Publication Server of Kempten University
Not a member yet
2351 research outputs found
Sort by
An Analysis of Camera Lens Contaminants in Vehicle Vision Systems: Characterizing Dust, Mud, and Droplet Patterns
No full textThe contamination of vehicle camera lenses by artifacts such as mud, dust, or rain droplets presents a safety critical challenge that must be addressed by Advanced Driver Assistance Systems (ADAS). These obstructions, which impair the camera's functionality, are categorized and analyzed into three distinct classes: opaque, translucent, and water droplets. Using the publicly available Soiling-WoodScape (1) dataset, we investigate the biases and effects associated with these different types of contaminations. Our analysis focuses on the position, shape, and other attributes of the affected regions. This detailed understanding enables the development of improved simulation methods and provides valuable insights into the biases inherent in learning based models trained on this dataset. Such knowledge is essential for enhancing the robustness and reliability of ADAS under real world soiling conditions
Automating 3D printing for mass production
No full textThe development and evaluation of an automation concept for high-performance 3D printers in an industrial environment is presented. The paper’s special characteristic is the multi-domain approach, which combines development with parallel simulation studies of the cell. The 3D printing robot cell was designed with an individual gripping system and a magazine for printing plates, and the production performance of the concept was evaluated with simulation studies of the robot performance and cycle time simulation
Motion planning for an autonomous agricultural robot in unstructured terrain
No full textDespite the wide body of literature on motion planning for autonomous robots targeting structured agricultural environments, there remains a need for more efficient and reliable approaches for unstructured terrains. This work is motivated by a cooperation with the startup Paltech, which develops a weed-removal robot for grasslands. In this work, we employ a typical layered approach for robot navigation. However, in contrast to standard navigation tasks, such as point-to-point navigation, the global planner must solve the more challenging problem of visiting multiple targets in an optimal way while considering the kinematic constraints of the vehicle
Ego-Vehicle Localisation using Single-Camera Trilateration and Computer Vision
No full textThis thesis investigates monocular ego-vehicle localisation using single-camera trilateration and transformer based geometric reasoning. A real-time localisation pipeline is proposed that estimates the global 3D position of a vehicle-mounted monocular camera from a single RGB image and a structured 3D point cloud map. The core contribution is the Single-Camera Trilateration Transformer (SCTT), a transformer-based model that infers position from per-keypoint descriptors, ray directions, and precomputed spatial priors. The architecture fuses these inputs via cross-attention and confidence-weighted aggregation, enabling robust single-frame inference without requiring visual odometry, GNSS, or LiDAR. To construct spatial priors, an offline embedding pipeline is developed based on semantic masking, static keypoint matching, and geometric filtering. Top-ranked matches are projected into fixed-length scene embeddings via supervised pretraining and referenced by a spatial database for runtime retrieval. SCTT is evaluated on a real-world dataset collected in Landsberg am Lech, Germany. Experiments demonstrate sub-metre accuracy under diverse urban conditions including occlusion, dynamic traffic, and adverse weather. The pipeline operates in real time on GPU hardware and integrates natively with ROS 2. These results show that accurate monocular localisation is feasible when static geometric context is fused with learned representations. The SCTT architecture provides a scalable and modular foundation for GNSS-independent ego-positioning in urban environments
Comprehensive Analysis of Neural Network Inference on Embedded Systems: Response Time, Calibration, and Model Optimisation
No full textThe response time of Artificial Neural Network (ANN) inference is critical in embedded systems processing sensor data close to the source. This is particularly important in applications such as predictive maintenance, which rely on timely state change predictions. This study enables estimation of model response times based on the underlying platform, highlighting the importance of benchmarking generic ANN applications on edge devices. We analyze the impact of network parameters, activation functions, and single- versus multi-threading on response times. Additionally, potential hardware-related influences, such as clock rate variances, are discussed. The results underline the complexity of task partitioning and scheduling strategies, stressing the need for precise parameter coordination to optimise performance across platforms. This study shows that cutting-edge frameworks do not necessarily perform the required operations automatically for all configurations, which may negatively impact performance. This paper further investigates the influence of network structure on model calibration, quantified using the Expected Calibration Error (ECE), and the limits of potential optimisation opportunities. It also examines the effects of model conversion to Tensorflow Lite (TFLite), highlighting the necessity of considering both performance and calibration when deploying models on embedded systems
Power Indices with Threats in Precoalitions
No full textWe investigate power indices for simple games with precoalitions which distribute power among players in an external and an internal step. We extend an existing approach which uses the Public Good index both on the external level in the quotient game as well as on the internal level for measuring the leverage of players to threaten their peers through departing the precoalition. We replace the Public Good index in that model by five other efficient power indices, i.e., the Shapley–Shubik index, the Deegan–Packel index, the Johnston index and two indices based on null player free winning coalitions. Axiomatizations of the novel power indices with threat partitions are presented. We also propose a slight modification to the existing framework for threat power indices which guarantees that null players are always assigned zero power. Numerical results for all power indices combined with different threat partitions are presented and discussed
Retrieval-Augmented Generation for ICD-10 Coding in German Clinical Texts – A Technical Case Report
No full textIntroduction: Manual ICD-10 coding of German clinical texts is time-consuming
and error-prone. This project aims to develop a semi-automated pipeline for efficient coding of unstructured medical documentation. State of the art: Existing approaches often rely on fine-tuned language models that require large datasets and perform poorly on rare codes, particularly in low-resource languages such as German. Concept: The proposed system integrates Named Entity Recognition, semantic and lexical retrieval, abbreviation resolution, and context-aware normalization within a Retrieval-Augmented Generation (RAG) framework using a compact generative model. Implementation: The pipeline utilizes Sentence-BERT embeddings, FAISS indexing, and the Mistral-Small-Instruct model. ICD codes are assigned through a combination of semantic similarity and generative refinement among the top retrieval candidates. Lessons learned: Major sources of error were found in semantic retrieval and diagnosis normalization. Future improvements should focus on domain-specific German embeddings, more robust abbreviation handling, and enhanced context-aware prompting to increase accuracy and usability
in clinical environments
Standardized Data Exchange for Industrial Dataspaces Using Asset Administration Shell, AutomationML and OPC UA
No full textThis paper presents an integration methodology for Asset Administration Shell (AAS), AutomationML (AML), and OPC Unified Architecture (OPC UA) to address challenges in achieving seamless interoperability across engineering, runtime, and lifecycle data in Industry 4.0 ecosystems. The proposed approach combines AML for engineering data structuring, OPC UA for real-time communication, and AAS for unified lifecycle management, enabling seamless and efficient data exchange. The methodology was validated using a lab-scale production system, demonstrating its potential to streamline lifecycle management processes and enhance interoperability. While the integration was successful, challenges remain, highlighting the need for further standardization. This research provides a scalable framework for advancing digital transformation in manufacturing systems
Bridging Sensors and Scenarios: A Method for Reliable Driving Functions
No full textThe development of driving functions for advanced driver assistance systems and automated/autonomous driving relies on the interaction between the driving function and its environment. While driving function objectives can be well-defined, systematic descriptions of relevant scenarios are often lacking. A comprehensive scenario catalog remains challenging due to uncertainties in identifying critical environmental elements. Sensor modeling is crucial for driving function validation, as sensor reliability directly impacts vehicle homologation. Realistic sensor behaviour models enable closed-loop simulations, integrating driving functions, sensor dynamics, and known scenarios. This approach accelerates validation and enhances system understanding. This study presents an approach to accurately model both sensor impairments and the vehicle environment for localization with a camera sensor. Camera behaviour models are implemented within the software library sensor models, with parameters derived from datasheets or precise test setups. The environment is represented by a digital twin of the city of Kempten, independent of specific vehicle applications. Scenarios are modeled and then simulated with the Python-based software library osi3test in a fixed time-step loop, visualized via software library panda3d. Standardized interfaces such as ASAM OSI, FMI and OGS standards facilitate interoperability and deeper insights into model interactions that indicate real interactions worth investigating
Nonlinear causal discovery for grouped data
No full textInferring cause-effect relationships from observational data has gained significant attention in recent years, but most methods are limited to scalar random variables. In many important domains, including neuroscience, psychology, social science, and industrial manufacturing, the causal units of interest are groups of variables rather than individual scalar measurements. Motivated by these applications, we extend nonlinear additive noise models to handle random vectors, establishing a two-step approach for causal graph learning: First, infer the causal order among random vectors. Second, perform model selection to identify the best graph consistent with this order. We introduce effective and novel solutions for both steps in the vector case, demonstrating strong performance in simulations. Finally, we apply our method to real-world assembly line data with partial knowledge of causal ordering among variable groups