Refine
Year of publication
- 2020 (50) (remove)
Document Type
- Article (21)
- Conference Proceeding (17)
- Working Paper (4)
- Book (2)
- Report (2)
- Part of a Book (1)
- Doctoral Thesis (1)
- Other (1)
- Part of Periodical (1)
Is part of the Bibliography
- no (50) (remove)
Keywords
- GIS (2)
- remote sensing (2)
- windthrow (2)
- Abrechnung von Bauleistungen (1)
- Alois Alajos Unger (1)
- Bauabwicklung (1)
- Bürgerfamilie (1)
- CNNs (1)
- CSR (1)
- Controller (1)
Camera Re-Localization with Data Augmentation by Image Rendering and Image-to-Image Translation
(2020)
Self-localization of cars, robots or Unmanned Aerial Vehicles (UAVs) as well as self-localization of pedestrians is and will be of high interest for a wide range of applications.
A major task is autonomous navigation of vehicles, whereas the localization in the surrounding scene is a key component.
Since cameras are well-established built-in sensors in cars, robots and UAVs, there is little to no extra cost in utilizing them for subsequent localization.
The same applies for pedestrian localization, where smartphones serve as mobile platforms for cameras.
Camera re-localization, where the pose of a camera is determined with respect to a certain scene, is therefore a valuable process to solve or support localization solutions of such vehicles or pedestrians.
Cameras are low-cost sensors that are established commonly in the everyday life of humans and machines.
The support of camera re-localization is not limited to applications related to navigation but can in general be applied to support image analysis or image processing tasks as scene reconstruction, detection, classification or alike.
For these purposes, this thesis concerns the improvement of the camera re-localization task.
As Convolutional Neural Networks (CNNs) and hybrid pipelines to regress camera poses are recently competing against established hand-crafted designed pipelines reaching similar or superior accuracies, the focus is set on the former in this thesis.
The main contributions of this thesis include the design of a CNN to regress camera poses, with focus on a lightweight architecture fitting the requirements to be applicable on mobile platforms.
This network achieves accuracies in the same order as CNNs with larger model sizes.
Furthermore, the performance of CNNs is highly depending on the quantity and quality of training data utilized for their optimization.
Hence, further contributions are considering image rendering and image-to-image translation to extend such training data in terms of Data Augmentation (DA).
3D models are utilized for image rendering to valuable extend training datasets.
Generative Adversarial Networks (GANs) serve for DA by image-to-image translation. Whereas image rendering is increasing the quantity of images in datasets, image-to-image translation on the other hand aims to enhance the quality of such data.
Experiments are carried out on datasets augmented by image rendering and image-to-image translation. It is shown, that both approaches valuable enhance the localization concerning accuracy.
Therefore, state-of-the-art localization is improved by DA in this thesis
One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for the data generations. We propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), which learns to differentiate between the individual mixture components and therefore allows for generations from the distributional data clusters. We assume a two-level generative process with a continuous (Gaussian) latent variable sampled conditionally on a discrete (categorical) latent component. The new variational objective naturally couples the learning of the posterior and prior conditionals, and the learning of the latent categories encoding the multimodality of the original data in an unsupervised manner. The data-dependent conditional priors are then used to sample the continuous latent code when generating new samples from the individual mixture components corresponding to the multimodal structure of the original data. Our experimental results illustrate the generative performance of our new model comparing to multiple baselines.
One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for data generations. In this paper, we propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), with a two-level generative process for the observed data where continuous 𝐳 and a discrete 𝐜 variables are introduced in addition to the observed variables 𝐱. By learning data-dependent conditional priors, the new variational objective naturally encourages a better match between the posterior and prior conditionals, and the learning of the latent categories encoding the major source of variation of the original data in an unsupervised manner. Through sampling continuous latent code from the data-dependent conditional priors, we are able to generate new samples from the individual mixture components corresponding, to the multimodal structure over the original data. Moreover, we unify and analyse our objective under different independence assumptions for the joint distribution of the continuous and discrete latent variables. We provide an empirical evaluation on one synthetic dataset and three image datasets, FashionMNIST, MNIST, and Omniglot, illustrating the generative performance of our new model comparing to multiple baselines.
Lifelong generative modeling
(2020)
Lifelong learning is the problem of learning multiple consecutive tasks in a sequential manner, where knowledge gained from previous tasks is retained and used to aid future learning over the lifetime of the learner. It is essential towards the development of intelligent machines that can adapt to their surroundings. In this work we focus on a lifelong learning approach to unsupervised generative modeling, where we continuously incorporate newly observed distributions into a learned model. We do so through a student-teacher Variational Autoencoder architecture which allows us to learn and preserve all the distributions seen so far, without the need to retain the past data nor the past models. Through the introduction of a novel cross-model regularizer, inspired by a Bayesian update rule, the student model leverages the information learned by the teacher, which acts as a probabilistic knowledge store. The regularizer reduces the effect of catastrophic interference that appears when we learn over sequences of distributions. We validate our model’s performance on sequential variants of MNIST, FashionMNIST, PermutedMNIST, SVHN and Celeb-A and demonstrate that our model mitigates the effects of catastrophic interference faced by neural networks in sequential learning scenarios.
Protocol for the management and the monitoring of the measurements for the COVID19 containment
(2020)
Forest damage due to storms causes economic loss and requires a fast response to prevent further damage such as bark beetle infestations. By using Convolutional Neural Networks (CNNs) in conjunction with a GIS, we aim at completely streamlining the detection and mapping process for forest agencies. We developed and tested different CNNs for rapid windthrow detection based on PlanetScope satellite data and high-resolution aerial image data. Depending on the meteorological situation after the storm, PlanetScope data might be rapidly available due to its high temporal resolution, while the acquisition of high-resolution airborne data often takes weeks to a month and is, therefore, used in a second step for more detailed mapping. The study area is located in Bavaria, Germany (ca. 165 km2), and labels for damaged areas were provided by the Bavarian State Institute of Forestry (LWF). Modifications of a U-Net architecture were compared to other approaches using transfer learning (e.g., VGG19) to find the most efficient architecture for the task on both datasets while keeping the computational time low. A custom implementation of U-Net proved to be more accurate than transfer learning, especially on medium (3 m) resolution PlanetScope imagery (intersection over union score (IoU) 0.55) where transfer learning completely failed. Results for transfer learning based on VGG19 on high-resolution aerial image data are comparable to results from the custom U-Net architecture (IoU 0.76 vs. 0.73). When using both architectures on a dataset from a different area (located in Hesse, Germany), however, we find that the custom implementations have problems generalizing on aerial image data while VGG19 still detects most damage in these images. For PlanetScope data, VGG19 again fails while U-Net achieves reasonable mappings. Results highlight the potential of Deep Learning algorithms to detect damaged areas with an IoU of 0.73 on airborne data and 0.55 on Planet Dove data. The proposed workflow with complete integration into ArcGIS is well-suited for rapid first assessments after a storm event that allows for better planning of the flight campaign followed by detailed mapping in a second stage.
he reduction of emission is a key goals for the aviation industry. One enabling technology to achieve this goal, could be the transition from conventional gas turbines to hybrid-electric drive trains. However, the requirements concerning weight and efficiency that come from applications like short range aircraft are significantly higher than what state-of-the-art technology can offer. A key technology that potentially allows to achieve the necessary power and volume densities for rotating electric machines is superconductivity. In this paper we present the concept of a high power density generator that matches the speed of typical airborne turbines in its power class. The design is based on studies that cover topology selection and further electromagnetic, HTS, thermal, structural and cryogenics aspects. All domains were analyzed by means of analytical sizing and 2D/3D FEA modeling. With the help of our digital twin that is a synthesis of these models, we can demonstrate for the first time that under realistic assumptions on material properties gravimetric power densities beyond 20 kW kg−1 can be achieved.
Internationale Soziale Arbeit neu denken - Zur Verschärfung Globaler Ungleichheit durch COVID-19
(2020)
Covid-19 hat sich über die ganze Welt ausbreitet und kann sie an den Rand eines ökonomischen und politischen Kollapses führen. Das Virus trifft zwar alle, doch mit Unterschieden. Die Krankheit wird zum einen die sozialen Probleme in den reichen Gesellschaften des Nordens verändern und auch vergrößern, doch insbesondere in den Ländern des Globalen Südens kann der Ausbruch in seinen ökonomischen und sozialen Folgen verheerend sein. Kann der Globale Norden noch eindämmende und die Folgen abfedernde Maßnahmen etablieren, so schlagen die Auswirkungen im Globalen Süden voll durch. In diesen Gesellschaften sind verheerende Folgen zu erwarten, die Armut, Elend, Hunger und Migrationsbewegungen intensivieren werden. Dadurch wird aber auch die Globale Ungleichheit zwischen den Nationen größer. In dieser Krise wäre Solidarität über nationale Grenzen hinweg gefragt; doch genau diese ist kaum erkennbar. Die Internationale Soziale Arbeit ist aufgefordert, ihre Ausrichtung neu zu definieren.
This paper deals with a tank level probing scenario using electromagnetic waves where the vessel supports a significant number of higher order modes which deteriorate the accuracy of the measurement. A signal model is presented which allows to estimate the accuracy of interferometric measurements taking higher order modes into account. Based on the predicted estimation an appropriate waveguide taper with a high suppression of parasitic modes is designed, fabricated and measured. The presented design approach is validated by measurements.
Research attention on decentralized autonomous energy systems has increased exponentially in the past three decades, as demonstrated by the absolute number of publications and the share of these studies in the corpus of energy system modelling literature. This paper shows the status quo and future modelling needs for research on local autonomous energy systems. A total of 359 studies are investigated, of which a subset of 123 in detail. The studies are assessed with respect to the characteristics of their methodology and applications, in order to derive common trends and insights. Most case studies apply to middle-income countries and only focus on the supply of electricity in the residential sector. Furthermore, many of the studies are comparable regarding objectives and applied methods. Local energy autonomy is associated with high costs, leading to levelized costs of electricity of 0.41 $/kWh on average. By analysing the studies, many improvements for future studies could be identified: the studies lack an analysis of the impact of autonomous energy systems on surrounding energy systems. In addition, the robust design of autonomous energy systems requires higher time resolutions and extreme conditions. Future research should also develop methodologies to consider local stakeholders and their preferences for energy systems.
Electrification of transport in RES-based power system will support the decarbonisation of the transport sector. However, due to the increase in energy demand and the large peak effects of charging, the passive integration of electric cars is likely to undermine sustainability efforts. This study investigates three different charging strategies for electric vehicle in Europe offering various degrees of flexibility: passive charging, smart charging and vehicle-to-grid, and puts this flexibility in perspective with the flexibility offered by interconnections. We use the Balmorel optimization tool to represent the short-term dispatch and long-term investment in the energy system and we contribute to the state-of-the-art in developing new methodologies to represent home charging and battery degradation. Our results show how each step of increased charging flexibility reduces system costs, affects energy mix, impacts spot prices and reduces CO2 emissions until the horizon 2050. We quantify how flexible charging and variable generation mutually support each other (>100 TWh from wind and solar energy in 2050) and restrict the business case for stationary batteries, whereas passive charging results in a substitution of wind by solar energy. The comparison of each charging scheme with and without interconnection expansion highlights the interplay between European countries in terms of electricity prices and CO2 emissions in the context of electrified transport. Although the best outcome is reached under the most flexible scenario at the EU level, the situation of the countries with the cheapest and most decarbonised electricity mix is damaged, which calls for adapted coordination policy at the EU level.
Understanding how innovations are accepted in a dynamic and complex market environment is a crucial factor for competitive advantage. To understand the relevant factors for this diffusion and to predict success, empirically grounded agent-based models have become increasingly popular in recent years. Despite the popularity of these innovation diffusion models, no common framework that integrates their diversity exists. This article presents a flexible, modular and extensible common description and implementation framework that allows to depict the large variety of model components found in existing models. The framework aims to provide a theoretically grounded description and implementation framework for empirically grounded agent-based models of innovation diffusion. It identifies 30 component requirements to conceptualize an integrated formal framework description. Based on this formal description, a java-based implementation allowing for flexible configuration of existing and future models of innovation diffusion is developed. As a variable decision support tool in decision-making processes on the adoption of innovations the framework is valuable for the investigation of a range of research questions on innovation diffusion, business model evaluation and infrastructure transformation.
Decision makers strive for optimal ways of production and usage of energy. To adjust their behavior to the future situation of markets and technology, the execution of different models predicting e.g. energy consumption, energy production, prices and consumer behavior is necessary. This execution is itself time-consuming and requires input data management. Furthermore, since different models cover different aspects of the energy domain, they need to be linked. To speed up the linkage and reduce manual errors, these linkage needs to be automated. We present IRPsim, an infrastructure for computation of different models and their linkage. The IRPsim-infrastructure enables management of model data in a structured database, parallelized model execution and automatic model linkage. Thereby, IRPsim allows researchers and practitioners to use energy system models for strategic business model analysis.
Community electricity storage systems for multiple applications promise benefits over household electricity storage systems. More economical flexibility options such as demand response and sector coupling might reduce the market size for storage facilities. This paper assesses the economic performance of community electricity storage systems by taking competitive flexibility options into account. For this purpose, an actor-related, scenario-based optimization framework is applied. The results are in line with the literature and show that community storage systems are economically more efficient than household storage systems. Relative storage capacity reductions of community storage systems over household storage systems are possible, as the demand and generation profiles are balanced out among end users. On average, storage capacity reductions of 9% per household are possible in the base scenario, resulting in lower specific investments. The simultaneous application of demand-side flexibility options such as sector coupling and demand response enable a further capacity reduction of the community storage size by up to 23%. At the same time, the competition between flexibility options leads to smaller benefits regarding the community storage flexibility potential, which reduces the market viability for these applications. In the worst case, the cannibalization effects reach up to 38% between the flexibility measures. The losses of the flexibility benefits outweigh the savings of the capacity reduction whereby sector coupling constitutes a far greater influencing factor than demand response. Overall, in consideration of the stated cost trends, the economies of scale, and the reduction possibilities, a profitable community storage model might be reached between 2025 and 2035. Future work should focus on the analysis of policy frameworks.
Despite the intensive research on residential photovoltaic adoption, there is a lack of understanding regarding the social dynamics that drive adoption decisions. Innovation diffusion is a social process, whereby communication structures and the relations between sender and receiver influence what information is perceived and how it is interpreted. This paper addresses this research gap by investigating stakeholder influences in household decision-making from a procedural perspective, so-called stakeholder dynamics. A literature review derives major influence dynamics which are then synthesized based on egocentric network maps for distinct process stages. The findings show a multitude of stakeholders that can be relevant in influencing photovoltaic adoption decisions of owner-occupied households. Household decision-makers are mainly influenced by stakeholders of their social network like family, neighbors, and friends as well as PV-related services like providers and civil society groups. The perceived closeness and likeability of a stakeholder indicate a higher level of influence because of greater trust involved. Furthermore, the findings indicate that social influence shifts gradually from many different stakeholders to a few core stakeholders later on in the decision-making process. These insights suggest that photovoltaic (PV) adoption may be more reliably predicted if a process perspective is taken into account that not only distinguishes between different stakeholders but considers their dynamic importance along the process stages. In addition, especially time- and location-bound factors affect the influence strength. This clearly shows the importance of local and targeted interventions to accelerate the uptake.
In addition to question prompts for information transfer, we also used prompts to facilitate the expression of emotions. Our aim was to investigate how a question prompt list (QPL) is accepted by patients and whether it enhances interactional empowerment of the patients in the consultation with the radio-oncological treatment team before the beginning of radiotherapy.
Methodology
Adult cancer patients before the beginning of radiotherapy were randomly assigned to the intervention group (IG) or control group (CG). The patients in the IG received a QPL with predefined subsets and subject areas. After the physician’s consultation, both groups completed a self-developed, content validated questionnaire on interactional empowerment. The IG evaluated the QPL using a self-developed instrument.
Result
A total of 279 adult cancer patients participated in the study (IG n = 139/CG n = 140). The participants of the IG reported a significantly higher interactional empowerment compared with those of the CG (t(277) = − 2.71, p = .007, 95% CI [− 1.61, − 0.26], d = 0.29). 60.4% of the IG agreed “rather” or “very” that they used the QPL in consultation with the medical team.
Conclusion
The QPL used in the consultation improved the self-assessed competence for interaction with the medical team and strengthened the interactional empowerment. The QPL was well accepted by the patients and is to be introduced into a routine as a practicable and simple instrument in the future. The support of patients in addressing concerns and fears is an important innovation.
We present TrueAdapt, a model-free method to learn online adaptations of robot trajectories based on their effects on the environment. Given sensory feedback and future waypoints of the original trajectory, a neural network is trained to predict joint accelerations at regular intervals. The adapted trajectory is generated by linear interpolation of the predicted accelerations, leading to continuously differentiable joint velocities and positions. Bounded jerks, accelerations and velocities are guaranteed by calculating the valid acceleration range at each decision step and clipping the network's output accordingly. A deviation penalty during the training process causes the adapted trajectory to follow the original one. Smooth movements are encouraged by penalizing high accelerations and jerks. We evaluate our approach by training a simulated KUKA iiwa robot to balance a ball on a plate while moving and demonstrate that the balancing policy can be directly transferred to a real robot with little impact on performance.
This book focuses on enabling mobile robots to recognize scenes in indoor environments, in order to allow them to determine which actions are appropriate at which points in time. In concrete terms, future robots will have to solve the classification problem represented by scene recognition sufficiently well for them to act independently in human-centered environments. To achieve accurate yet versatile indoor scene recognition, the book presents a hierarchical data structure for scenes – the Implicit Shape Model trees. Further, it also provides training and recognition algorithms for these trees. In general, entire indoor scenes cannot be perceived from a single point of view. To address this problem the authors introduce Active Scene Recognition (ASR), a concept that embeds canonical scene recognition in a decision-making system that selects camera views for a mobile robot to drive to so that it can find objects not yet localized. The authors formalize the automatic selection of camera views as a Next-Best-View (NBV) problem to which they contribute an algorithmic solution, which focuses on realistic problem modeling while maintaining its computational efficiency. Lastly, the book introduces a method for predicting the poses of objects to be searched, establishing the otherwise missing link between scene recognition and NBV estimation.
Flexible pick-and-place is a fundamental yet challenging task within robotics, in particular due to the need of an object model for a simple target pose definition. In this work, the robot instead learns to pick-and-place objects using planar manipulation according to a single, demonstrated goal state. Our primary contribution lies within combining robot learning of primitives, commonly estimated by fully-convolutional neural networks, with one-shot imitation learning. Therefore, we define the place reward as a contrastive loss between real-world measurements and a task-specific noise distribution. Furthermore, we design our system to learn in a self-supervised manner, enabling real-world experiments with up to 25000 pick-and-place actions. Then, our robot is able to place trained objects with an average placement error of 2.7 (0.2) mm and 2.6 (0.8)°. As our approach does not require an object model, the robot is able to generalize to unknown objects while keeping a precision of 5.9 (1.1) mm and 4.1 (1.2)°. We further show a range of emerging behaviors: The robot naturally learns to select the correct object in the presence of multiple object types, precisely inserts objects within a peg game, picks screws out of dense clutter, and infers multiple pick-and-place actions from a single goal state.