B-spline surfaces are a well-established tool to analytically describe objects. They are commonly used in various fields, e.g., mechanical and aerospace engineering, computer aided design, and computer graphics. Obtaining and using B-spline surface models of real-life objects is an intricate process. Initial virtual representations are usually obtained via scanning technologies in the form of discrete data, e.g., point clouds, surface meshes, or volume images. This data often requires pre-processing to remove noise and artifacts. Even with high quality data, obtaining models of complex or very large structures needs specialized solutions that are viable for the available hardware. Once B-spline models are constructed, their properties can be utilized and combined with application-specific knowledge to provide efficient solutions for practical problems. This thesis contributes to various aspects of the processing pipeline. It addresses pre-processing, creating B-spline models of large and topologically challenging data, and the use of such models within the context of visual surface inspection. Proposed methods improve existing solutions in terms of efficiency, hardware restrictions, and quality of the results. The following contributions are presented: Fast and memory-efficient quantile filter: Quantile filters are widely used operations in image processing. The most common instance is the median filter which is a standard solution to treat noise while preserving the shape of an object. Various implementations of such filters are available offering either high performance or low memory complexity, but not both. This thesis proposes a generalization of two existing algorithms: one that favors speed and one that favors low memory usage. An adaptable hybrid algorithm is introduced. It can be tuned for optimal performance on the available hardware. Results show that it outperforms both state-of-the-art reference methods for most practical filter sizes. Robust B-spline reconstructions of isosurfaces in volume images: The micro-structure of wood-based thermal insulation materials is analyzed to research heat conductivity properties. High-quality scans reveal a complex system of cellulose fibers. B-spline models of individual fibers are highly desirable to conduct simulations. Due to the physical processing of the material, the surfaces of those fibers consist of challenging elements like loose filaments, holes, and tunnels. Standard solutions fail to partition the data into a small number of quadrilateral cells which is required for the B-spline construction step. A novel approach is presented that splits up the data processing into separate topology and geometry pipelines. This robust method is demonstrated by constructing B-spline models with 236 to 676 surfaces from triangulated isosurfaces with 423628 to 1203844 triangles. Local method for smooth B-spline surface approximations: Constructing smooth B-spline models to approximate discrete data is a challenging task. Various standard solutions exist, often imposing restrictions to knot vectors, spline order, or available degrees of freedom for the data approximation. This thesis presents a local approach with less restrictions aiming for approximate \(G^1\)-continuity. Nonlinear terms are added to standard minimization problems. The local design of the algorithm compensates for the higher computational complexity. Results are shown and evaluated for objects of varying complexity. A comparison with an exact \(G^1\)-continuous method shows that the novel method improves approximation accuracy on average by a factor of 10 at the cost of having small discontinuities in normal vectors of less than 1 degree. Model-based viewpoint generation for surface inspection: Within modern and flexible factories, surface inspection of products is still a very rigid process. An automated inspection system requires the definition of viewpoints from which a robot then takes pictures during the inspection process. Setting up such a system is a time-intensive process which is primarily done manually by experts. This work presents a purely virtual approach for the generation of viewpoints. Based on an intuitive definition of analytic feature functionals, a non-uniform sampling with respect to inspection-specific criteria is performed on given B-spline models. This leads to the definition of a low number of viewpoint candidates. Results of applying this method to several test objects with varying parameters indicate that good viewpoints can be obtained through a fast process that can be performed fully automatically or interactively through the use of meaningful parameters.

Interconnection networks enable fast data communication between components of a digital system. The selection of an appropriate interconnection network and its architecture plays an important role in the development process of the system. The selection of a bad network architecture may significantly delay the communication between components and decrease the overall system performance. There are various interconnection networks available. Most of them are blocking networks. Blocking means that even though a pair of source and target components may be free, a connection between them might still not be possible due to limited capabilities of the network. Moreover, routing algorithms of blocking networks have to avoid deadlocks and livelocks, which typically does only allow poor real-time guarantees for delivering a message. Nonblocking networks can always manage all requests that are coming from their input components and can therefore deliver all messages in guaranteed time, i.e, with strong real-time guarantees. However, only a few networks are nonblocking and easy to implement. The simplest one is the crossbar network which is a comparably simple circuit with also a simple routing algorithm. However, while its circuit depth of O(log(n)) is optimal, its size increases with O(n^2) and quickly becomes infeasible for large networks. Therefore, the construction of nonblocking networks with a quasipolynomial size O(nlog(n)^a) and polylogarithmic depth O(log(n)^b) turned out as a research problem. Benes [Clos53; Bene65] networks were the first non blocking networks having an optimal size of O(nlog(n)) and an optimal depth of O(log(n)), but their routing algorithms are quite complicated and require circuits of depth O(log(n)^2) [NaSa82]. Other nonblocking interconnection networks are derived from sorting networks. Essentially, there are merge-based (MBS) and radix-based (RBS) sorting networks. MBS and RBS networks can be both implemented in a pipelined fashion which leads to a big advantage for their circuit implementation. While these networks are nonblocking and can implement all n! permutations, they cannot directly handle partial permutations that frequently occur in practice since not every input component communicates at every point of time with an output component. For merge-based sorting networks, there is a well known general solution called the Batcher-Banyan network. However, for the larger class of radix-based sorting networks this does not work, and there is only one solution known for a particular permutation network. In this thesis, new nonblocking radix-based interconnection networks are presented. In particular, for a certain permutation network, three routing algorithms are developed and their circuit implementations are evaluated concerning their size, depth, and power consumption. A special extension of these networks allows them to route also partial permutations. Moreover, three general constructions to convert any binary sorter into a ternary split module were presented which is the key to construct a radix-based interconnection network that can cope with partial permutations. The thesis compares also chip designs of these networks with other radix-based sorting networks as well as with the Batcher-Banyan networks as competitors. As a result, it turns out that the proposed radix-based networks are superior and could form the basis of larger manycore architectures.

This thesis investigates how smart sensors can quantify the process of learning. Traditionally, human beings have obtained various skills by inventing technologies. Those who integrate technologies into daily life and enhance their capabilities are called augmented humans. While most existing augmenting human technologies focus on directly assisting specific skills, the objective of this thesis is to assist learning -- the meta-skill to master new skills -- with the aim of long-term augmentations. Learning consists of cognitive activities such as reading, writing, and watching. It has been considered that tracking them by motion sensors (in the same way as the recognition of physical activities) is a challenging task because dynamic body movements could not be observed during cognitive activities. I have solved this problem with smart sensors monitoring eye movements and physiological signals. I propose activity recognition methods using sensors built into eyewear computers. Head movements and eye blinks measured by an infrared proximity sensor on Google Glass could classify five activities including reading with 82% accuracy. Head and eye movements measured by electrooculography on JINS MEME could classify four activities with 70% accuracy. In a wild experiment involving seven participants who wore JINS MEME more than two weeks, deep neural networks could detect natural reading activities with 74% accuracy. I demonstrate Wordometer 2.0, an application to estimate the number of rear words on JINS MEME, which was evaluated in a dataset involving five readers with 11% error rate. Smart sensors can recognize not only activities but also internal states during the activities. I present an expertise recognition method using an eye tracker which performs 70% classification accuracy into three classes using one minute data of reading a textbook, a positive correlation between interest and pupil diameter (p < 0.01), a negative correlation between mental workload and nose temperature measured by an infrared thermal camera (p < 0.05), an interest detection on newspaper articles, and effective gaze and physiological features to estimate self-confidence while solving multiple choice questions and spelling tests of English vocabulary. The quantified learning process can be utilized for feedback to each learner on the basis of the context. I present HyperMind, an interactive intelligent digital textbook. It can be developed on HyperMind Builder which may be employed to augment any electronic text by multimedia aspects activated via gaze. Applications mentioned above have already been deployed at several laboratories including Immersive Quantified Learning Lab (iQL-Lab) at the German Research Center for Artificial Intelligence (DFKI).

In this paper we present the results of the project “#Datenspende” where during the German election in 2017 more than 4000 people contributed their search results regarding keywords connected to the German election campaign. Analyzing the donated result lists we prove, that the room for personalization of the search results is very small. Thus the opportunity for the effect mentioned in Eli Pariser’s filter bubble theory to occur in this data is also very small, to a degree that it is negligible. We achieved these results by applying various similarity measures to the result lists that were donated. The first approach using the number of common results as a similarity measure showed that the space for personalization is less than two results out of ten on average when searching for persons and at most four regarding the search for parties. Application of other, more specific measures show that the space is indeed smaller, so that the presence of filter bubbles is not evident. Moreover this project is also a proof of concept, as it enables society to permanently monitor a search engine’s degree of personalization for any desired search terms. The general design can also be transferred to intermediaries, if appropriate APIs restrict selective access to contents relevant to the study in order to establish a similar degree of trustworthiness.

Activity recognition has continued to be a large field in computer science over the last two decades. Research questions from 15 years ago have led to solutions that today support our daily lives. Specifically, the success of smartphones or more recent developments of other smart devices (e.g., smart-watches) is rooted in applications that leverage on activity analysis and location tracking (fitness applications and maps). Today we can track our physical health and fitness and support our physical needs by merely owning (and using) a smart-phone. Still, the quality of our lives does not solely rely on fitness and physical health but also more increasingly on our mental well-being. Since we have learned how practical and easy it is to have a lot of functions, including health support on just one device, it would be specifically helpful if we could also use the smart-phone to support our mental and cognitive health if need be. The ultimate goal of this work is to use sensor-assisted location and motion analysis to support various aspects of medically valid cognitive assessments. In this regard, this thesis builds on Hypothesis 3: Sensors in our ubiquitous environment can collect information about our cognitive state, and it is possible to extract that information. In addition, these data can be used to derive complex cognitive states and to predict possible pathological changes in humans. After all, not only is it possible to determine the cognitive state through sensors but also to assist people in difficult situations through these sensors. Thus, in the first part, this thesis focuses on the detection of mental state and state changes. The primary purpose is to evaluate possible starting points for sensor systems in order to enable a clinically accurate assessment of mental states. These assessments must work on the condition that a developed system must be able to function within the given limits of a real clinical environment. Despite the limitations and challenges of real-life deployments, it was possible to develop methods for determining the cognitive state and well-being of the residents. The analysis of the location data provides a correct classification of cognitive state with an average accuracy of 70% to 90%. Methods to determine the state of bipolar patients provide an accuracy of 70-80\% for the detection of different cognitive states (total seven classes) using single sensors and 76% for merging data from different sensors. Methods for detecting the occurrence of state changes, a highlight of this work, even achieved a precision and recall of 95%. The comparison of these results with currently used standard methods in psychiatric care even shows a clear advantage of the sensor-based method. The accuracy of the sensor-based analysis is 60% higher than the accuracy of the currently used methods. The second part of this thesis introduces methods to support people’s actions in stressful situations on the one hand and analyzes the interaction between people during high-pressure activities on the other. A simple, acceleration based, smartwatch instant feedback application was used to help laypeople to learn to perform CPR (cardiopulmonary resuscitation) in an emergency on the fly. The evaluation of this application in a study with 43 laypersons showed an instant improvement in the CPR performance of 50%. An investigation of whether training with such an instant feedback device can support improved learning and lead to more permanent effects for gaining skills was able to confirm this theory. Last but not least, with the main interest shifting from the individual to a group of people at the end of this work, the question: how can we determine the interaction between individuals within a group of people? was answered by developing a methodology to detect un-voiced collaboration in random ad-hoc groups. An evaluation with data retrieved from video footage provides an accuracy of up to more than 95%, and even with artificially introduced errors rates of 20%, still an accuracy of 70% precision, and 90% recall can be achieved. All scenarios in this thesis address different practical issues of today’s health care. The methods developed are based on real-life datasets and real-world studies.