Springe direkt zu Inhalt

2019

Zabel, Sophie: Detection and localization of bees located in honeycomb cells using convolutional neural networks

Betreuer
Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
30.10.2019
Sprache
eng

Honey bees are of particular interest to research in computer assisted behavioral analysis because of their complex social behavior and the large number of individuals in a colony. An important part for such studies is the detection and localization of bees during their activities in the hive. This thesis specifically focuses on bees located in honeycomb cells. A ground truth dataset was created by manually marking bees located inside cells in images of  whole honeycombs. Detail images of the marked position and random, unmarked positions were respectively extracted as positive and negative examples. A convolutional neural network was designed and trained on this dataset to classify detail images on whether they show bees in cells or not. The quality of the training was evaluated and heatmaps were created to visualize prediction strength in the original images. Local maxima of the heatmaps were determined as the predicted positions of bees in honeycomb cells. Several modifications were made during the  work to increase the precision of the network predictions. Testing the classifier used on the test sets made it possible to achieve a detection rate of 1 while testing on detail images  and a detection rate of about 0.6 while testing on images of an entire honeycomb.

Casares, Fabian: Datenfusion von Kamera und LiDAR unter Verwendung von YOLO in Verkehrsszenarien

Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
17.09.2019

Die Wahrnehmung der Umgebung ist für autonomes Fahren eine fundamentale Eigenschaft. Mit Sensoren wie LiDAR (Light Detection and Ranging) ist es möglich, Objekte zu extrahieren und zu klassifizieren und so die bestmöglichen Entscheidungen zu treffen. Fehlerhafte Informationen können zu falschen Entscheidungen des autonomen Fahrzeuges führen, was fatale Folgen haben kann. Daher ist es sinnvoll, verschiedene Sensoren zu verwenden, um ein Ereignis mit unterschiedlichen Informationen zu bewerten und die Nachteile eines Sensors durch die eines anderen Sensors auszugleichen.

Doch wie für den Menschen kann es auch für Algorithmen schwierig sein, eine Situation nur anhand eines bestimmten Augenblicks zu bewerten. Das liegt meist daran, dass keine ausreichenden Informationen vorliegen, um mit hoher Sicherheit die richtige Entscheidung zu treffen. Daher bietet es sich an, eine Situation zu verfolgen und die gesammelten Informationen für eine Entscheidung des Fahrzeuges zu verwenden. Dafür ist es sinnvoll, die Bewegung von Objekten zu berücksichtigen, um die korrekte Extrahierung und Klassifizierung von Objekten zu gewährleisten.

In dieser Arbeit wird ein echtzeitfähiges System vorgestellt, das LiDAR, Kamera und Bewegungsdaten miteinander kombiniert mit dem Ziel, fehlerhafte Informationen der LiDAR Objekterkennung zu filtern. Dafür werden die LiDAR Objekte auf dem Kamerabild mit einer 2D Objekterkennung verglichen und über mehreren Frames verfolgt, um so mit größerer Sicherheit falsche Informationen zu löschen.

Doch auch das hier untersuchte Verfahren hat Schwächen, denn verschiedene Experimente haben gezeigt, dass der alleinige Vergleich  der Objekterkennungen im Kamerabild nicht ausreichend ist, um falsche Informationen fehlerfrei zu filtern. Die in dieser Arbeit entwickelte High-Level-Datenfusion kann jedoch leicht in Zukunft durch neue Funktionen erweitert werden.

Stastny, Julian: Towards solving the RoboFish Leadership Problem with Deep Reinforcement Learning

Betreuer
Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
12.09.2019
Sprache
eng

The RoboFish project investigates interactions of swarm agents by using a robot fish to interact with real fish. This work explores the use of Deep Reinforcement Learning to train a robot policy that is able to lead real agents. To this end, a simulation with handcrafted swarm agents as well as an appropriate observation and action space for the robot is set up. Then, Neural Networks with Convolutional layers and Long-Short Term Memory cells are trained to maximize reward via Proximal Policy Optimization. The reward function is defined such that it measures the ability to make agents follow the robot to different places in the environment. Domain Randomization is used on the simulation in order to obtain a policy that is robust against different agent dynamics. The method is evaluated on its ability to lead one agent in a randomized simulation where interestingly, the robot seems to display a form of meta learning by being able to dynamically adapt its policy by ad-hoc inference about the agent dynamics. Its success in randomized simulations provides  reason to be optimistic about a transfer from simulated to real environment in future work.

Zahoransky, Valeria: Modelling the US Constitution in HOL

Betreuer
Christoph Benzmüller, Jan von Plato (University of Helsinki)
Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
23.08.2019
Sprache
eng

The thesis aims at modelling parts of the US Constitution with higher order logic (HOL) in theorem prover ISABELLE/HOL in order to verify the possibility of a legal dictatorship in the USA. The basis for the argument is a notorious anecdote on how, at his US citizenship hearing, logican Kurt Gödel informed the judge that the US Constitution was in fact faulty and allowed for the erection of a constitutional dictatorship. We shall explore both the argument Gödel might have had in mind when saying this and a verified version of the supposed argument, modelled on the computer.

Before delving into the argument, we give a short overview on the tools used, including an introduction to Isabelle/HOL and the manner in which we are going to use it.

The ensuing section of this work is concerned with Gödel's supposed argument on the Constitution's shortcomings. This also encompasses a quick overview of the Constitution and a more detailed consideration of the articles most relevant to the argument.

After having laid a theoretical foundation, we will devise and implement a HOL model for the argument in the main part of this work. Being mindful of the technical restrictions, we shall choose a suitable logic embedded into Isabelle's HOL-language and map the relevant parts of the Constitution to their equivalents in the proposed logic. Havind succeeded in this, we shall prove that it is possible to build a dictatorship without violating the Constitution, thus verifying Gödel's argument. The main part concludes with a few remarks on what to avoid when modelling a concept with Isabelle/HOL.

The last section will present a few further problematic properties of the US Constitution in addition to the one modelled in the main part.  We then name a few questions not yet addressed and conclude the thesis.

For convenience, the terms "US Constitution" and "(the) Constitution shall be used interchangeably.

Wolf, Valentin: EwaldBlocks: Translating the Idea of Ewald Summation to Neural Networks for Global and Local Feature Extraction

Betreuer
Frank Noe, Tim Landgraf
Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
20.08.2019
Sprache
eng

This Bachelor's thesis introduces the EwaldBlock, a component for NeuralNetworks to efficiently extract local and global features from images. Inspired by the Ewald Summation, a method for efficient computation of long-range interactions in particle systems, EwaldBlocks split their input into high and low frequency components and process them separately. The high frequency part contains local features that can be captured well with standard convolutional layers. Global features remain in the low frequency part of the image. Convolutional layers are unsuited to capture them, as they can only process features as large as their kernel size. EwaldBlocks therefore do pointwise multiplication of the lowest frequency components with learned kernels in the Fourier space. By the circular convolutional theorem, this is an approximated convolution with global kernel size. EwaldBlocks are designed to be able to replace convolutional layers in existing Convolutional Neural Networks (CNN) architectures. Experiments show, that the EwaldBlock can improve performance of shallow architectures  and that computation in the spectral space could lead to models that are robust against noise perturbations and less biased towards local textures compared to CNNs.

Gette, Eduard: Automated ID-Tag Localization and Statistical Analysis of Foraging Behavior of a Beehive

Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
06.08.2019
Sprache
eng

The goal of this thesis is to gain insights from a given dataset into the foraging behavior of a beehive. The dataset is a collection of video recordings of bees exiting and entering a hive through a transparent pipe. Each bee is marked with an individual ID-tag. To this end, first an already existing ID-tag localizer, which was originally trained on data with different  illumination and background, is adapted to improve its detection performance on the new dataset. Next, the fine-tuned localizer is used in combination with the Beesbook tracking pipeline to extract exit and entry tracks for each individual bee from the raw video data. The results are presented, using descriptive statistics, and visualized through several plots. Finally, an algorithm to cluster bees into peer groups, based on the tracking data, is suggested.

Tiemens, Lucca: Computer-supported Exploration of a Categorical Axiomatization of Miroslav Benda's Modeloids

Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
19.06.2019
Sprache
eng

A modeloid, a certain equivalence relation with an operation inspired by Ehrenfeucht-Fraïssé games, formulated by Miroslav Benda is generalized first to an inverse semigroup and then to an inverse category. It is shown that a categorical modeloid on the category of a finite vocabulary can be used to plan an Ehrenfeucht-Fraïssé game. On the way the Wagner-Preston representation theorem and the Ehrenfeucht-Fraïssé theorem are proven. The whole work is supported by using computer-based theorem proving.

Meyer, Janis Jendrik: End-to-End Learning of Steering Wheel Angles for Autonomous Driving

Abgabedatum
23.05.2019
Sprache
eng

Autonomous driving has been an active area of research for many years, especially since the rise of deep learning for computer vision tasks. In the past several years, many different works used end-to-end learning in an attempt to deal with the complexity of the problem, relying on datasets of collected driving data to train their models.

This thesis presents an overview of some of these papers and highlights decisions, considerations and difficulties encountered when creating an autonomous driving model using end-to-end learning. A convolutional neural network is trained to predict steering wheel angles using a dataset of front view images of human driving collected by AutoNOMOS Labs. Different configutations and settings for dataset composition and network input and output are explored and compared. The models are evaluated on a validation subset and with the use of the VisiualBackProp algorithm, which visualizes regions of the input with the biggest influence on the network prediction.

The best performing model is able to follow the lane on highways and is shown to have learned to detect and recognize lane markings to be used in the decision-making process.

Szynal, Janek: Behavioral clues for forager transitions in a fully tracked honeybee colony

Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
11.04.2019
Sprache
eng

The BeesBook system is a state-of-the-art system for research into the honeybee (Apil Melifera). It collects data from cameras continuously observing a beehive. All individuals are marked by tags the system can decode. Its main advantages are extensibility and scale (being able to track an entire colony over its lifetime). In this work I attempt an investigation into the honeybee division of labor, focusing on the transition to foragers (the workers that survey the environment for food and transport it back to the hive). I attempt to conduct it in a way that makes any future examinations of following topics easier and I make recommendations for such methodology.

Rau, Franziska: Automatisierte Klassifikation digitalisierter Spiralzeichnungen zur Erkennung von Parkinson

Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
01.04.2019

Parkinson ist neben Alzheimer eine der häufigsten neurodegenerativen Erkrankungen,die einen erheblichen Teil der älteren Bevölkerung betrifft. Die Ursachen der Parkinson-Krankheit sind noch weitgehend unerforscht und selbst frühe Symptome können leicht unentdeckt bleiben. Eines der ersten Symptome ist Tremor in verschiedenen Gliedmaßen, insbesondere der Hände in Ruhelage. Dies hat zur Folge, dass sich auch die Handschrift verschlechtert, was durch Änderungen kinematischer Eigenschaften, wie Geschwindigkeit und Beschleunigung, gekennzeichnet ist.

Je nachdem, wie weit die Krankheit fortgeschritten ist, ist der Tremor mehr oder weniger stark und Abweichungen in Zeichnungen werden deutlicher. Aufgrund von Bradykinesien benötigen Menschen, die an der Parkinson-Krankheit leiden, länger um Zeichnungen zu erstellen.

Da in den meisten Fällen mit frühen Anzeichen einer Parkinson-Krankheit keine eindeutigen Diagnosen gestellt werden können, kann eine computergestützte Klassifikation hilfreich sein.

In dieser Arbeit benutze ich Tablet generierte und digitalisierte Daten von 77 Probanden, von denen 62 an Parkinson erkrankt sind. Die Probanden wurden gebeten verschiedene Tests zu absolvieren, in denen Spiralen nachgezeichnet werden sollten. Die aufgezeichneten Daten habe ich benutzt, um gesunde und an Parkinson erkrankte Teilnehmer zu klassifizieren, dazu habe ich verschiedene Attribute generiert, die auch Auskunft über bestimmte Symptome geben können. Anschließend habe ich eine Attributauswahl vorgenommen und dann ein Random Forest Modell pro Test erstellt.

Ich habe eine 5-fache Kreuzvalidierung vorgenommen, bei der die Klassen im Trainings- und Testdatensatz prozentual ungefähr gleich wie im Original Datensatz verteilt sind. Danach habe ich die Ergebnisse evaluiert und mit AUC-Werten von 0.88 und 0.95, sowie Cohen’s Kappa von 0.76 und 0.73 gute Ergebnisse erhalten.

Digitale Daten von Zeichnungen können derzeit keine eindeutige Diagnose ersetzen. Meine Untersuchungen haben jedoch ergeben, dass sie eine gute Möglichkeit bieten, Symptome einer Parkinson-Erkrankung zu erkennen und ärztliche Diagnosen zu unterstützen.

Hörner, Georg: Development and Prototypical Implementation of a Camera Based Body Pose Recognition Inside of a Vehicle

Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
08.02.2019
Sprache
eng

Driver distraction is a significant factor in the occurence of car crashes. Its detection is a topic of interest in the automotive industry. The purpose of driver distraction detection is to reduce risk of crashes by giving advanced warning of driver inattentiveness, enabling warning signals or automotive system actions.  Using in-vehicle cameras, the body pose of a driver can be determined. With that an estimation of attentiveness could be made by a software system.

In this thesis, a concept for a body pose estimation approach using a depth camera is described. Sensor choice and the subsequent development of a frame acquisition pipeline is laid out. A test video generation plan is introduced and its merits discussed. A static background filter algorithm is designed and implemented. Estimation of arm positions in point cloud data is explored using the sample consensus functionality provided by the Point Cloud Library. The methods for generating and finding parameters used for arm position estimation are presented. A measurement for the performance is presented and intermediate results are evaluated. A further iteration on paramteter generation is shown to improve the performance of the approach by narrowing down parameters. The final results are presented and factors leading to differences in the performance are discussed.

The thesis is concluded by a summary of the results and in an outlook further developments of the system are presented and discussed.

Göbel, Frederik: Simulation of Communication Models for Headway Minimisation in the Context of Car2Car Energy Sharing

Abschluss
Bachelor of Science (B.Sc.)
Abgabedatum
07.01.2019
Sprache
eng

With the advancement of electric and autonomous cars new ways to charge and distribute energy are researched. One method researched at the Biorobotics Lab of the FU Berlin is Car2Car energy sharing. For efficient and safe energy transmission a minimal headway with high stability between two cars needs to be guaranteed.

In the scope of the thesis Car2Car communication will be evaluated in comparison with sensor based information gathering. Two Car2Car communication models will be implemented. The first will utilise the communication to transmit data as a replacement for the sensor. In the second model the planned future behaviour of the leading car is shared with the following car. The second model will be evaluated with the sensor and the data transmitting model as fallback systems.