Explore CTU's Digital Library

Access research publications, theses and academic works

Faculty of Civil Engineering Faculty of Mechanical Engineering Faculty of Electrical Engineering Faculty of Nuclear Sciences and Physical Engineering Faculty of Architecture Faculty of Transportation Sciences Faculty of Biomedical Engineering Faculty of Information Technology Klokner Institute Masaryk Institute of Advanced Studies Central Library Acta Polytechnica Acta Polytechnica CTU Proceedings CTU Publishing Activity

Recent Submissions

Weakly-supervidsed learning with applications in biomedicial image analysis
Slabě řízené učení s aplikacemi v analýze biomedicínských obrazů
(Czech Technical University in Prague) Baručić, Denis; Kybic, Jan; Brázdil, Tomáš; Cheplygina, Veronika
Supervised learning has achieved remarkable success in biomedical image analysis, but it typically requires large amounts of fully annotated data. Acquiring such annotations is laborious and, in the biomedical domain, often requires expert knowledge. In many practical scenarios, weak annotations of various forms are easier to obtain. This thesis studies several examples of learning from data consisting of labeled groups of unlabeled instances. We first consider Multiple Instance Learning (MIL), where a group is labeled positive if it contains at least one positive instance. Building upon an existing MIL method, we propose a learning objective that explicitly incorporates instance-level supervision through pseudo-labeling. The resulting method improves instance classification performance while preserving strong group-level accuracy. Next, we address Learning from Label Proportions (LLP). In LLP, each group is annotated with the proportion of positive instances it contains. We focus on problems with small training groups and model the number of positive instances in a group using the Poisson binomial distribution. We propose an expectationmaximization algorithm for training a deep network that parametrizes this model. The method converges faster than existing deep LLP approaches while maintaining competitive predictive performance. Third, we study weakly supervised image segmentation with annotation in the form of object sizes. Our approach is based on a distance function that is non-differentiable with respect to the pixel probabilities, which prevents the application of gradient descent. Therefore, instead of optimizing the loss directly, we inject virtual noise into the system, and optimize the expected loss. The approach improves segmentation performance when only a few fully (pixel-wise) annotated images and many weakly annotated images are available. Finally, we propose a group-based formulation for recognizing chemical compounds applied to cells in microscopy images. The formulation allows the classifier to account for varying concentration and exposure conditions of the predicted compound. The resulting classification method matches or outperforms existing alternatives. Together, our contributions demonstrate that, although weakly supervised learning is inherently challenging, it can be successfully applied to practical biomedical image analysis problems.
Towards Robust Long-Term Visual Teach and Repeat Navigation
Cesta k robustní dlouhodobé vizuální navigaci
(Czech Technical University in Prague) Rozsypálek, Zdeněk; Krajník, Tomáš; Christoforis, Pablo de; Balogh, Richard
Autonomous mobile robotics is a rapidly developing field of research with many real-world applications, where the viability of mobile system deployment is heavily dependent on the ability to reliably navigate within the operational area. Autonomous navigation is a complex problem that requires advanced data processing and decision-making on multiple levels. One widely used framework is Visual Teach and Repeat (VT&R) navigation, which divides mapping and navigation into two separate processes and allows a precise definition of the robots operational space. These simplifications make VT&R a popular choice for various deployment scenarios. However, unstructured or highly dynamic environments still pose a significant challenge to most existing autonomous mobile systems. This thesis focuses on improving navigation robustness in long-term autonomous setups, where uncontrolled environments can undergo significant appearance or structural changes. We propose a novel visual model with state-of-the-art performance for horizontal image registration under time-specific environmental changes. Building on this model, we design a probabilistic state estimator that significantly improves navigation precision. We also present an alternative approach to VT&R that can substantially reduce mapping requirements. All proposed methods address common problems encountered in real-world deployments and have practical implications for the applicability of VT&R in autonomous mobile robotics.
Effect of surface preparation and adhesive type on CFRP–steel bridge joints
(Ernst & Sohn, 2026) Boutar Y.; Ryjáček P.; Hovorka J.; Jarolím T.; Montalbano A.; Tichá P.
Strengthening steel bridges using carbon fibre reinforced polymer (CFRP) laminates is an effective technique to increase load-carrying capacity and extend the service life of ageing infrastructure. The performance of CFRP strengthening systems depends strongly on the quality of the adhesive bond between steel and CFRP, making adhesive selection and surface preparation critical design parameters. This study experimentally investigates the influence of adhesive type and primer application on the bond behaviour of CFRP–steel joints manufactured using a wet-layup technique. Mild steel substrates were sandblasted, with half of the specimens receiving an additional primer coating. Three commercially available structural epoxy adhesives were used to bond CFRP fabrics to steel substrates. A total of 18 CFRP–steel double-lap joints were tested under quasi-static shear loading. Load–displacement behaviour and shear strength were evaluated. The results show that both adhesive type and primer application significantly affect joint performance. While primer application improved consistency for some adhesive systems, it did not universally increase shear strength and, in certain cases, resulted in reduced capacity. The findings highlight the adhesive-dependent nature of surface treatment effects and provide practical guidance for the design of durable CFRP strengthening systems for steel structures.
Revitalization of Ruzyně Square - IMPRINT
Revitalizace Ruzyňského náměstí - OTISK
(Czech Technical University in Prague, 2026-06-02) Mazalová, Natálie; Sklenář, Tomáš; Trevisan, Jitka; Imramovská, Martina
The bachelor's thesis titled Revitalization of Staré náměstí in Ruzyně: Imprints Beneath the Surface is based on a landscape study developed in the summer semester of the 2025/2026 academic year. It focuses on the revitalization of the public space of Staré náměstí in Ruzyně, in connection with the remand prison complex and the course of the Litovický stream. The thesis centers on detailing a selected part of the area to the level of project documentation corresponding to the building permit stage. The study explores the relationship between humans and the landscape through traces created by the long-term effects of natural processes and human activity. The design works with the memory of the place, the historical context of the area, and natural landscape-forming processes, particularly the water regime of the Litovický stream and its role within the structure of the public space. The bachelor's thesis focuses on designing a water feature as a stable landscape intervention that enables a more natural integration of the watercourse into the territory, supports the landscape's retention capacity, and contributes to the enhancement of the public space. The water feature is conceived as a memory structure anchored in the landscape, the significance of which stems from working with material, time, and natural processes. The proposal includes the design of the public space in front of the prison complex as a place for meeting, transit, and lingering, connecting the needs of Ruzyně's residents, visitors, and broader landscape links. The design strives to create a legible, safe, and sustainably usable space. The solution also includes the restoration of the granary building, which provides operational and technical facilities for the use of the public space. The aim of the project is to create a functional and stable public space responding to the historical, social, and landscape context of the area. The design works with water, terrain modifications, vegetation, material solutions, and stormwater management, with an emphasis on its sensitive integration into the existing structure of the site and respect for its layered character.
Proposal for Modifications to the Masaryk Circuit in Brno to Accommodate the F1 Race
Návrh úprav Masarykova okruhu Brno pro zajištění závodu F1
(Czech Technical University in Prague, 2026-08-04) Machačný, Vladimír; Filip, Josef; Kočárková, Dagmar; Tichý, Tomáš
This thesis aims to identify non-compliant parameters of the Masaryk Circuit with respect to FIA Grade 1 requirements and to propose specific structural solutions for their remediation. The analysis is based on direct fieldwork in cooperation with STRABAG, a.s., which carried out a comprehensive reconstruction of the circuit pavement in 2025, as well as on data from the Czech Office for Surveying, Mapping and Cadastre (ČÚZK) and data processed in Autodesk Civil 3D. The thesis systematically compares the existing geometric, safety, and material parameters of all fourteen directional corners, the pit lane, and the circuit infrastructure against the normative requirements of Appendix O to the 2026 FIA Sporting Regulations. Based on maximum achievable single-seater speed calculations derived from real telemetry data of the 2025 Japanese Grand Prix, braking distances and required run-off area dimensions are determined. The results show that the key non-compliant parameter is the run-off area of the first directional corner, named after František Šťastný, where the missing run-off length reaches the order of tens of meters depending on the chosen run-off surface type. The proposed structural solution combines the extension of the asphalt surface (tarmac) beyond the outer edge of the track with the installation of TecPro absorption systems certified for Formula 1 races, while respecting the spatial constraints posed by the presence of the III/6021 road in the immediate vicinity of the corner. Furthermore, the thesis proposes modifications to the starting grid and pit lane road markings in accordance with the single seater dimensions for the 2026 season.