Since the invention of U-net architecture in 2015, convolutional networks based on its encoder-decoder approach significantly improved results in image analysis challenges. It has been proven that such architectures can also be successfully applied in different domains by winning numerous championships in recent years. Also, the transfer learning technique created an opportunity to push state-of-the-art benchmarks to a higher level. Using this approach is beneficial for the medical domain, as collecting datasets is generally a difficult and expensive process.
In this thesis, we address the task of semantic segmentation with Deep Learning and make three main contributions and release experimental results that have practical value for medical imaging.
First, we evaluate the performance of four neural network architectures on the dataset of the cervical spine MRI scans. Second, we use transfer learning from models trained on the Imagenet dataset and compare it to randomly initialized networks. Third, we evaluate models trained on the bias field corrected and raw MRI data. All code to reproduce results is publicly available online.
In current research of the autonomous mobile robots, path planning is still a very important issue.
This master's thesis deals with various path planning algorithms for the navigation of such mobile systems. This is not only to determine a collision-free trajectory from one point to another. The path should still be optimal and comply with all vehicle-given constraints. Especially the autonomous driving in an unknown and dynamic environment poses a major challenge, because a closed-loop control is necessary and thus a certain dynamic of the planner is demanded.
In this paper, two types of algorithms are presented. First, the path planner, based on A*, which is a common graph search algorithm: A*, Anytime Repairing A*, Lifelong Planning A*, D* Lite, Field D*, hybrid A*. Second, the algorithms which are based on the probabilistic planning algorithm Rapidly-exploring Random Tree (Rapidly-exploring Random Tree, RRT*, Lifelong Planning RRT*), as well as some extensions and heuristics. In addition, methods for collision avoidance and path smoothing are presented. Finally, these different algorithms are evaluated and compared with each other.
Mit der Microsoft Kinect waren die ersten Aufnahmen von synchronisierten Farb- und Tiefendaten (RGB-D) möglich, ohne hohe finanzielle Mittel aufwenden zu müssen und neue Möglichkeiten der Forschung eröffneten sich. Mit fortschreitender Technik sind auch mobile Endgeräte in der Lage, immer mehr zu leisten. Lenovo und Asus bieten die ersten kommerziell erwerblichen Geräte mit RGB D-Wahrnehmung an. Mit integrierten Funktionen der Lokalisierung, Umgebungserkennung und Tiefenwahrnehmung durch die Plattform Tango von Google gibt es bereits die ersten Tests in verschiedenen Bereichen des Rechnersehens z.B. Mapping. In dieser Arbeit wird betrachtet, inwiefern sich ein Tango Gerät für die Objekterkennung eignet. Aus den Ausgangsdaten des Tango Geräts werden RGB D-Daten extrahiert und für die Objekterkennung verarbeitet. Es wird ein Überblick über den aktuellen Stand der Forschung und gewisse Grundlagen bezüglich der Tango Plattform gegeben. Dabei werden existierende Ansätze und Methoden für eine Objekterkennung auf mobilen Endgeräten untersucht. Die Implementation der Erkennung wird anhand einer selbst erstellten Datenbank von RGB-D Bildern gelernt und getestet. Neben der Vorstellung der Ergebnisse werden Verbesserungen und Erweiterungen für die Erkennung vorgeschlagen.
Particle swarm optimization is an optimization technique based on simulation of the social behavior of swarms.
The goal of this thesis is to solve 6DOF local pose estimation using a modified particle swarm technique introduced by Khan et al. in 2010. Local pose estimation is achieved by using continuous depth and color data from a RGB-D sensor. Datasets are aquired from different camera poses and registered into a common model. Accuracy and computation time of the implementation is compared to state of the art algorithms and evaluated in different configurations.
The purpose of this master thesis is to enable the Robot Lisa to process complex commands and extract the necessary information in order to perform a complex task as a sequence of smaller tasks. This is intended to be achieved by the improvement of the understanding that Lisa has of her environment by adding semantics to the maps that she builds. The complex command itself will be expected to be already parsed. Therefore the way the input is processed to become a parsed command is out of the scope of this work. Maps that Lisa builds will be improved by the addition of semantic annotations that can include any kind of information that might be useful for the performance of generic tasks. This can include (but not necessarily limited to) hierarchical classifications of locations, objects and surfaces. The processing of the command in addition to some information of the environment shall trigger the performance of a sequence of actions. These actions are expected to be included in Lisa- currently implemented tasks and will rely on the currently existing modules that perform them.
Nevertheless the aim of this work is not only to be able to use currently implemented tasks in a more complex sequence of actions but also make it easier to add new tasks to the complex commands that Lisa can perform.
Pedestrian Detection in digital images is a task of huge importance for the development of automaticsystems and in improving the interaction of computer systems with their environment. The challenges such a system has to overcome are the high variance of the pedestrians to be recognized and the unstructured environment. For this thesis, a complete system for pedestrian detection was implemented according to a state of the art technique. A novel insight about precomputing the Color Self-Similarity accelerates the computations by a factor of four. The complete detection system is described and evaluated, and was published under an open source license.