 sensors
survey is saved by the drone. The method first aligns all the (see Figures 2 and 4). On both, ground and aerial vehicles, the
available images using the features from the individual image. data processing is done onboard (see Figure 3).
The GPS info available in the images stored as ExifTags is First step in data processing pipeline is registration of the actual
further used to aid in the alignment process. Once the images sensor data within a common frame of reference. In a first step
are aligned, a dense cloud is generated. The images are ortho- the actual reading from the sensor is pre-aligned using
rectified and an ortho-mosaic is generated. The process exports GNSS/IMU. Since the obtained position and orientation
the results in various different formats including the dense cloud (hereafter a combination of translation and orientation is
and ortho map which is later used in the web services to referred to as pose) may be corrupted due to noise or shading
visualize (Figure 10). effects, a fine registration step is applied afterwards.
Incoming scans are matched to a local reference map. An earth
fixed reference is obtained using Graph-SLAM approach, which
relates a locally constructed pose graph with a global one
(Schmiemann, 2017). Interconnecting both, mainly based on
time information, and minimizing the residual in least mean
squares sense, leads to an optimal alignment and eventual a
reference pose can be obtained, which transforms the local
Pose-Graph into a common frame of reference (e.g. UTM).
The mapping approach illustrated above, works well for single
vehicles, but tends to non-consistency, if multiple maps are
generated by independent agents. To generate more consistent
maps, matching algorithms were developed and implemented.
Main challenges in this domain include handling of unorganized
point cloud data, the huge amount of points per map and
generation of meaningful features to match with little overlap
and huge difference in perspective.
To achieve accurate matching performance, semi supervised
learning algorithms were utilized. Feature learning is performed
directly on point clouds using PointNet (Qi, 2016), which
originally was designed for classification tasks. The structure
can be subdivided into two components.
For each point within the map a representation is learned, which
then is pooled, to derive a global feature for the complete map.
Figure 10: The relative integrity of the maps generated from Those map-wise features are then fed to a classifier, which can
SLAM in reference to the PhotoScan after registration be trained e.g. to approximate a probability distribution over
class affiliation.
As can be seen from the Figure 9 qualitatively, the ortho-map We use the feature abstraction layer only and train with triplet
from the Photoscan is the most consistent in comparison to the loss, which aims to minimize feature distance to positive
SLAM based solution. This advantage comes at the cost of matches whilst maximizing feature distance to negative matches
computational resources required to for the processing. SLAM (Arandjelovic, 2015). A match can then be found by querying
solution which uses the 3D reconstruction on the other hand is the nearest neighbors within a given set of database maps and
slightly worse but has the advantage that the map is available thresholding. Figure 11 exemplary shows four different maps of
for the user to see in real time which can be critical in some overlapping areas, which shall be merged, whereby color
scenarios. The relative integrity of the Slam based solution can encodes height.
be quantified by taking Agisoft ortho-map as the ground truth.
The deviations of all the object locations from Figure 10 are
then averaged and calculated for all the methods. The results are a) light red b) yellow
presented in Table 1. It’s clear that the planar assumption results
in the highest error and when the elevation is considered by
taking into account the sparse or dense cloud, the relative
integrity is much better.
c) dark red d) purple

Relative Error (m)
Approach Assumption
Average Std.dev.
GNSS + Heading Planar 2.60 1.60
Visual SLAM Planar 5.10 6.30
Sparse Elevated 0.86 0.74
Interpolation
3D- Elevated 0.89 0.85
Reconstrauction Figure 11: Four maps generated by different agents in urban
scenario
Table 1: Relative orthophoto integrity errors for all approaches

Figure 12 provides bird’s-eye view of all four maps (see Figure

11), merged within a common frame of reference in overlay
with satellite imagery. Additionally the trajectories of the agents
collecting the map data is drawn in light red, yellow, dark red
and purple respectively.
Figure 12: Merged maps of Figure 11 overlaid with satellite

imagery. Trajectories of agents are drawn in color. Trajectory
for map (a) is shown in light red, trajectory for map (b) is shown
in yellow, trajectory for map (c) is shown in dark red and
trajectory for map (d) is shown in purple.
Further data processing is done to increase the Figure 13: Multiple views of the same map (top left / light
comprehensibility of generated maps. Therefore the raw point red map in Figure 11). Top: Feature response with red
cloud data, which only contains positional information and an channel representing linearity, green channel represents
additional channel for intensity, shall be enriched with sematic planarity and blue channel represents verticality. Center:
information. The enrichment process can be subdivided into Result of the clustering process, whereby clusters are colored
three crucial tasks, which are abstraction of geometrical randomly. Bottom: Prediction result for different classes.
features, clustering and predicting semantics. Red: Building, Light Green: High Vegetation, Dark Green:
Natural Terrain, Yellow: Low Vegetation, Gray: Man-Made
For each point, a covariance matrix of the local neighborhood is Terrain, Purple: Artefacts
generated. The ratios of the dominant eigenvalues indicate
spatial relations, e.g. if one eigenvalue is orders of magnitude
greater than all the others, one can assume a one dimensional 5. GEODATABASE MANAGEMENT AND GEODATA
neighborhood in which all points form a somewhat linear shape. DEPLOYMENT
On the other hand e.g. if all three dominant eigenvalues are
equally ranged, one can assume the local neighborhood to form 5.1. Geodatabase: concept and workflow
some kind of ellipsoidal shape (Guinard, 2017).
In every moment during emergencies, the situation can rapidly
With those feature abstractions, clustering algorithms can be change and therefore a large amount of geo-datasets containing
used, to group geometrically homogenous areas. Results of the great number of files and object data should be respected, stored
clustering process, using algorithms introduced by (Landrieu, and analyzed with such a multi sensor – multi platform system
2016), are shown in the center of Figure 13. One can see that as implemented in the research project. In such environment, a
the clustering process produces geometrically homogeneous key challenge is to provide the needed spatial and also non-
areas on different scales, e.g. complete facades of buildings spatial data in an easy way based on accurate and simple
containing multiple ten thousands of points are grouped to a solutions for an identification and retrieval of, often distributed
single cluster, whilst small groups of less than hundred points, and heterogeneous, geodata. From this aspect, thinking in a
e.g. representing a bush, are grouped in single clusters, as well. creation of geodatabases is essential, but the design of a
The result of the last step is shown in the bottom of Figure 13. geodatabase is a critical stage in ensuring its effectiveness and
With the clusters as nodes and their geometric relations as proficiency. If a geodatabase concept is designed incorrectly,
edges, a graph is constructed. Using graph convolutions the data provision will not function to its full potential,
(Landrieu, 2017), a model is trained to recognize structures and essentially creating problems for the users. Therefore we
classify them to semantic classes respectively. concentrate on the design of geodatabases with respect to
pertaining technical functionalities, as well as with a focusing

on Postgres1 and PostGIS2. The potential is that Postgres

enables on the one hand an easy access of the data available and
on the other hand there is an efficient compatibility with other
software that deals with data processing, retrieving and viewing
like QGIS, Geoserver, etc. Moreover, Postgres comes with a
quite liberal licence policy. PostGIS is a spatial database
extender for Postgres object-relational database. It adds support
for geographic objects allowing location queries to be run in
SQL.
Within this project, the main datasets include images and point
clouds. From a logical point of view, and to be smart in the data
provision, a classification of the datasets based on data type has
been achieved. For example, RGB images, thermal photos,
orthophotos, DSM, UAV’s/UGV’s Lidar data, etc. Each type
should be stored in a specific folder labelled with the
corresponding ID of the actual mission (Figure 14). One can
store a dataset within Postgres/PostGIS in a spatial database
allowing a smart and easy data access and retrieval, but storing
images directly as raster layer in a PostGIS-database is complex
and still not easy task because the Postgres/PostGIS concept
does not support this issue yet. To overcome this challenge, a
Python based mechanism has been developed to store the
geographic location of the image center points and other image
metadata into Postgres database, and linking them to the
original images. The mentioned mechanism can be structured in
three stages as follow (Figure 14):
 CSV-files related to images will be created in the first stage.

These files include metadata of images like: image name,
the geographic coordinates of the image center points, etc.
 The second stage is CSV-files conversion into shape files
that to be implemented as layers in QGIS. Figure 14: Storage process of images and the derived data in
Postgres/PostGIS and viewing in QGIS
 Finally, new database tables (sql-tables) will be created
based on the shape files (converted in the second step) and
loaded into the target database where the sql-table should be
stored. These sql-tables have a specific spatial index enables
an improved performance in terms of data archiving and
retrieving.
Based on a potential linkage between Postgres and QGIS

(Figure 15, left), the images of interest can be retrieved and
viewed. In QGIS, one can connect a specific database and select
the table of interest (namely: sql-table in Postgres database) and
import the corresponding shape file as layer showing the main
image points (Figure 15, middle). Each point has its own image
linkage based on the path to dataset, and therefore an image
view in QGIS is possible using the “Pop-up windows” through
a click event on the point of interest (Figure 15, right). At the
same time, orthophotos and DSMs can be directly imported and
viewed as georeferenced raster layers in QGIS independent Figure 15: QGIS-Postgres/PostGIS connection (left), main
from pop-ups functionality. The advantage is that the users and image points as shape layer (middle) and an image viewer with
operators can edit and process the orthophotos and DSMs in an its metadata like image name, latitude, longitude, etc. (right).
easy way.
Concerning on the point clouds provided from ANKommEn

capturing system, on the one hand from sensors mounted on
UAV/UGV system, and on the other hand based on
photogrammetric processing using RGB images, we classify the
point clouds available based on sensor type used as well as
labelled with flight mission ID. For e.g. UAV-Point Cloud,
UGV-Point Cloud and RGB-Point Cloud. In the same manner
for image data storage into Postgres databases, point clouds
1
https://www.postgresql.org available can be also imported into a target database using a
2
https://postgis.net/ python based script (Figure 16).

Here, the local geodata content (images, point clouds, etc.) is

implemented into a remote web map service for end users. The
images and the derived data such as orthophotos, DSMs, etc. are
published based on Geoserver as georeferenced raster layers.
These layers could be retrieved using their own Uniform Resource
Locator (URLs) that to be used for the creation of friendly
interactive maps. For this purpose, Leaflet is the leading open-
source JavaScript library that enables creating interactive maps
using Hypertext Markup Language Mark-up Language (HTML)
standards and Cascading Style Sheets (CCS). The Figure (18)
shows a leaflet based web map including an orthophoto
projected into basemap “Open Street Map”.
Figure 16: Storage process of point clouds in Postgres/PostGIS
5.2. Data deployment
Once having a complete geodatabase, it should address needs

for data provision to end users in an easy way like web
application that enables users/operators getting their work done
without having to learn a lot of skills or technical issues.
Therefore, in this section we introduce an initial framework for
geodata deployment and geo-web based application.
Beside databases, geo-web applications need other components Figure 18: A leaflet based web map created for the orthophoto
like: a geoserver for data hosting, a digital basemap to provide of the studied area. The orthophoto is published using
the geographical contexts, operational layers for editing layers, Geoserver and defined as a layer with its own URL, which is
queries, etc. The Figure (17) illustrates the principle of used in leaflet based map.
distributing geospatial information stored in ANKommEn
databases.
In contrast, the point clouds are converted directly into the html
file format using Potree Converter1 and can be retrieved as
layers using their URLs (Figure 19).
Figure 19: 3D rendering of large point cloud generated of the

studied area using Potree Converter Tools
1
Potree is a free open-source WebGL (Web Graphic Library) based
Figure 17: Principle of distributing geospatial information point cloud renderer for large point clouds, developed at the Institute of
Computer Graphics and Algorithms, TU Wien.

6. CONCLUSION of the 19th International Conference on Artificial Intelligence

and Statistics, in PMLR 51:1384-1393.
Within this paper, the current work accomplished in the joint
research project ANKommEn and its successor ANKommEn 2 Landrieu, L. & Simonovsky, M., 2017. Large-scale Point Cloud
has been introduced. The project idea embodies an exploratory Semantic Segmentation with Superpoint Graphs.. CoRR,
investigation to provide correct and timely geodata that can help abs/1711.09869.
in emergency cases; especially in support decision making in
emergency risk management. Starting by description of the Mur-Artal, R., Montiel, J. M. M. and Tardós, J. D., 2015.
system design and its components, then an intelligent ORB-SLAM: A Versatile and Accurate Monocular SLAM
communication structure implemented in the system has been System. IEEE Transactions on Robotics, vol. 31, no. 5, pp.
addressed. Furthermore, two mapping methods of the desired 1147-1163.
area are used, SLAM and photogrammetry based methods.
Finally, a geodatabase design has been presented so that all of Qi, C. R., Su, H., Mo, K. & Guibas, L. J., 2016. PointNet: Deep
datasets available such as images, point clouds, etc. can be Learning on Point Sets for 3D Classification and
connected and retrieved. A future task is to suggest additional Segmentation (cite arxiv:1612.00593)
ideas i.e. future design considerations that can help improving
system communication, as well as to enhance the geodatabase Schmiemann, J., Harms, H., Schattenberg, J., Becker, M.,
model designed; especially in the terms of advanced Batzdorfer, S., and Frerichs, L., 2017. A DISTRIBUTED
relationships between geospatial data and database tables. ONLINE 3D-LIDAR MAPPING SYSTEM,” in ISPRS
Archives of the Photogrammetry, Remote Sensing and Spatial
ACKNOWLEDGEMENTS Information Sciences.
The presented work was done within the joint research projects Umeyama, U., 1991. Least-Squares Estimation of
ANKommEn and ANKommEn 2 funded by the German Federal Transformation Parameters Between Two Point Patterns”. In:
Ministry of Economic Affairs and Energy administrated by the IEEE Trans. Pattern Anal. Mach. Intell. 13.4 (Apr. 1991), pp.
Space Administration of the DLR (ANKommEn - funding code: 376–380. issn: 0162-8828. doi: 10.1109/34.88573.
50NA1518, ANKommEn 2 - funding code 50NA1807).
REFERENCES
Agisoft Photoscan Python API Reference 1.5.0,

https://www.agisoft.com/pdf/metashape_python_api_1_5_0.pdf
Arandjelovic, R., Gronát, P., Torii, A., Pajdla, T. & Sivic, J.

2015. NetVLAD: CNN architecture for weakly supervised place
recognition.. CoRR, abs/1511.07247
Guinard, S., and Landrieu, L., 2017. Weakly supervised

segmentation-aided classification of urban scenes from 3d
LiDAR point clouds.
Haene, C., Heng, L., Lee, G. H., Sizov, A. and Pollefey, M.

2014. Real-Time Direct Dense Matching on Fisheye Images
Using Plane-Sweeping Stereo.” In: 3DV. IEEE Computer
Society, pp. 57–64. isbn: 978-1-4799-7000-1.
Kern, A., 2018. Real-time Photogrammetry using monocular

SLAM for Unmanned Aerial Vehicles; Master thesis at Institute
of Flight Guidance, TU Braunschweig, Source:
https://github.com/laxnpander/OpenREALM
Landrieu, L. & Obozinski, G., 2016. Cut Pursuit: Fast

Algorithms to Learn Piecewise Constant Functions. Proceedings

Copyright of International Archives of the Photogrammetry, Remote Sensing & Spatial
Information Sciences is the property of Copernicus Gesellschaft mbH and its content may not
be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for
individual use.

Uav Ugv 2

Uploaded by

Copyright:

Available Formats

Uav Ugv 2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Uav Ugv 2

Uploaded by

Copyright:

Available Formats

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-2/W13, 2019

THE JOINT RESEARCH PROJECT ANKommEn – EXPLORATION USING

A. Alamouri1, , M. Gerke1, S. Batzdorfer2, M. Becker2, U. Bestmann2, M. Bobbe2, Y. Khedar2, T. Blume3, J. Schattenberg3,

KEY WORDS: Emergency, Exploration, UAV/UGV, Database management, Web application

1. ANKOMMEN - SYSTEM DESIGN ANKommEn-system using different data links, “Nav-Com-

This contribution has been peer-reviewed.

On the left the combined RGB-IR camera payload is depicted,

For mission and task definition, controlling the UAV/UGV

Figure 2: UGV Robotnik Summit XL carrying ANKommEn

The configuration shown in Figure 2 is equipped with a LiDAR

Figure 5: Central Ground and Control Station (GCS)

Therefore an industrial computer, GNSS and a communication

The lower tier of the stack consists of a breakout board carrying

For fast and reliable data exchange, an intelligent

This contribution has been peer-reviewed.

relief displacement induced by the terrain and perspective

Figure 8 shows the architecture of the library. As described

Figure 7: Flexible communication structure between arbitrary

The system adaptively aggregates multiple physical

 Geotagged images are first passed to a pose estimation

This contribution has been peer-reviewed.

c) dark red d) purple

This contribution has been peer-reviewed.

Figure 12 provides bird’s-eye view of all four maps (see Figure

Figure 12: Merged maps of Figure 11 overlaid with satellite

This contribution has been peer-reviewed.

on Postgres1 and PostGIS2. The potential is that Postgres

 CSV-files related to images will be created in the first stage.

Based on a potential linkage between Postgres and QGIS

Concerning on the point clouds provided from ANKommEn

This contribution has been peer-reviewed.

Here, the local geodata content (images, point clouds, etc.) is

Figure 16: Storage process of point clouds in Postgres/PostGIS

5.2. Data deployment

Once having a complete geodatabase, it should address needs

Figure 19: 3D rendering of large point cloud generated of the

This contribution has been peer-reviewed.

6. CONCLUSION of the 19th International Conference on Artificial Intelligence

Agisoft Photoscan Python API Reference 1.5.0,

Arandjelovic, R., Gronát, P., Torii, A., Pajdla, T. & Sivic, J.

Guinard, S., and Landrieu, L., 2017. Weakly supervised

Haene, C., Heng, L., Lee, G. H., Sizov, A. and Pollefey, M.

Kern, A., 2018. Real-time Photogrammetry using monocular

Landrieu, L. & Obozinski, G., 2016. Cut Pursuit: Fast

This contribution has been peer-reviewed.

You might also like