Filtern
Dokumenttyp
- Video (53)
- Konferenzveröffentlichung (4)
Volltext vorhanden
- ja (57) (entfernen)
Schlagworte
- Robotik (27)
- Flugkörper (18)
- UAV (18)
- 3D Modell (7)
- Rettungsrobotik (5)
- Rasenmäher (4)
- DRZ (3)
- Deutsches Rettungsrobotik-Zentrum (3)
- SLAM (3)
- Kartierung (2)
This technical report is about the mission and the experience gained during the reconnaissance of an industrial hall with hazardous substances after a major fire in Berlin. During this operation, only UAVs and cameras were used to obtain information about the site and the building. First, a geo-referenced 3D model of the building was created in order to plan the entry into the hall. Subsequently, the UAVs were used to fly in the heavily damaged interior and take pictures from inside of the hall. A 360° camera mounted under the UAV was used to collect images of the surrounding area especially from sections that were difficult to fly into. Since the collected data set contained similar images as well as blurred images, it was cleaned from non-optimal images using visual SLAM, bundle adjustment and blur detection so that a 3D model and overviews could be calculated. It was shown that the emergency services were not able to extract the necessary information from the 3D model. Therefore, an interactive panorama viewer with links to other 360° images was implemented where the links to the other images depends on the semi dense point cloud and located camera positions of the visual SLAM algorithm so that the emergency forces could view the surroundings.
This technical report is about the architecture and integration of very small commercial UAVs (< 40 cm diagonal) in indoor Search and Rescue missions. One UAV is manually controlled by only one single human operator delivering live video streams and image series for later 3D scene modelling and inspection. In order to assist the operator who has to simultaneously observe the environment and navigate through it we use multiple deep neural networks to provide guided autonomy, automatic object detection and classification and local 3D scene modelling. Our methods help to reduce the cognitive load of the operator. We describe a framework for quick integration of new methods from the field of Deep Learning, enabling for rapid evaluation in real scenarios, including the interaction of methods.
At the beginning of the pandemic in Feb. 2020 I had a little time and wanted to do something new i.e. bring my 3D printer, AI and computer science together somehow. The result is a printed portrait with a lot of computer science. Using style transfer I transferred the etching style of a Göthe portrait to a young girl I call Carolin. By means of image processing I made a black and white picture out of it. Then, using the problem of the traveling salesman, each black point in the picture is interpreted as a city and the whole picture is drawn by only one line. Since this line is very long, it is optimized and shortened by a so-called simulated annealing algorithm. The result is printed in 5 layers on a 3D printer.
Durch Panoramen in Kombination mit dem ORB-SLAM ist ein schnelles Tracking möglich, liefert jedoch ausschließlich spärliche Daten. Durch die Kombination mit einem neuronalen Netz soll der SLAM Algorithmus zu einem RGBD-SLAM erweitert werden, um ein besseres Tracking und eine dichtere Punktwolke zu gewährleisten.
360° Camera at a small UAV
(2021)
The video shows a very high resolution 3D point cloud !!! of the outdoor area of the German Rescue Robotics Center. For the recording, a 25-second POI flight was performed with a Mavic 3. From the 4K video footage captured during this flight, 77 images were cropped and localized within 4 minutes using colmap and processed using Neural Radiance Fields (NeRF). The nerfacto model of Nerfstudio was trained on an Nvidia RTX 4090 for 8 minutes. In summary, a top 3D model is available to task forces after about 13 minutes. The calculation is performed locally on site by the RobLW of the DRZ. The video shown here shows a free camera path rendered at 60 hz (Full HD).
Nerf(acto) for the 3D modeling of the Computer Science building of Westfälische Hochschule GE
(2023)
The video shows a very high resolution 3D point cloud !!! of the computer science building of the University of Applied Science Gelsenkirchen. For the recording a 3 minute flight with a M30T was performed. The 105 images taken by the wide-angle camera during this flight were localized within 3 minutes using colmap and processed using Neural Radiance Fields (NeRF). The nerfacto model of Nerfstudio was trained on an Nvidia RTX 4090 for 8 minutes. Thus, a top 3D model is available after about 15 minutes.
The video shown here shows a free camera path rendered at 60 hz (Full HD).
From the 360° images of the former video (
• German rescue robotic center captured... ) we now generate the 3D point cloud. The UAV needs 3 minutes to capture the outdoor scenario and the hall from inside and outside. The 3D point cloud generation is 5x slower than the video. It uses a VSLAM algorithm to localize the k-frames (green) and with 3 k-frames it use a 360° PatchMatch algorithm implemented at a NVIDIA graphic card (CUDA) to calculated the dense point clouds.The hall ist about 70 x 20 meters.
The video shows the first test of a small spherical UAV (35 cm) with 4 rotors for missions in complex environments such as buildings, caves or tunnels. The spherical design protects the vehicle's internal components and allows the UAV to roll over the ground when the environment allows. The drone can land and take off in any position and come into contact with objects without endangering the propellers and can restart even after crashes.