University of Siegen
Permanent URI for this communityhttps://dspace.ub.uni-siegen.de/handle/ubsi/7
Browse
201 results
Search Results
- Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Deterministic Superscalar AI Hardware Accelerator for Dependable Embedded Real-Time Systems(2026-10-12)Safety-critical real-time systems that rely on AI-based object classification and detection for critical decision-making require strict guarantees of reliability and timing correctness, as even minor deviations in behavior can lead to catastrophic consequences. Integrating AI hardware accelerators into such real-time environments therefore demands not only high computational throughput but also predictable and analyzable execution behavior. Conventional accelerators, such as the Versatile Tensor Accelerator (VTA), meet performance requirements but fall short in predictability due to their event-driven execution model and limited timing control. This variability introduces significant risks in scenarios where timely and deterministic inference is essential. To address these limitations, this work adopts a time-triggered execution paradigm for AI acceleration. It first introduces the Time-Triggered VTA (TT-VTA), which replaces the non-deterministic behavior of the original VTA with a deterministic counterpart based on statically defined, pattern-based timing schedules. These schedules are generated offline at design time using a scheduler with a built-in simulator for the VTA architecture. The scheduler further integrates additional simulation frameworks: DRAMSim2 is used to model memory-related instructions, while a cycle-accurate simulator captures compute-bound operations. Together, these components enable precise timing analysis, deterministic resource utilization, and predictable execution behavior. Evaluation with a CNN classification model shows that the TT-VTA preserves classification accuracy while ensuring predictable execution, however with a slight increase in inference latency (53 ms compared to 51 ms for conventional VTA), indicating room for further optimization. Building on this foundation, the Superscalar Time-Triggered VTA (STT-VTA) is proposed to improve performance without sacrificing determinism. STT-VTA extends the architecture by exploiting parallelism through multiple coordinated accelerator functional-units, whose configuration is determined at compile time via an enhanced architecture management and scheduling framework. This framework analyzes the computational structure needed for the target AI model and generates optimized schedules that maintain strict timing guarantees while increasing throughput. Experimental results demonstrate that the STT-VTA achieves the same classification accuracy as both TT-VTA and conventional VTA, while reducing inference latency by 17-41% compared to the TT-VTA. With the STT-VTA, another challenge arose in managing memory-intensive workloads, which introduce performance bottlenecks. To address this, the work further incorporates deterministic caching and pre-fetching strategies, leading to the Superscalar Time-Triggered VTA with Caching and Pre-fetching (STA-CP). Unlike conventional dynamic approaches, these mechanisms are integrated directly into the static scheduling process. At compile time, all tensor operations and memory transactions are fully specified, enabling predictable data movement and efficient overlap between computation and memory communication. By eliminating redundant memory accesses and hiding memory latency within a fully analyzable execution model, STA-CP achieves improved throughput while preserving strict temporal determinism. Experimental evaluation on different CNN classification models demonstrates up to a 30% reduction in deterministic inference latency compared to TT-VTA. Overall, the proposed STA-CP architecture demonstrates that the integration of time-triggered execution, controlled parallelism, and deterministic memory-management techniques can simultaneously achieve high performance and predictable execution behavior. These characteristics are fundamental for deploying AI accelerators in safety-critical real-time systems, where both computational correctness and strict timing guarantees are essential.Source Type:Doctoral Thesis55 14 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Speculative Design in HCI: A Methodological Investigation into Research Practice, Process, and Quality(2026)Speculative design has become a prominent approach in Human–Computer Interaction (HCI) for exploring not-yet-existing technologies and their ethical and societal implications. Yet, despite its growing use, it remains a contested research approach, often criticized for lacking methodological rigor. Positioned between art and academia, it is more a research stance than a method. It brings together a range of heterogeneous approaches, including design fiction, critical design, material speculation, and speculative enactments. What unites these approaches is not a shared methodology, but a common orientation. They all use design to explore possible futures in order to reflect on, question, and potentially change present assumptions, values, and trajectories. While prior work has clarified what speculative design is and introduced a variety of methods for its application, far less attention has been paid to foundational methodological questions. In particular, it remains unclear what defines “good” or successful speculative research and how speculative design and its diverse practices and methods can be effectively applied in research contexts. As a result, researchers face ongoing challenges in applying speculative design to the study of emerging technologies and in evaluating its outcomes, for instance within academic peer-review processes. This dissertation responds to this situation by examining how speculative design can be effectively applied to explore the design, roles, and implications of emerging technologies. It does so using the example of otherware, that is, autonomous and interactive technologies experienced as socially present others (e.g., voice assistants, social robots). Across five interrelated studies, it investigates speculative design’s methodological potentials, challenges, and “quality” both through practice and the systematic analysis of speculative design knowledge in HCI. A first set of three case studies investigates speculative design (particularly design fiction and speculative enactments) within academic HCI contexts. These studies examine how speculative approaches and methods can be applied to explore potentially meaningful futures with voice assistants and social robots. Across the cases, speculations were grounded in empirical and theoretical insights and materialized through tangible artefacts such as fictional product websites, documentaries, and flyers. These materializations enabled participants to imagine, inhabit, and reflect on possible technological futures in situated and experiential ways. Through participant engagement, the studies surfaced values, tensions, and assumptions that informed the design of otherware. Together, these studies revealed recurring methodological strategies for effective speculative inquiry (e.g., grounding speculation, materializing imaginaries) and a transferable three-part research process: (1) informing the design, (2) (co-)creating fictional worlds, and (3) empirically engaging with these worlds through debate and reflection. A fourth case study extends this investigation by situating speculative design within a real-world organizational context through an 18-month collaboration with an automotive company on the future of self-driving cars. By examining how speculative design practice changes when applied beyond academic and expert settings, it shows how business objectives and organizational structures can steer speculative design toward product-driven innovation rather than critical exploration. This shift affected speculation’s capacity to generate new knowledge about the potentials, challenges, and long-term implications of envisioned products and services. Speculations that closely resembled existing products and services proved less effective in stimulating discussion and uncovering assumptions about their future role and challenges than more provocative speculations. This finding highlights the importance of quality criteria for maintaining the exploratory, critical, and epistemically productive character of speculative design, particularly in organizational contexts. Building on these findings, the fifth study addresses the question of quality directly. Through a systematic scoping review of 63 theoretical and methodological publications on speculative design, design fiction, and critical design, it synthesizes the widely scattered characteristics scholars associate with “good” and successful speculations. This analysis results in a descriptive taxonomy that comprises nine quality criteria (e.g., critical, thought-provoking, participative), organized into three categories: speculative, discursive, and process. This taxonomy provides researchers, designers and practitioners with a shared vocabulary for planning, creating, and evaluating high-quality speculative designs. Across its five studies, the thesis makes three connected methodological contributions: (1) it conceptualizes speculative design as a positive, participatory, and empirically informed approach to design research for studying emerging technologies; (2) it operationalizes this approach through a transferable three-part research process; and (3) it introduces a descriptive taxonomy for the quality of speculation that provides orientation within this process. Together, these contributions offer recommendations and methodological guidance for researchers, designers, and practitioners in structuring and evaluating speculative design in HCI.Source Type:Doctoral Thesis35 16 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Automated Reliability Optimization for SRAM-based FPGAs(2026)Today’s most densely integrated Field Programmable Gate Arrays (FPGAs) store their configuration in integrated volatile SRAM cells and thereby offer virtually limitless reconfigurability. However, randomly occurring bit flips in an SRAM-based FPGA’s configuration memory can alter the implemented circuit. Furthermore, bit flips in user-defined storage elements such as flip-flops and block memory can corrupt the design’s state. Both effects may compromise system functionality, making SRAM-based FPGAs especially vulnerable to bit flips. As a mitigation measure, on-chip spatial redundancy is often introduced. This, however, increases FPGA resource usage and thereby system cost. Selectively protecting the design parts that are most vulnerable to bit flips allows to balance cost and reliability. This requires identifying especially vulnerable parts of the design to co-optimize a design’s bit-flip resilience with system cost. Existing reliability prediction approaches for FPGA designs often focus on logic resources, overlooking the significant and harder-to-predict effects of interconnect bit flips. Additionally, they do not account for process, voltage, and temperature variations on the outcome of configuration memory bit flips. These gaps in fault modeling limit current design-space exploration methods, preventing them from evaluating the impact of fault mitigation measures on bit-level reliability. This dissertation addresses these challenges with the following contributions. 1. This work provides an in-depth analysis of configuration memory bit flips in SRAM-based FPGAs. Based on this analysis, it develops and validates criticality prediction models to classify bitstream bits as benign or malignant when flipped. This reduced the need for time-consuming fault injection experiments by 60% to 98% with minimal false negatives in four case studies. In this course, this dissertation provides the first systematic study of how chip temperature and operating voltage influence configuration-level fault injection outcomes in FPGAs, which revealed significant inter- and intra-device variations in bit flip effects. 2. Based on the fault modeling work discussed above, an automated module-based design-space exploration approach for partial redundancy insertion is proposed. It co-optimizes FPGA resource usage and vulnerability to single-bit flips by encoding the module selection problem for genetic algorithms. In this course, it defines fitness metrics based on cost and vulnerability. Evaluated on six case-study designs, the approach generates Pareto-optimal partially redundant design variants, improving the cost-effectiveness of solutions by 13 to 59 percentage points compared to random starting sets. This method provides a diverse range of design variants along the cost-reliability trade-off. This enables the designer to select a variant that best fulfills the application’s, mission’s, or mission phase’s requirements. 3. A profiling approach is proposed for FPGA designs generated by high-level synthesis. The approach aims at identifying critical user-defined storage elements by profiling the design’s memory access patterns at the high-level language or intermediate representation level. Fault injection evaluations of case-study designs based on well-known benchmarks revealed a potential link between this memory access behavior and the in-hardware vulnerability of storage elements. This enables cost-effective design variants with partial bit-flip protection for user-defined storage elements. Furthermore, this high-level analysis may guide the design-space exploration process mentioned above, accelerating convergence. In summary, this dissertation contributes multiple methods for automating reliability improvements for SRAM-based FPGA designs. These methods are based on empirical analyses of FPGA vulnerabilities and address effects previously overlooked in existing research.Source Type:Doctoral Thesis12 12 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Boundary Handling Evaluation and Boundary Handling Enhancement for Adaptive Fluid Simulation in SPH(2026-08-07)One of the most popular fluid simulation methods in recent years is Smoothed Particle Hydrodynamics (SPH). Its flexibility, easy implementation, and natural free surface representation make it a state-of-the-art method in computer graphics. While fluid simulation is the main aspect of the SPH method, boundary handling also demands significant attention and effort to contain the fluid and prevent its penetration. There are a lot of boundary handling methods in SPH, and these methods differ in means of evaluation and comparison. It makes it difficult to comprehend which method is better in concrete circumstances. Moreover, each method also contributes differently to the fluid behavior that is not directly related to the fluid-boundary interaction. It causes different types of fluid behavior disruptions near the boundary for different methods. Despite the variety of boundary handling methods in SPH, there is no particle-based method that allows working with adaptive fluid. In this thesis, we address all aforementioned problems and propose our solutions to all of them. First of all, we developed a comprehensive protocol for boundary handling method testing. The protocol includes several testing scenarios with clear and unambiguous parameters of the simulation. It provides all the needed timing and all the needed values to collect from the simulation results. It also contains the format of data analysis and an example of the protocol's application in a real method evaluation. Secondly, we developed a decoupling method for the fluid-boundary interaction to prevent boundary handling from influencing fluid internal interaction and hence increase fluid stability near the boundary and obtain predictability of fluid behavior. The decoupling happens first of all in terms of fluid density, which is split into two densities, and based on the two densities, we obtain two independent pressures for the fluid particles near the boundary. The last thing we invented, as part of this work, is the adaptivity method for boundary particles. Currently, it is the only work on boundary adaptivity with a particle-based representation of boundaries. This method allows the use of particle-based boundaries with adaptive fluid in SPH. It significantly increases the potential of boundaries with respect to adaptive fluids.Source Type:Doctoral Thesis7 18 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Improving Imaging Pipelines by Sensor Optimization and Low-Level Image Enhancement(2026)Modern digital imaging pipelines transform incident light from the physical world into high-level semantic information for a variety of applications, such as medical imaging, robotics, visualization, and autonomous driving. They consist of multiple stages, including the focusing of light on an image sensor by optical systems, the measurement on the sensor itself, and further processing by an image signal processor (ISP) that also reduces noise and other errors from the measurement process, before the image is consumed by a task-specific algorithm such as a deep neural network. Conventionally, the training of the network is largely decoupled from the rest of the components, which are typically optimized for general-purpose use cases or human perception. This thesis investigates how imaging pipelines can be improved (i) as a whole by jointly optimizing sensor parameters and downstream learning-based algorithms, and (ii) by developing efficient low-level reconstruction and error-removal methods to handle the unavoidable imperfections introduced in the measurement process. The first part of this thesis focuses on the task-driven design of sensors. It develops methods to optimize the physical position of pixels on a sensor and the shape of spectral color filters, jointly with a downstream neural network to solve specific tasks. Although learned sensor layouts show a significant improvement in a variety of tasks compared to uniform grids, no such improvement could be found for learned spectral color filters in experiments for the specific task of semantic segmentation in autonomous driving, suggesting that, in this setting, the information gained from the spatial positioning of pixels is more impactful than fine-grained spectral information. The second part of the thesis improves the handling of imperfect measurements through several methods. It includes a provably convergent plug-and-play variational scheme with learned priors for solving inverse problems, a generic depth image enhancement method utilizing normal consistency constraints, and an outlier-robust point-based rendering algorithm that can directly render very noisy point clouds produced by depth sensors.Source Type:Doctoral Thesis12 10 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Pictogram-Based Eye-Tracking-Driven Robot Control(2026-06-17)People with severe disabilities strive for a more independence to improve their quality of life. Due to an aging society, the number of people dependent on daily care is expected to increase in the coming years. In response, research in human-robot interaction (HRI) focuses on investigating factors in user experience with robot controls and developing user-oriented robot applications to automate daily tasks. Providing people with severe disabilities with robots that can perform everyday tasks would reduce the workload of caregivers and family members while increasing users' independence. This work focuses on eye-tracking-driven controls for assistive robotic arms. Using eye-tracking as an input modality offers several advantages: it is noninvasive, requires minimal equipment investment, and allows for quick calibration compared to other input modalities. These benefits facilitate the deployment of such controls. This work identified many approaches to operate a robot with gaze. These approaches differed in the way gaze was employed to control the robot and in the level of effort required to issue commands. A major challenge of eye-tracking-driven controls is achieving three-dimensional control of the robotic arm using two-dimensional eye movements. One solution is to grant the robot more autonomy and enabling it to perform tasks independently. In this case, the user only needs to initiate the desired task. Such an outcome can be achieved through object-oriented gaze or user intent recognition methods. It was found that ambiguous information can result when gaze is tracked on objects that are used in different tasks. Considering how the system is used in everyday life, it is evident that robust task selection is difficult in an unstructured environment with many available tasks. A second challenge in designing eye-tracking controls is instructing the robot to interact correctly with an object. Many approaches use 3D-gaze estimation to determine the object's location relative to the eye-tracker. This requires high accuracy and precise calibration to minimize error. This doctoral thesis investigated the user experience with eye-tracking-driven controls and feasible control strategy solutions to address the challenges that were identified. The focus rested on creating a scalable control architecture that addresses usability and robustness challenges. This approach overcomes the key limitations identified in object-oriented selection methods. The resulting framework combines gaze-based interaction with semantically structured task representations and architectural safeguards against perceptual uncertainty. The framework allows for free head movement, reducing dependency on precise head tracking and additional sensors, thereby supporting accessible robotic systems.Source Type:Doctoral Thesis4 8 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Time-Triggered Organic Computing(2026-06-10)Qosja, MarioThe complexity of embedded computing systems is growing due to applications with increasing functional and non-functional requirements, including performance, reliability, and energy efficiency. Organic Computing (OC) has emerged as a research area for handling complex distributed embedded computer systems and is inspired by the organization of biological entities. In biology, Deoxyribonucleic Acid (DNA) is responsible for the structure and organization of living organisms. The same concept can be used in embedded systems, where a single computer file, known as Artificial DNA (ADNA), contains information for the whole system (e.g., sensors, actuators). In addition, following the hormone system in mammals, an Artificial Hormone System (AHS) has been developed that has the ability to read the ADNA and self-organize the system by exchanging hormones. It provides improved adaptability, robustness, and reduces development efforts. However, organic computing has drawbacks in terms of determinism, composability, and dependability, which are key features for safety-critical applications. Distributed computing systems using time-triggered communication networks have these characteristics and therefore offer distinct advantages for safety-critical systems. By combining organic computing technologies, such as artificial DNA and hormone features, with time-triggered communications, we can make these systems more adaptive, safe, reliable, and suitable for safety-critical applications while also providing self-x properties. Therefore, in this thesis, we present a Time-Triggered Organic Computing architecture (TTOC). In TTOC, computing nodes, also known as Processing Elements (PEs), exchange hormones with each other in order to perform the spatial allocation of application tasks. After a task is allocated, the chosen PE utilises a list-based scheduling algorithm that is designed specifically according to TTOC requirements to calculate the system schedule dynamically. The list-based algorithm calculates the temporal allocation of application tasks together with the spatial and temporal assignments of application messages. The task dispatcher is responsible for the execution of application tasks and messages as defined by the time-triggered schedule. In the case of a PE failure, the remaining active nodes will switch to a reconfiguration state where the execution of the affected task is halted. The system recomputes the application graph, and the AHS starts the process of hormone exchange to determine task allocation based on the newly computed graph. The system will return to a normal state when all the affected tasks from the PE failure have been reallocated, and new schedules are generated. To validate our proposed solution, an environment consisting of network and computation simulations is constructed. In addition, a co-simulation controller has been developed to couple both simulations. The results obtained from the evaluation of the use cases show improved temporal predictability and fault containment of the system.Source Type:Doctoral Thesis6 3 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Solving Combinatorial Optimization Problems in Computer Vision via Quantum Annealing(2025)Seelbach Benkner, MarcelSince the ability of quantum computing to speed up certain, specific computations has a solid theoretical underpinning, it is a natural question to ask if problems in computer vision also benefit from it. This doctoral thesis is in particular concerned with the question if the recently emerging optimization method of quantum annealing is useful for solving combinatorial optimization problems that arise in computer vision. The computer vision problem that is mainly investigated here is the problem of shape matching. The task of shape matching is to obtain the correspondences of two objects in a scene that are described with point clouds. Most often, one is concerned with finding the correspondences for a person who is recorded in 3D two times but with a different pose. The reason why this computer vision problem was picked as an interesting candidate for an application of quantum annealing is because it can be formulated as a difficult combinatorial optimization problem with a quadratic objective. This is close to the type of optimization problem that recent quantum annealing devices try to solve. The biggest difference is that in quantum annealing one a priori has unconstrained problems so that one cannot include equality constraints in the optimization. In the first work of this thesis, multiple methods are presented to overcome this difference and for really small shape matching problems experiments with simulated quantum annealing and with quantum annealing on D-Wave quantum computers were presented. In the second work of this thesis an iterative method was developed so that it was even possible to solve shape matching problems with over 500 vertices. Finally, a third publication looked at problems where one needs to find the correspondences between multiple shapes. Besides applications of quantum annealing for shape matching, this thesis also contains work on solving the motion segmentation problem with quantum annealing and work about integrating a quantum annealer in a neural network to solve computer vision problems. The integration of the quantum annealer in a neural network can also be viewed as the task to learn an appropriate formulation of a computer vision problem in a way that is solvable with a current quantum annealer. Finally, we also present work about an iterative method for compensating limited connectivity in quantum annealing devices. This work has no particular computer vision application in mind, but investigates the performance of the different approaches on Max-Cut problems. The thesis is structured in a cumulative way, so that the individual publications are presented separately. On a broader picture, the works are motivated by the following questions: What could be the impact of quantum annealing on computer vision and how could computer vision algorithms that use quantum annealing look like? And from another perspective: How would the field of computer vision be changed if there were massive advances in building quadratic, unconstrained, binary optimization problem solvers with or even without utilizing quantum computing?Source Type:Doctoral Thesis4 10 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , The Impact of Restarts in Online Machine Scheduling(2026)This dissertation investigates the impact of job preemptions and restarts in online machine scheduling, with a focus on three fundamental objectives. An online scheduling algorithm determines whether to process a job or not without knowing the future events. One important extension to the classical online scheduling framework is the use of restarts. Allowing restarts means that the processing of a job may be interrupted, but in this case the interrupted job loses all its previous progress and must be started again later, until it is completed without interruptions. This model also known as preemption with restarts. Although restarts can potentially be very beneficial in the context of online scheduling, there has been relatively little research on this topic up until now. First, we study the problem of minimizing the total completion time on a single machine. We present a simple deterministic online algorithm that achieves a competitive ratio better than 1.4568, improving upon the previous best bound of 3/2. The algorithm follows an increasing-size processing order and interrupts a running job only when doing so reduces the completion time of an arriving job by more than a factor of 1.4568 compared to processing it immediately. Second, we extend the results to the minimization of total weighted completion time. For the case of equal processing times, we design a deterministic online algorithm that achieves a competitive ratio better than 1.325 based on carefully structured decision intervals. The algorithm uses at most three restarts for the entire schedule and redefines these decision intervals in certain situations to avoid further interruptions. Finally, we consider the weighted makespan minimization problem. This problem is generalization of the classical makespan minimization problem. We present a deterministic online algorithm that achieves a competitive ratio better than 1.3098 in the case where all jobs have equal processing times.Source Type:Doctoral Thesis3 13 - Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Variational L0 Regularization for Enhanced Depth Image Analysis(2025)Mathematical optimization is fundamental in computational imaging, where problems often center on minimizing an objective function under various constraints. An important class of optimization problems involves minimal partitioning, which aims to segment data into partitions which meet certain criteria, such as photometric or geometric similarity. This class of problems is computationally demanding as it is has a NP-Hard complexity, which means that the computational effort required to estimate a solution increases exponentially with the size of the data. Consequently, finding (nearly) optimal solutions within a reasonable amount of time is often infeasible. This dissertation presents an analysis of the efficient and effective utilization of minimal partitioning techniques in RGB-D image processing. This is achieved by constraining the underlying optimization problems with the L0 “norm”, which enforces piecewise constant solutions. The first case study analyses the utilization of the L0 “norm” for scene flow estimation, where 3D motions are modeled as rigid transformations and constrained to be piece-wise constant. By substituting the L0 “norm” with the truncated quadratic “norm”, this approach achieves both high accuracy as well as real-time performance. The next focus is on the estimation of geometric models on RGB-D images, specifically to obtain denoised depth and normal values. The novel geometric model definition along with the Cut Pursuit optimization leads to a higher accuracy compared to other model based methods. The third and final enhancement addresses the high computational demands of the preceding Cut Pursuit based method by introducing a stochastic optimization technique. It improves the convergence rate and also the overall runtime efficiency.Source Type:Doctoral Thesis1 5
