Aktuelle Trajektorienfolgeregelungskonzepte bestehend aus einer modellbasierten Vorsteuerung und einem modellfreien Regler ermöglichen bereits das automatisierte Fahren. Aufgrund von Ungenauigkeiten in den Systemmodellen und äußeren, sich ändernden Störeinflüssen werden zusätzlich vermehrt adaptive Regelungsverfahren eingesetzt, um eine kontinuierlich hohe Regelgüte zu garantieren.

Im Rahmen dieser Masterarbeit sollen verschiedene Reinforcement Learning Algorithmen bezüglich ihrer Eignung zur Erweiterung eines bestehenden Trajektorienfolgeregelungskonzeptes evaluiert werden. Mithilfe von künstlichen neuronalen Netzen werden sogenannte Agenten abgebildet, welche als additives Regelglied ein System bestehend aus Vorsteuerung und PID-Regler ergänzen sollen, um die Querdynamikregelung eines automatisierten Fahrzeugs zu optimieren.

In einer fahrzeugnahen Simulationsumgebung werden Fahrversuche durchgeführt, bei welchen Aganten sowohl mit verschiedenen Hyperparametern aus auch verschiedenen Reinforcement Learning Algorithmen trainiert werden. Im Rahmen dieser Untersuchungen konnte eine Methode gefunden werden, welche in Kombination mit dem bestehenden Regelungskonzept zu einer signifikanten Verbesserung des Fahrzeugverhaltens führt. Zur Bewertung der Regelung wird vorwiegend die Abweichung zwischen Soll- und Isttrajektorie betrachtet, welche deutlich reduziert werden konnte.

Die Entwicklung von Fahrerassistenzsystemen und automatisierten Fahrfunktionen erfordert zunehmend einen hohen Testaufwand. Der virtuelle Fahrversuch mithilfe der Simulation ergänzt und entlastet die reale Fahrerprobund, darf sich aber in Qualität und Aussagekraft nicht wesentlich vom realen Test auf der Versuchsstrecke oder im öffentlichen Straßenverkehr unterscheiden.

Mit steigenden Anforderungen müssen die virtuellen Teststrecken immer realistischer das wirkliche Straßenbild nachstellen. Dazu gehört auch, möglichst wirklichkeitsnahe Teststrecken zu generieren, die nicht nur einzelne Testfälle abbilden. Reale Straßen und ihre Markierungen entsprechen häufig nicht den Idealvorstellungen und variieren in ihren Eigenschaften. Grundvoraussetzung, eine Teststrecke aus Bilddaten realer Straßen zu generieren, sind akkurate Bilderkennung- und unterscheidung. Die klassische Spurerkennung ist hier nur bedingt hilfreich.

Gegenstand dieser Arbeit ist es, verschiedene Ansätze zum Erkennen von Spurmarkierungen auf Bilddaten zu analysieren und zu evaluieren. Die beste erreichbare und effizienteste Lösung wurde mittels einer semantischen Segmentierung, die die verschiedenen Arten der Spurmarkierung erkennt und pixelweise klassifiziert, implementiert. Mithilfe der Segnet-Architektur konnte ein IOU Score von 0.3169 und ein Dice Koeffizient von 0.4790 erreicht werden. Darüber hinaus wurden weitere unabdingbare Arbeitsschritte diskutiert, die erforderlich sind, um Teststrecken generieren zu können.

In this thesis a deep reinforcement learnng agent is trained in an environment made with the Unity game engine. The new ML Agents APl allows communication with a Python backend, which enables research with familiar tools. The simulation features a huge variety of different adjustable parameters and can be run in parallel. The agent successfully learned to follow the track and is robust to various environmental changes. Furthermore, the future deployment of Neural Networks on the AutoMiny car is prepared. The AutoMiny cars are developed by the FU Berlin (Institute of Computer Science). The model car features the NvidiaJetson Nano, which allows the hardware accelerated inference via Deep Neural Networks.

In the field of collective behavior group-level phenomena emerge from interactions between individuals. To study inter-individual rules the Landgraf lab has built a robotic guppy that replaces live anmials in the shoal, RoboFish. The primary purpose of this thesis is to examine if the pair interaction behavior of female guppies can be learned by recurrecnt neural networks via supervised learning and to develop the software components required to have the RoboFish system run the resulting models in the real world. Two distinct datasets are studied and RNN models trained to try to imitate the behavior seen in them: One dataset was synthetically generated from a simple deterministic model as a baseline and one was captured from live fish. Training different kinds of RNNs on the datasets revealed a capability of a simple stacked RNN to learn swarm behavior, further improved upon by putting a ConvLSTM input network in front of it. I was also successful in showing that a more complex network was able to learn some basic interaction behavior as seen in real fish.

Despite their small brain size, honeybees are able to orientate themselves exceptionally well in the environment. In order to study their navigational capabilities, correlations between brain activity of a bee and its visual perception are analysed. Neural activity of the animal was recorded while flying with a quadcopter. First, it is shown that similar brain activity was measured in repeated flight routes. Next, the flight of the copter is simulated in a virtual 3D environment of the experiment's fields and its surroundings. Images were rendered from the perspective of a bee and processed by an autoencoder. Activations of the encoder were correlated with the bee's neural activity. Additionally, the brain activity was projected onto a field in the range of a bee's field of view. Both approaches revealed high correlations during flight turns.

Model logic is a formalism suitable for modelling problems in artificial intelligence, computer science, philosophy and many other disciplines. Its automation in all its flavours can easily be achieved employing the semantic embedding approach and resuing existing reasoning software: This thesis presents a procedure that encodes a problem of higher-order modal logic in higher-order  logic which is implemented in the Model Embedding Tool for the input language TPTP THF which was extended for this purpose. By combining this software with one or more compliant reasoners a competitive system is created that can handle countless more variants of modal logic than any other theorem prover available to date.

Motion Planning in autonomous driving defines the task of planning the desired movement of a vehicle through a dynamically moving environment. A plan is stored as trajectory, saving spatial and temporal information about the future vehicle movement.  Path-Speed decomposition is a planning method for finding such a trajectory. A path is planned in a first step, followed by an according speed profile.

This aims to implement and evaluate a planner for finding a rough speed profile in a descretized search space. A graph is created and a single source shortest paths algorithm is used to find the optimal speed profile within the limited search space,  evaluated by cost functions representing the requirements of speed planning. The rough speed profile can serve as initial solution for numerical optimization, which is not part of this thesis.

The implemented approach is evaluated in sumulation of various urban traffic scenarios, showing promising collision free and low-jerk trajectories. It is able to find a speed profile in real-time. Therefore, the planner seems useful for practical application in an autonomous driving vehicle.

In this work, a Prediction model, using LSTMs and mixture density networks was trained to predict trajectories. The model is tested on simple models and then applied to bee trajectories. A simulation was implemented that can run different movement models. The simulation offers a visualization of the movement models. After the Prediction models were trained an analysis that made use of the hidden states of the model was done. Plotting T-SNE and UMAP projections revealed interesting clusters in the hidden states. Furthermore, a classification task was solved to see if the hidden states of the Prediction model  are able to boost classification performance. The results revealed that if the Prediction model is able to predict realistic trajectories, classification performance can be improved for problems where few labels are available. All the easy movement models were to some extend successfully learned by the Prediction model. The Prediction model was not able to predict realistic bee trajectories.

Artificial pacemakers  and implantable cardioverter-defibrillators (ICDs) provide life-sustaining therapies for those affected by cardiac diseases. With  the aim of supporting follow-up care modern implants transmit diagnostic data to the manufacturer. The data includes an electrocardiogram (ECG) and is made available to the attending physician. An additional monitoring by means of machine learning methods is largely unresearched but holds great potential considering the recent advances in medical data science.

In the first part of the thesis an infrastructure that enables research on the transmitted ECGs was developed. It allows the data to be accessed  language-independent and approximately 120 times faster compared to a formerly utilized approach. Secondly, two machine learning methods originally designed for surface ECGs were tested on intracardiac signals. The first classifier was based on manual feature engineering. The second employed a convolutional neural network (CNN). Both were used to distinguish between supraventricular tachycardia (SVT) and ventricular tachycardia (VT) addressing current issues in ICD therapy. Compared to conventional SVT/VT detection algorithms they received a significantly smaller part of the ECG. With an accuracy of 94.9% the CNN outperformed the manual feature engineering (86.0%) and equaled the performance of the classifier used to annotate the training data.

The tested classifiers indicate not only that research on surface ECGs is transferable to intracardiac signals but also showed the importance of the new infrastructure.

Stereo matching (SM) is a well researched method for generating depth information from camera images. Efficient implementations of SM algorithms exist as part of widely used computer vision libraries, such as OpenCV. Typically, SM is being performed on pairs of images from a stereo camera in which intrinsic and extrinsic parameters are fixed and determined in advance by calibration.  Furthermore, the images are usually taken at approximately the same time by triggering the shutters simultaneously. In this thesis a different approach is being pursued: stereo pairs are selected from a video sequence of a monocular  camera, which is mounted on a moving vehicle. Two scenarios are covered: one where the camera is facing sidewards and one where it is facing forwards in relation to the driving direction. Extrinsic transformatins between frames are computed by visual odometry. Images out of a series can each be rectified with the same reference image; the resulting image pairs are therefore effectually taken by a virtual stereo camera with variable baseline. Stereo matching and three-dimensional reconstruction can be applied to these images in the same way as to those of a binocular camera with fixed extrinsic calibration. Apart from the development of the virtual stereo principle itself, two main contributions have been developed in this thesis: Firstly, it has been shown that the fusion of disparity images (according to Hirschmüller) taken at varying baselines improves quality in terms of density and error rate. Secondly, a new rectification procedure has been developed for the scenario of the forward facing camera; here the standard procedure developed for conventional stereo cameras is not applicable.

The goal of this thesis was to create a HLA-based software framework for the creation of distributed simulations which will be used in animal behavior experiments conducted by the BioroboticsLab of the Freie Universität Berlin. For this target, specifications based on the expected use cases have been specified, coupled with a market analysis. This led to the decision to develop a custom software. The thesis describes the created software, the key elements of its architecture and its main features, as well as evaluating several performance aspects.

In dieser Arbeit wird untersucht, ob sich mit Hilfe maschinellen Lernens der Tod von Honigbienen vorhersagen lässt. Dazu werden verschiedene Merkmale des Verhaltens einer Biene definiert und für einen bestimmten Zeitraum vor dem Tod mit ihren durchschnittlichen Werten verglichen, bevor abschließend unter Verwendung des Random Forest Classifiers ein Modell trainiert und auf Testdaten angewendet wird. In beiden Fällen wird dabei die Relevanz der einzelnen Parameter untersucht um herauszufinden, welche Verhaltensweisen ein Indikator für den bevorstehenden Tod einer Biene sein können. Hierbei wird gezeigt, dass neben den erwarteten Einfluss des Alters der Bienen noch andere Parameter eine sogar größere Aussagekraft für die Vorhersage haben. Abschließend werden darauf basierend Schlussfolgerungen zum Sterben von Bienen aufgezeigt, sowie weitere daraus resultierende Ansätze formuliert.

The ultimative objective of this thesis is to tackle the lack of CNN-based detectors specialiazed on fisheye images. Thereby experiments are required, therefore we will compare the performance of several state-of-the-arts and optimize one of them. Along the way we construct a new object detection framework to perform our experiments as convenient as possible. Generally the entire work done in this thesis can be divided into two main parts as follows:

• Optimization of CNN-based detector for fisheye images.
• Construction of an object detection framework.