Research Topics for Prospective PhD Students

Department:
Branch of study:
Supervisor:
Date
Topic
Supervisor
Branch of study
Department
Date
Computer vision for autonomous cars
doc. Dr. Ing. Radim Šára
Computer Science
Department of Cybernetics
16. 5. 2015
In this project concepts and efficient computer vision methods will be applied to object detection problems and geometric computation for distance and velocity measurements from cameras and/or LiDARs mounted on a vehicle. The exact problem will be specified upon starting the PhD project. We will select a concrete scenario and use case for the sensoric system in an autonomous car or for a driver assistance system. We will define formal detection/measurement problems and design, implement, and test efficient real-time algorithms. This work will be done in close collaboration with a leading car manufacturer. More information at http://www.interactive-ip.eu/ http://up-drive.eu/ and http://cmp.felk.cvut.cz/~sara/
Computer vision methods for astronomical image processing
doc. Dr. Ing. Radim Šára
Computer Science
Department of Cybernetics
20. 5. 2016
The core of this project is detection of faint objects in optical astronomical images. Modern machine learning methods have not yet penetrated this field deep enough, although the problem is important, in particular when detecting small debris pieces on Earth orbit. Extremely low contrast and high level of noise make the problem difficult, especially in obtaining a detection certificate (posterior probability that an image object corresponds to a real object as opposed to a random configuration of image data). This project will develop efficient algorithms for statistical inference and suitable approximations to Bayesian model selection that can provide detections with such certificates. We plan collaboration with astronomers from the Japanese space agency JAXA.
Probabilistic Model Selection for Computer Vision Problems
doc. Dr. Ing. Radim Šára
Computer Science
Department of Cybernetics
3. 3. 2014
One of the fundamental problems in computer vision is the selection of a statistical model that explains an image or its part relevant to the visual task at hand. Such a mechanism is essential in a machine that “can see”, in the sense of object detection and scene understanding. In this project the student will study the problem of Bayesian model selection or a similar mechanism. The project will focus on structured model design, and on statistical inference and learning algorithms for such models. An application domain may be selected based on student’s preferences (e.g. parsing façade images for understanding street scenes, detection of structured objects in street traffic scenarios, faint object detection in astronomical images, etc).
More information at http://cmp.felk.cvut.cz/~sara/
Complex measures in internal medicine
doc. Ing. Daniel Novák, Ph.D.
Bioengineering
Department of Cybernetics
23. 4. 2019
Complexity measures can describe dynamic of physiological process such as heart beat or level of glycemia in blood. After publication of seminal paper in Nature predicting heart stoke, this theory found its way into common signal processing methods in clinical practice. The main aim of this topic is processing of neurological and electrocardiogram data to distinguished between different nuclei in basal ganglia or assess level of fragmentation of internal ECG signal [1,2,3].
[1] Eva M. Cirugeda–Roldan, D. Novak, V. Kremen, D. Cuesta–Frau, M.W. Keller, C. Schilling, O. Doessel, C. Schmitt, A. Luik, Characterization of Complex Fractionated Atrial Electrograms by Sample Entropy: An International Multi–Center Study, Entropy, 17(11), p.7493-7509,2015, IF
1.579
http://www.mdpi.com/1099-4300/17/11/7493/htm
[2] Andres Orozco-Duque, Daniel Novak, Vaclav Kremen and John Bustamante, Multifractal analysis for grading complex fractionated electrograms in atrial fibrillation, Physiological Measurement, Physiological Measurement, IF
1.808, 36(11), p. 2269-84,2015
http://www.ncbi.nlm.nih.gov/pubmed/26450345
[3] Juan Pablo Ugarte Macías; Andrés Orozco-Duque; Catalina Tobón; Vaclav Kremen; Daniel Novak; Javier Saiz; Tobias Oesterlein; Clauss Schmitt; Armin Luik; John Bustamante , Dynamic approximate entropy electroanatomic maps detect rotors in a simulated atrial fibrillation model. PLOS One, 9(12), 2014, IF
3.79
http://journals.plos.org/plosone/article?id
10.1371/journal.pone.0114577
Data processing and analysis in neuroscience
doc. Ing. Daniel Novák, Ph.D.
Bioengineering
Department of Cybernetics
23. 4. 2019
Computation neuroscience are one of the most developing areas in the field of physiological modeling being backed up nowadays by two big EU and USA initiatives. The goal of this topic is processing of neurological data, most of them having origin in human basal ganglia. Several sub-topics are addressed as correlation between objective neural recordings and so called UPDRS score, or modeling electrophysiology of basal ganglia [1,2,3].
More information can be found on neuro.felk.cvut.cz
[1] Tomas Sieger , Tereza Serranova, Pavel Vostatek, Jiri Wild, Daniela Stastna, Cecilia Bonnet, Daniel Novak, Evzen Ruzicka, Dusan Urgosik,Robert Jech, Distinct Populations of Neurons Respond to Emotional Valence and Arousal in the Human Subthalamic Nucleus, Proceedings of the National Academy of Sciences (PNAS), 112(10), p.3116-3121, 2015
[2] Sieger T, Bonnet C, Serranová T, Wild J, Novák D, Růžička F, Urgošík D, Růžička E, Gaymard B, Jech R., Basal Ganglia Neuronal Activity during Scanning Eye Movements in Parkinson’s Disease, PlosOne, 8(11), 2013, IF
3.79
http://journals.plos.org/plosone/article?id
10.1371/journal.pone.0078581
[3] J.Wild, Z.Prekopcsak, T.Sieger, D.Novak, ,R.Jech, Performance comparison of spike sorting algorithms for single-channel recordings, Journal of Neuroscience Methods, 203(2), p.369-76, 2012 , IF
2.114
http://www.ncbi.nlm.nih.gov/pubmed/22037595
Methods for active perception in partially observable environments.
doc. Ing. Karel Zimmermann, Ph.D.
Computer Science
Department of Cybernetics
14. 5. 2015
Accurate local 3D perception is an essential component for many fundamental capabilities such as emergency braking, active damping or self-localization from offline maps. Consequently, all autonomous vehicles require a sensor providing high resolution and long range 3D measurements. Since state-of-the-art rotating lidars are very expensive, heavy and contain moving parts prone to mechanical wear, several manufacturers have announced development of cheaper, lighter, smaller and motionless Solid State Lidars (SSL). SSLs should become available before the end of 2017 with target cost of $250 at automotive scale production, which make them affordable for ordinary cars in a close future.
In contrast to rotating lidars, the SSL can independently steer pulses of light by shifting and focus its attention on the parts of the scene important for the current task. Task-driven reactive control of hundreds of thousands rays per second using only an on-board computer is a challenging problem, which calls for highly efficient parallelizable algorithms.
We are looking for students who want to cooperate with us in the research of active mapping/segmentation/detection methods, which simultaneously (i) learns to reconstruct a dense 3D map from sparse depth measurements and (ii) optimize the reactive control of depth-measuring rays in order to minimize reconstruction error.
Další informace lze nalézt na http://cmp.felk.cvut.cz/~zimmerk/
Advanced methods of heterogeneous multidimensional data processing in electrophysiology
doc. Ing. Lenka Lhotská, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Current medicine, in particular electrophysiology, allow acquisition of large volumes of heterogeneous data and signals that must be evaluated in interrelated context. Research is focused on suitable methods of data and knowledge representation that serve for efficient storing and communication on one side and on advanced methods of processing that allow searching for mutual relations in data and reveal hidden information on the other side. Representation methods are based on requirements on semantic interoperability. Processing methods are inspired by advanced mathematical transforms, methods of digital signal processing and data mining methods.
Advanced methods of long-term multi-channel signal processing in neurosciences
doc. Ing. Lenka Lhotská, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
In neurosciences and neurology we meet more and more frequently long-term recordings (mostly so-called polygraphic), where in various channels different physiological signals and additional data are recorded, usually with different sampling frequency. Main aim is design and implementation of advanced methods of biomedical signal preprocessing and processing. The work will be performed using neurological and neurophysiological data provided by cooperating university hospitals. New approaches will be studied, as for example application of advanced data mining algorithms, including metaheuristics and integration of medical background knowledge in the decision support process. An important issue that must be considered is high interpersonal variability in the recordings, which poses a restriction on utilization of classical machine learning methods.
Decision support systems in medical domain
doc. Ing. Lenka Lhotská, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
With the advancement of health care information systems, electronic health records and acquisition of large volumes of data (big data), computer supported decision making is more required in medicine. It is necessary to consider medical guidelines, international coding systems, taxonomic structures, etc. The research aim is to design and develop relevant data models and create ontological structure on their top. Ontologies will allow representation of all relations in data and process areas in a suitable way. Another complex issue that deserves attention is ontology verification.
3D Scene Reconstruction from Images
doc. Ing. Tomáš Pajdla, Ph.D.
Computer Science
Department of Cybernetics
3. 3. 2014
3D scene reconstruction from images is a fundamental problem of computer vision. Our goal will be to advance the state of the art, publish at CVPR, ICCV, ECCV, in IJCV and in PAMI. We will collaborate with ETH Zurich, JR Graz, NII Tokyo, Google, Microsoft and Leica. The topic is suitable for students with interest in algorithms, experimental work, and engineering of really working systems. ½-1 year stay abroad expected.

people.ciirc.cvut.cz/~pajdla/
Advanced Algebraic Methods in Computer Vision
doc. Ing. Tomáš Pajdla, Ph.D.
Computer Science
Department of Cybernetics
3. 3. 2014
Algebraic techniques have proved very useful in solving difficult problems in geometry of computer vision. We will aim at studying more advanced elements of algebraic geometry and applying them to real engineering problems. We will publish in CVPR, ICCV, ECCV, in IJCV and in PAMI. We will collaborate with University of Washington, University of California in Berkeley, and INRIA. The topic is suitable for students with interest in applied mathematics. ½-1 year stay abroad expected.

people.ciirc.cvut.cz/~pajdla/
Image-based Scene Recognition and Visual Localization
doc. Ing. Tomáš Pajdla, Ph.D.
Computer Science
Department of Cybernetics
3. 3. 2014
Visual scene recognition and image based localization is an important problem in computer vision and machine learning. We will aim developing new approaches to place representation and its search. We will publish in CVPR, ICCV, ECCV, NIPS, in IJCV, PAMI. We will collaborate with INRIA Willow Paris, TiTech Tokyo and ETH Zurich. The topic is suitable for students with interest computer vision and machine learning applied to real engineering problems. ½-1 year stay abroad expected.

people.ciirc.cvut.cz/~pajdla/
Polynomial Optimization in Computer Vision and Robotics
doc. Ing. Tomáš Pajdla, Ph.D.
Computer Science
Department of Cybernetics
3. 3. 2014
Polynomial optimization techniques proved very useful in solving interesting problems in geometry of computer vision and robotics. We will aim at studying more polynomial optimization techniques and applying them in computer vision and robotics. We will publish in CVPR, ICCV, ECCV, ICRA, in IJCV, PAMI, IRR. We will collaborate with CNRS Toulouse, INRIA, University of Washington, University of California in Berkeley. The topic is suitable for students with interest in applied mathematics but used on real engineering problems. ½-1 year stay abroad expected.
people.ciirc.cvut.cz/~pajdla/
Body and peripersonal space representations of robots
doc. Ing. Tomáš Svoboda, Ph.D.
Computer Science
Department of Cybernetics
20. 4. 2016
In humans, interaction with the environment is mediated mainly by visual, auditory, and tactile sensations that need to be integrated with information about the current position and posture of the body and prior knowledge about its size and shape. In psychology and the neurosciences the key terms used are body schema, body image, and peripersonal space representations. In robots, multimodal sensing has received comparatively less attention and tactile information has been typically concentrated on the end-point only. With the advent of artificial tactile systems and their application to robots - covering not only the end-effector, but large areas of the body -, it is now possible to study multisensory integration that is necessary for safe interaction of robots with their surroundings, including humans. The goal of this work is to develop or newly apply machine learning methods (such as Restricted
Boltzmann Machine) that will warrant Bayes optimal behavior of robots in complex environments, in which physical contact of different body parts with other objects is unavoidable.

https://sites.google.com/site/matejhof/research/body-schema
Multimodal data analysis for autonomous systems
doc. Ing. Tomáš Svoboda, Ph.D.
Computer Science
Department of Cybernetics
3. 12. 2018
Multimodal data like RGB images, depth images, thermal images, Lidar measurements are vital for autonomous systems like robots or cars. We aim at combining various modalities optimally, with a particular interest in possibly missing data. For machine learning methods we use, it is necessary to create large realistic dataset in an automated way. Many simulators exist however, generating realistic multimodal data is still a research challenge we want to address.
Algorithms for Large-scale Optimization in Computer Vision and Machine Learning
doc. Ing. Tomáš Werner, Ph.D.
Computer Science
Department of Cybernetics
24. 7. 2018
Optimization is ubiquitous in nowadays computer vision and machine learning, examples being energy minimization in
computer vision, inference in probabilistic graphical models, or training of classiťrers (such as SVM or deep neural
networks). A recent feature ofthese tasks is their size - they can easily have millions ofvariables and (sparse) constraints.
Due to this, classical optimization theory and algorithms are often inadequate. For instance, popular simplex and interior
point methods are inapplicable to linear programs ofthat size already for their superJinear space complexity. Therefore,
large-scale distributed optimization algorithms muts be used, such as coordinate minimization minimization, alternating
direction method of multipliers (ADMM) or stochastic gradient descent. The aim of the proposed doctoral thesis is to develop
novel theory and algorithms for large- scale optimization in computer vision and machine leaming, with emphasís on
coordinate minimization.
Structured Statistical Models for Image Analysis
doc. Ing. Tomáš Werner, Ph.D.
Computer Science
Department of Cybernetics
12. 6. 2014
In machine learning and computer vision, structural statistical models of images describe collections of images with certain characteristics in a concise way, which then facilitate tasks like image segmentation, scene reconstruction, and object detection/recognition. Two classes of such models have been widely successful: graphical models and pictorial structures. The student should design and analyze new models within these two classes (possibly combining them) and algorithms for inference and learning.
Publications on prestigious international conferences are expected.
http://cmp.felk.cvut.cz/~werner/
Machine Learning in Learning Analytics
doc. Ing. Zdeněk Zdráhal, CSc.
Computer Science
Department of Cybernetics
23. 10. 2018
Learning analytics refers to the measurement, collection, analysis and reporting of data about the progress of learners and the contexts in which leaming takes place. The thesis goal is an investigation of machine learning techniques when applied to the data collected in Virtual Learning Environments and the identification of the behavioural patterns created by students at risk of failing. This will allow the educational institutions to optimize the process of learning using virtual e-leaming platforms. Important application area of the thesis is an analysis of data from traditional universities in order to improve the understanding of study process and detect the students in need of additional pedagogical support.
Visual search in large image collections
doc. Mgr. Ondřej Chum, Ph.D.
Computer Science
Department of Cybernetics
3. 3. 2014
This topic adresses efficient methods of visual search (query by example). Different aspects, such as efficient indexing, feature detection and description, image representation, similarity measures, model learning, and other will be will be studied and developed.

cmp.felk.cvut.cz/~chum/
Advanced decomposition of biosignals
doc. MUDr. Jakub Otáhal, PhD
Bioengineering
Department of Cybernetics
23. 4. 2019
Biological signals (EEG, EMG etc.) usually contain summation of signals from several sources. For instance recordings of electrical activity of skeletal muscle for body surface consists of signals arising from individual motor units (an aggregate of muscle fibers innervated by the same neuron). Single muscle thus contain tens to hundreds of such motor units. The aim of the thesis is to develop robust algorithms allowing identification of activity of individual sources. The results will be used in diagnostics of severe neuromuscular disorders but also in new approaches in myoelectric prosthesis control, human machine interfaces and top level sport.
Algorithms for analysis of vascular changes from microscopy images
doc. MUDr. Jakub Otáhal, PhD
Bioengineering
Department of Cybernetics
23. 4. 2019
Function of the vascular system is to cover metabolic needs of body tissues and organs to maintain their proper functionality. Blood flow through the tissue is variable and highly regulated by several physiological mechanisms controlling diameter of precapillaries. Such regulation controls immediate capillary blood flow. Lasting insufficient blood supply induces growth of new capillaries especially in areas with low levels of oxygens. This process take place in tumors or tissues with high activity like in epileptic foci. The aim of the thesis is to develop algorithms for detecting changes in capillary bed from time series of microscopy images. The results will be used in diagnostics of cancer or epilepsy.
Advanced methods for unstructered data processing
Ing. Jiří Kubalík, Ph.D.
Computer Science
Department of Cybernetics
26. 5. 2016
Today the Internet, social networks, and others generate large amounts of data. The Internet allows us to collect, store, and process large quantities of unstructured data. Current advanced information retrieval and extraction systems do not offer a simple way to process the knowledge contained in these data. The research will focus on methods for data and knowledge representation, which allow efficient storage and processing. In addition, we will focus on the advanced processing methods that discover the semantics, relationships, and uncover hidden information. New methods will be inspired by advanced mathematical transformations, methods, knowledge representation, RDF database, and data mining methods.
Advanced methods of heterogeneous multidimensional data processing in electrophysiology
Ing. Jiří Kubalík, Ph.D.
Computer Science
Department of Cybernetics
28. 5. 2014
Current health system, especially in the area of electrophysiology, enables to collect large amount of heterogeneous data and signals. The data and signals have to be evaluated in proper context. This research is therefore focused to develop appropriate methods of data and knowledge representation that allows efficient storage and processing of these data and signals. Extra attention is given to the advanced processing methods that enable to search for mutual relation and connectivity in the data and reveal hidden information. Such methods are based on requirements for semantic interoperability and are inspired by advanced mathematical transformations, digital signal processing algorithms and data-mining techniques.
Advanced world models and representations for inteligent mobile robots. Building and keeping world models under uncertainty and their usage for robot activity/mission planning. Models and maps of larg
Ing. Libor Přeučil, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Models of the world in robotics represent repositories of geometric and/or topological information about the surrounding world of the robot. Herein, the models are primarily created to serve as knowledge base for robot mobility accomplishment - in particular for robot localization and planning as well as its’ navigation through the environment to fulfill the given task.
The necessity to keep the model valid addresses problems to assure extraction of problem-relevant knowledge from the sensory data and methods to update content of the model on a runtime basis. Many real-life cases need to cope with incomplete world representation, or models with introduced uncertainty.
In real situations, the afore represents the backbone task of exploration – intentional discovery of unknown environments, or robot navigation through partly known/highly variable environments. In these terms, the world model design has to take into account the type, size and desired resolution of the robot operating space (i.e. outdoor, indoor, strictly controlled or human oriented/uncontrolled environments, sparse or dense models, etc.)
The topic targets analysis and elaboration of various types of world representations and models, and their keeping, to allow efficient autonomous operation of mobile robots in various types of environments and to serve for diverse tasks.
More at: http://imr.ciirc.cvut.cz
Collective robotics. Cooordinated and cooperating robots of UAV and UGV type, robot swarms and formations. Resousre fusion and multi-robot systems for efficient localization, navigation and mapping t
Ing. Libor Přeučil, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Multi-robot systems may offer substantial improvements of performance in comparison to single robot operation. This allows i.e. better performing sensing systems, use of low-end sensors, or fostered manipulation abilities. Another possibility to improve performance of multi-robot systems is to embed (mainly nature-inspired) capabilities into robot units, The are rules, that assure various usable behaviors through very simple algorithms and low cost HW. The embedded functionalities may create i.e. reflex-like behaviors, allow robot swarming and evolution in cooperating groups of robots, and many more. Therefore, such approaches allow attaining specific functionalities, which may be hard to obtain using classical methods. Besides, these technologies bring scalability of multi-robot systems (swarms), create and improve robustness of multi-robot systems in use against a unit malfunction, stand-in behaviors, etc.).
Subjected research addresses techniques of multi-robot control and autonomy based on advanced principles of collective and cooperative robotics (robot formations and swarms) in UAV(MAV) and UGV domains.
More at: http://imr.ciirc.cvut.cz
Planning and scheduling for mobile robotics, strategies for robot navigation and operation in realistic environments, single robots and multi-robot systems.
Ing. Libor Přeučil, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Mobile robot planning and scheduling comprises mainly the core task of trajectory generation (planning) and their chaining (scheduling) to attain the given goal (target position of the robot). In the case of mobile robots, the tasks are posed into locomotion problems resolution at given constrains and even uncertainty in terms of a varying environment. In general, these tasks are of high computational intensity and therefore need to be simplified and approximated. Similarly, the scheduling tasks (allowing more complex problem solution as exploration, or surveillance and inspection tasks (possibly of (M)TSP complexity), etc.) exhibit hardness of an NP-complete problem and may be non-solvable in its’ pure form.
The goal of the research is to elaborate and develop new and fast planning and scheduling algorithms that approximate computationally intensive (and even the NP-hard) problems and deliver a usable approximate results in finite time (close to real-time) for practical problem solution in mobile robotics.
More at: http://imr.ciirc.cvut.cz
Robot control at malfuction states. Advanced failure detection and failure recovery methods. Long-term autonomy.
Ing. Libor Přeučil, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Autonomous robots, that act in real environments may fail to accomplish their task due to high complexity of the problem, diversity of conditions, unexpected event occurrence, and uncertainty in general. These cases typically end-up in a robot being trapped by environmental structures, dropping into cycles, etc. Such undesired situation needs to be detected and resolved. The malfunction situations may be recognized via analysis and classification of “act-and-response” behaviors of the given robot, since recovery from these cases may be obtained by a combination and/or learning of diverse fall-back procedures. The whole procedure leads to adaptive re-planning to obtain a self-recovery functionality of the malfunctioning robot, or a group of robots. Targeting of this topic is to research, design and develop advanced approaches to detect and resolve functional failures of autonomous robots, that are to exhibit life-long autonomy in realistic environments and problems.
More at: http://imr.ciirc.cvut.cz
Sensor fusion and SLAM methods and tools for building world models in mobile robotics. Robust approaches to mobile robot navigation in general (outdoor) environments. Mobile robot vision.
Ing. Libor Přeučil, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Sensor fusion and simultaneous localization and mapping (SLAM) belong to basic methods in robotics. These methods mainly serve for processing of the gathered sensory data and allow localization of the robot in the environment. The primary goal of the sensor fusion is to adjust sensing failures and noise of the respective sensor, whereas the SLAM optimizes robots’ position determination and helps to keep the world model up-to-date. The task addresses fusion of knowledge and recovery of the robot position to derive major geometric properties of the operating environment, i.e. its’ shape and structure and therefore to build up the world model. The research targets further development of advanced methods and tools for sensor fusion, world model building and keeping - all based on standard mobile robot sensors (LIDAR, UWB RADAR, IMU, ordinary camera, RGBD camera, etc. )
More at: http://imr.ciirc.cvut.cz
Unmanned and autonomous flying robots. Advanced control of UAV systems in real, constrained and obstructed environments (indoor and outdoor). Problems of localization-navigation and maping with limite
Ing. Libor Přeučil, CSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Unmanned aerial systems (UAV), or their “micro” outfit (MAV), belong to robotic platforms, that allow addressing robotic challenges in 3D space. Nevertheless, due to current technology state-of-the-art, aerial robotics exhibits certain specifics and limitations as: limited operating time of the UAV, limit on payload, that constrains use of sensing systems and computational equipment to light-weight and not very powerful systems, as well as exhibits constrained maneuvering capabilities. The goal herein is elaboration and research in advanced methods and tools (ranging from algorithms and SW solutions to specialized HW design) for onboard sensing, navigation, mapping and collision avoidance systems for aerial robots. The ultimate goal is to target optimization of the suggested solutions for UAVs, or MAVs, that allow creating of robot swarms, collision-free operation of aerial robots in tight spaces, cooperation with ground robots (UGV), etc.
More at: http://imr.ciirc.cvut.cz
Active machine learning for improving efficiency of long time series annotation
Ing. Martin Macaš, Ph.D.
Computer Science
Department of Cybernetics
19. 5. 2017
An expert annotation of long signals can be very time and money consuming.
It can also cause lack of experts and reduction of annotation quality because of
expert's fatigue or replacement of expert by a less skilled annotator.
The whole annotation process can be improved by semi-automatic methods using
an active learning paradigm. The active learning can be used in terms of selection
of some signal segments to be annotated and subsequent automatic annotation
of the rest of the signal. Another approach can be detection of erroneously
annotated parts of the signal and a query for their re-annotation. The thesis should
identify and solve selected crucial issues of active learning in real domain like
unbalanced character of classes, lack of knowledge of ground truth in testing
or importance of temporal or another context. Typical example of suitable application
is annotation of sleep EEG signals. However, proposed methods can be evaluated using any application domain.
Context aware active learning
Ing. Martin Macaš, Ph.D.
Computer Science
Department of Cybernetics
3. 6. 2016
The thesis wil focus on the usage of (mainly temporal) contextual information within active learning and optimal experimental design for data mining and machine learning methods. Instances of temporal context data typically violate the assumption of being independently and identically distributed. They exhibit significant sequential interdependencies, where instances that are close to each other in time are likely to be related to each other. In part-of-speech tagging, sequence of parts of speech are constrained by the grammar of the language. In EEG sleep staging, the transitions are constrained by physiological properties of the human brain. In prediction of thermal comfort, it is unlikely that a person would change its thermal comfort feeling from 'too hot' to 'too cold'. Those interdependencies and correlations can be exploited by the predictor and improve its accuracy. Different ways of active generation of such temporal training data will be proposed and analysed. One example is active annotation, where predictor’s inputs are selected or generated for which it is worth to measure an output. An issue is that such procedure can hardly benefit from knowledge about past or future annotations that are mostly unknown. Another example is determination of time moments suitable for queriing annotation of a complex dynamic process. For example, an efficiency of learning of hypoglycemia predictor in diabetes depends on selection of time instans, where the glycemia measurements (annotation) occurs. The thesis should provide a survey of different formulations of such temporal context aware active learning, analyse state of the art within this area, propose novel approaches and make some empirical and theoretical evaluations. Although the temporal context will be mainly focused, other information that can be used to characterize the situation of an entity will be also considered (e.g. spatial location, user’s identity, performed activity, or relations).
Cooperative sensing by group of unmanned aerial vehicles in environment with obstacles
Ing. Martin Saska, Dr. rer. nat.
Computer Science
Department of Cybernetics
14. 5. 2015
Progress in design of methods for stabilization of unmanned helicopters allows their deployment in specific robotic scenarios in real-world conditions. Aerial vehicles are mostly used for carrying standard robotic sensors (cameras, rang-finders) and also specialized sensory equipment (measurement of radiation, pollution). Utilization of small cooperating Micro Aerial Vehicles (MAVs) brings possibility of distributed measurement by heterogeneous sensors with extended fault-tolerance, while keeping constraints given by limited payload of MAVs.
This thesis will be aimed at continuation of research conducted by Multi-robot Systems group in the field of control and coordination of MAV groups. In particular, theory and methodology required for cooperative sensing by teams of MAVs will be designed and suited for multi-robot applications in environment with densely distributed obstacles, where MAVs bring added value in comparison with available systems of unmanned airplanes. Finally, conditions of stability of the MAV group will be theoretically and experimentally studied based on characteristic of environment (density of obstacles) and range of onboard sensors.
Requirements: practical and theoretical experience with MAVs, knowledge of programming in C, good knowledge of mathematics.
Motion planning for formations of micro aerial vehicles
Ing. Martin Saska, Dr. rer. nat.
Computer Science
Department of Cybernetics
31. 12. 2014
Recent progress in development and miniaturization of micro unmanned vehicles (MAVs) allows us to consider deployment of large systems of simple helicopters for cooperative solving a given task. State-of-the-art methods are limited to stabilization of MAV groups in laboratories equipped with precise external localization system in control feedback. Nevertheless, the task of stabilization and control of MAVs in real-world conditions requires developing of new methodology that relies on onboard sensors and onboard computational resources.
The aim of this PhD thesis is to continue with research of our group in the field of motion planning and optimal control for formations with compact shape. The designed methodology and consequently the implemented system should enable planning of complex maneuvers for MAV formations in demanding environment with dynamic obstacles. PhD student will focus on theoretical and practical determination of requirements on stability and convergence of the system, taking into account sensors being used for relative localization of team members.
Requirements: knowledge of programming in C and/or in MATLAB, good knowledge of mathematics, experience with MAVs is an advantage.

http://imr.felk.cvut.cz/People/Martin
Multi-robot persistent monitoring and spatial coverage
Ing. Martin Saska, Dr. rer. nat.
Computer Science
Department of Cybernetics
5. 1. 2016
Persistent monitoring of large environments by mobile robots is one of the most promising robotic applications. It is expected that the robots will be deployed in these scenarios in different complexity, from compact formations of closely cooperating robots to large swarms that enable information gathering simultaneously in different locations. These topics become attractive nowadays mainly due to the recent progress in development and miniaturization of Micro Aerial Vehicles (MAVs) and other aerial and ground vehicles.
The thesis will be aimed at theoretical research of challenges motivated by deployment of large groups of robots in these long-term applications and by using limited onboard computational resources. Onboard processing is crucial for achieving full autonomy. In addition to the optimal coverage and coordination tasks, new strategies for optimal usage of redundant robots with limited operational time and for their replacement in case of failures will be investigated. Theoretical analyses of designed principles experimentally verified in relevant multi-robot scenarios will be part of the thesis. The experimental verification of theoretical results will be realized with MAVs of Multi-Robot Systems group and with other robots of Departments of Cybernetics and Computer Science and Engineering within the Center for Robotics and Autonomous Systems.
Requirements: experience with robotics, knowledge of programming in C and in MATLAB, excellent knowledge of mathematics, experience with MAVs is an advantage.
Stabilization and control of swarms of micro aerial vehicles
Ing. Martin Saska, Dr. rer. nat.
Computer Science
Department of Cybernetics
18. 12. 2014
Swarms of low-cost micro unmanned vehicles (MAVs) may be efficiently deployed in numerous security and other commercial applications, such as e.g. robotic surveillance, search and rescue, mapping, and environment monitoring. A group of simple MAVs is usually able to cooperatively accomplish given tasks more effectively (faster and mainly more reliable), than if using a better equipped and in total more expensive single helicopter.
The aim of this PhD thesis is to design a mechanism for control and stabilization of MAV groups, which would fully exploit all key properties of swarm behavior, such as possibility of redundancy, interchangeability of swarm members, and decentralized control, with the aim to increase reliability and decrease requirements on communication. One of the possible approaches is to utilize rules of swarm behavior observed in nature, since the sense organs of animals in swarms may be described by a similar model as the sensors that can be used onboard of MAVs. The important part of the student’s work will be theoretical and experimental analysis of swarm stability in real-world conditions.
Requirements: knowledge of programming in C and/or in MATLAB, good knowledge of mathematics, experience with MAVs is an advantage.

http://imr.felk.cvut.cz/People/Martin
Visual relative localization and stabilization of groups of unmanned helicopters
Ing. Martin Saska, Dr. rer. nat.
Computer Science
Department of Cybernetics
24. 6. 2016
The topic is focused on design of a robust visual localization of neighboring swarm members without a necessity of placing artificial identification tags on particular unmanned aerial vehicles (UAVs). The goal of the work is design of a proper methodology for autonomous detection of moving objects in front of a moving background. The research will be targeted at sufficiently fast processing of images using limited computational power carried by UAVs aimed at direct integration of outputs into the feedback control. Beyond conventional approaches of visual detection and object tracking, coordination and cooperation of swarm members will be employed to increase precision of the relative localization and its robustness. Beside the research of real time embedded solution of onboard relative localization in control feedback, this topic enables an interesting multidisciplinary research of behavior of swarms in nature. It is very difficult or impossible to study influence of changing perception of swarm members on their collective behavior in nature. The proposed system of visual relative localization enables to emulate these changes and to study their impact in artificial swarms controlled by rules observed in nature.

Requirements: practical and theoretical experience with computer vision, knowledge of programming in C, good knowledge of mathematics.
Machine learning methods for malware detection
Ing. Vojtěch Franc, Ph.D.
Computer Science
Department of Cybernetics
20. 6. 2015
The topic of the thesis is motivated by a real-life problem of detecting
malicious behavior in a computer network communication based on monitoring data flows between a client computer and a server. The ultimate goal is to create and to update such detectors automatically from data with a minimal human intervention. The main obstacle preventing us from usage of conventional machine learning methods is the lack of sufficiently representative sample of legitimate
and malicious network communication. The creation and maintainance of an
annotated network communication requires a tedious forensic analysis done by computer security experts which does not scale with the number of constantly evolving malware. The main topic of the thesis will be learning of the detectors from easily available weakly annotated data, for example obtained from black-lists of Internet domains, and exploitation of active learning methods that can significantly reduce the number of precisely annotated training samples.
Collaborative robots, artificial touch, and automatic self-calibration
Mgr. Matěj Hoffmann, Ph.D.
Computer Science
Department of Cybernetics
22. 10. 2018
As robots are leaving safety fences and starting to share workspaces and even living spaces with humans, they need to dynamically adapt to unpredictable interactions with people and guarantee safety at every moment. On the rapidly growing market for collaborative robots, safety is ensured through specific technologies such as force limitations by design or contact detection and stopping relying on force measurements. Humans, however, possess awareness of their body in space drawing on dynamic, context-dependent fusion of multimodal sensory information, which makes them adaptive, flexible, and versatile. We will add important new dimensions to physical human-robot interaction: (1) we will use artificial electronic skins, developed only recently; (2) we will extend the skin space to the space around it, giving rise to a protective safety margin following the body parts in space - inspired by the peripersonal space in humans; (3) due to the multimodal nature of our approach, redundant information about the robot itself and its environment is typically available, which facilitates continuous self-calibration as well as reasoning with uncertainty. We're collaborating with leading research labs in Europe as well as industrial partners (e.g., KUKA Corporate Research, Airskin).

https://sites.google.com/site/matejhof/research/cobots-and-hri
https://sites.google.com/site/matejhof/research/touch-and-selfcalibration
http://cmp.felk.cvut.cz/projects/body-schema
Embodied computational models of body representations in primate brains and their development, with applications in robot self-calibration and human-robot interaction
Mgr. Matěj Hoffmann, Ph.D.
Computer Science
Department of Cybernetics
6. 9. 2018
How do babies learn about their bodies? Newborns probably do not have a holistic perception of their body; instead they are starting to pick up correlations in the streams of individual sensory modalities (in particular visual, tactile, proprioceptive). The structure in these streams allows them to learn the first models of their bodies in space. The mechanisms behind these processes are largely unclear. In collaboration with developmental and cognitive psychologists and neuroscientists (New Orleans, Paris, Bielefeld), we want to shed more light on this topic: through implementations in humanoid robots, the goal is to develop concrete, embodied models of the development of multimodal (tactileproprioceptive-visual) representations of the body and the space surrounding it. Access to humanoid robots with wholebody artificial sensitive skin provides a key enabling technology. At the same time, the algorithms developed pave the way for robots with "whole-body awareness" and find applications in automatic robot calibration and safe physical human-robot interaction.
Haptic exploration and categorization of objects using robotic grippers
Mgr. Matěj Hoffmann, Ph.D.
Computer Science
Department of Cybernetics
20. 9. 2019
The goal is to use different robotic arms and grippers to explore various objects and collect data from proprioceptive, tactile, and force feedback. Different clustering or classification algorithms will be employed on this data to differentiate between the objects, focusing in particular on properties that can only be extracted from haptic exploration (manipulating the objects) such as stiffness, elasticity, etc. In a second step, the choice of manual actions that aid recognition will be studied. Finally, priors extracted from vision or other sources (Internet – linguistic description) as well as visual feedback (normal or RGB-D cameras) can be also employed.
The main research questions are as follows:
1) What are the optimal exploratory actions for a given object, gripper and the object properties we want to learn about?
2) How are the object properties that can be learned dependent on (i) the gripper used, (ii) the exploratory action (sequence)?
Biomedical Image Analysis Algorithms
prof. Dr. Ing. Jan Kybic
Bioengineering
Department of Cybernetics
23. 4. 2019
The amount of data produced in medicine and biology increases steadily. The bottleneck is no longer the availability of the acquisition device but the ability to evaluate the data produced. Hence, the space for automatic image analysis algorithms is growing.
The PhD student will develop new algorithms for image analysis, especially for the tasks of registration, segmentation and classification. We will work with data provided by clinicians and biologists and solve their specific problems. We emphasize the algorithmic side of the problem. Potential candidates should have good knowledge of programming, mathematics, and at least basic knowledge of techniques of signal and image processing.

http://cmp.felk.cvut.cz/~kybic
Detection and recognition of text „in the wild“.
prof. Ing. Jiří Matas, Ph.D.
Computer Science
Department of Cybernetics
3. 3. 2014
The problem of text detection an recognition has many application. The open problems include: reliable detection of characters and word in any script, font and language; grouping of strokes, characters and words into a linear sequences, and modelling and detection of higher level structures such as blocks and displays.

The thesis will focus on one of the open problems listed above.

http://cmp.felk.cvut.cz/~matas/
Recognition of objects in images and videos
prof. Ing. Jiří Matas, Ph.D.
Computer Science
Department of Cybernetics
3. 3. 2014
Object recognition encompasses a broad range of problems ranging from two view matching of images of a given object, image retrieval and discovery of objects in collections of images and perhaps the most challenging problem of categorisation.

The thesis will focus on one of the open problems listed above.

http://cmp.felk.cvut.cz/~matas/
Visual tracking, motion estimation and segmentation
prof. Ing. Jiří Matas, Ph.D.
Computer Science
Department of Cybernetics
16. 1. 2018
Visual tracking is a broad area of research that deals with estimating the state of one or more a priori know or unknown objects in sequences captured by one or more cameras with overlapping on non-overlapping fields of view. The state might be as simple as a single point (estimating location), a rectangle – an approximation of the segmentation, location, scale and possibly orientation, full pixel-wise segmentation, or a per-pixel displacement field, especially in the case of articulated or deformable objects.
The problem can be restricted to sequences without significant occlusion (short-term tracking); the more general setting consider cases where the object leaves the field of view or is fully occluded (long-term tracking). If the object is not known a priori (model-free tracking), the problem entails learning the model of the tracked object, typically its appearance but potentially its shape as well. There are many open problems in visual tracking; the field is highly active with hundreds of papers published at major conference and top journals.
Logic of Quantum Systems
prof. Ing. Mirko Navara, DrSc.
Computer Science
Department of Cybernetics
3. 3. 2014
The classical probability theory studies only those random events which admit unlimited repetition. This assumption is violated not only in quantum theory, but also in sociology, psychology, medicine, etc. The corresponding mathematical theory brings many difficulties. Specialized computer algebra programs play an important role in this study. Thus the topic is appropriate both for theoreticians and for programmers capable of implementation of new procedures.
Non-classical Probability and Statistics
prof. Ing. Mirko Navara, DrSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Probability and statistics study a relatively narrow field of applications satisfying the assumptions of the classical theory. Among others, they assume repeatable experiments with yes-no results described by the classical logic. In many applications, these assumptions are violated and the classical approach fails. The development of mathematical tools handling these more general situations meets many problems. It is desirable to contribute by building new mathematical theories, as well as their applications in sociology, psychology, medicine, etc.
Principles of Fuzzy Control
prof. Ing. Mirko Navara, DrSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Fuzzy control has a lot of successful applications, sometimes restricted by only a partial knowledge of the underlying mathematical theory. A detailed analysis of the properties of fuzzy controllers leads to improvement of quality of control. This topic combines the mathematical and engineering approach, thus is it particularly convenient for those students of electrical engineering who are keen of mathematics.
Properties of Fuzzy Logical Operations
prof. Ing. Mirko Navara, DrSc.
Computer Science
Department of Cybernetics
3. 3. 2014
Fuzzy logic has a lot of successful applications, sometimes limited by our knowledge of the mathematical background; many theoretical questions remained open. New results and tools (including computer programs) enable a progress. This topic combines the mathematical and engineering approach with the use of modern computational tools.
Machine learning for information retrieval
prof. Ing. Václav Hlaváč, CSc.
Computer Science
Department of Cybernetics
5. 9. 2015
Machine learning for information retrieval,. anomaly detection, identification of items, events or observations which do not conform to an expected pattern or other items in a dataset, focusing on rare or burst occurrences using semi-supervised models traning and techniques to overcome the inherent unbalanced nature of datasets.
Perception in the force/torque compliant robotic manipulators
prof. Ing. Václav Hlaváč, CSc.
Computer Science
Department of Cybernetics
6. 6. 2019
Force/torque compliant robotic manipulators are the prerequisite of safe human-robot collaboration. The topic addresses the feedback (servoing) issues relying on computer vision and tactile sensing also in robotic manipulation context. The topic fits the Industry 4.0 paradigm too.
Qualitative reasoning under uncertainty in robotics context
prof. Ing. Václav Hlaváč, CSc.
Computer Science
Department of Cybernetics
6. 6. 2019
The intelligence of autonomous vehicles and in a robot manipulation relies on the representation of the robot world and reasoning in it. The mathematical logic-based reasoning in this context is the topic for the prospective Ph.D. thesis.
Advanced information retrieval
prof. Ing. Vladimír Mařík, DrSc., dr. h. c.
Computer Science
Department of Cybernetics
23. 5. 2014
Information retrieval is one of the very important Internet search technologies. It solves the problem of obtaining information resources relevant to an information need. It includes lot of other sub-problems such as learning to rank, meta search, recommendation, learning representations, exploiting large unlabelled data sets, information merging, etc. The thesis will study the new field of exploiting unstructured data sets to improve existing algorithms and use them in a new contexts.

http://isg.felk.cvut.cz/
Conversational AI
prof. Ing. Vladimír Mařík, DrSc., dr. h. c.
Computer Science
Department of Cybernetics
18. 1. 2018
With the advance of speech recognition and text to speech the Conversational AI became a very important research topic. The Conversational AI research includes application and development of new machine learning algorithms for all aspects of natural text processing i.e. text data acquisition, information retrieval and extraction, unstructured to structured text conversion, dialog management, question answering, semantic text similarity, natural language generation, words and text representation etc. The agent-based systems and relevant architectures (up-to-now used in industrial CPS and Electric Grids) will be used for the design of the corresponding architectures; the experience with SOA in industrial agent-based systems will be leveraged. Each task of Conversational AI (information retrieval, question answering etc.) will be represented by an autonomous module.
Ing. Jan Šedivý, CSc. is expected to serve as a supervisor-specialist as the topic is quite novel, on the borderline of the latest approaches.
Personalized Serious Games in Therapy, Training and Rehabilitation
prof. RNDr. Olga Štěpánková, CSc.
Computer Science
Department of Cybernetics
4. 12. 2014
Observation of patient´s reaction to specific situations and his/her behaviour during the course of problem solving provide valuable data that is often used for diagnostic purposes. If such data are collected automatically in a digital form, they can be interpreted automatically not only to simplify diagnostics but to serve as an input in a feedback loop that ensures highly personalized rehabilitation. Game-like scenarios used for this purpose are referred to as serious games. One type of a game can be applied for very diverse purposes (to strengthen muscular function or to stimulate brain functions) and by users with varying requirements on complexity of the user interface. Suggest a methodology for description of an ontology based user profile that makes it possible to create efficiently personalized version of a serious game through adaptation of its user interface and game complexity to match the needs of the user. The advantages of the suggested approach should be verified in a real application tested on several types of patients.

https://nit.felk.cvut.cz/drupal/
Navigation and planning for complex robots
RNDr. Miroslav Kulich, Ph.D.
Computer Science
Department of Cybernetics
4. 5. 2017
The topic is focused on navigation, trajectory planning, and high-level decision making for pure robotic or human-robot systems in high-dimensional state spaces. High dimension of these problems is caused by many degrees of freedom of robots, and/or by a number of involved robots. The PhD student will research and develop new algorithms for processing, analysis, and fusion of data produced by chosen sensors (2D and 3D range sensors or cameras), fast and robust navigation based on these data, and efficient planning in tasks like multi-robot surveillance and reconnaissance, formation keeping, dexterous cooperative manipulation or similar. Potential candidates should have good knowledge of programming, mathematics, and at least basic knowledge of techniques of data (signal and image) processing.
Routing problems in mobile robotics
RNDr. Miroslav Kulich, Ph.D.
Computer Science
Department of Cybernetics
4. 5. 2017
Mobile robots, as their autonomy increases, can be employed in more and more complex scenarios. A mission goal in these scenarios is not specified in the form 'go from A to B', but the task incorporates specification of particular sub-goals to be reached by a robot successively as well as determination of the order in which to reach the sub-goals. Typical tasks include inspection of a priori known environment, exploration of an unknown environment, or search for an object of interest. These tasks lead to solution of NP-hard optimization problems, for which many approaches exist. Although these approaches developed mainly by the operational research community generate feasible solutions, their complexity is high and thus not suitable for real-time decision making in mobile robotics.
The thesis will focus on research and development of novel approximation methods for these optimization problems, which will provide good quality results in a reasonable time and their application to routing problems in mobile robotics.
Data fusion in areal robotics
RNDr. Petr Štěpán, Ph.D.
Computer Science
Department of Cybernetics
12. 9. 2019
Multimodal data like RGB images, depth images, thermal images, Lidar measurements are vital for autonomous systems like aerial robots. The UAV, especially small UAV robots, are limited by sensors and computers weight, so new approaches for data fusion and data preprocessing are necessary. The specialised HW accelerators for Deep NN can be used together with improved data fusion methods.
Onboard localization of aerial robots
RNDr. Petr Štěpán, Ph.D.
Computer Science
Department of Cybernetics
12. 9. 2019
Design of new algorithms and techniques for onboard localization of autonomous aerial robots. The algorithms have to use only sensors, that are suitable for aerial robots. Another limit comes from limited computation power and storage sources of aerial robots.
Weakly supervised learning for visual recognition
Šivic, Josef, Dr. Ing.
Computer Science
Department of Cybernetics
11. 6. 2019
Building machines that can automatically understand complex visual inputs is one of the central problems in artificial intelligence with applications in autonomous robotics, automatic manufacturing or healthcare. The problem is difficult due to the large variability of the visual world. The recent successes are, in large part, due to a combination of learnable visual representations based on convolutional neural networks, supervised machine learning techniques and large-scale Internet image collections. The next fundamental challenge lies in developing visual representations that do not require full supervision in the form of inputs and target outputs, but are instead learnable from only weak supervision that is noisy and only partially annotated data. This thesis will address this challenge.
More details are on http://impact.ciirc.cvut.cz/
Learning for Conversation. Argumentation and Reasoning
Urban, Josef, Mgr., Ph.D.
Computer Science
Department of Cybernetics
11. 6. 2019
The goal of the work is to research combinat ions of Natural Language Processing, Conversational A1 and Automated Reasoning. The Ph.D. candidate will develop conversation, argumentation and reasoning methods to automatically create con versa tional assistants from an original text, mainly manuals that interactively help users with product configuration, trou bleshooting, etc. The research and development will include text processing, creating questions from unstructured text, disambiguation, summarization, query answering and reasoning, and related tasks.
Machine learning for saturation-based theorem proving
Urban, Josef, Mgr., Ph.D.
Computer Science
Department of Cybernetics
11. 6. 2019
The goal of the project is to investigate the topic of combining machine learning methods with state-of-the-art saturation-based Automated Theorem Provers. The Ph.D. candidate will apply modern machine learning techniques such as deep neural networks and reinforcement learning to improve the quality of prover's decisions at key heuristic choice points such as premise selection, clause selection, and the choice of strategy parameters. The work includes the design of new learning architectures particularly suited for dealing with logical formulas and related objects.
Responsible person: RNDr. Patrik Mottl, Ph.D.