Dissertation theses
Synthesis of artificial intelligence and machine learning models to programmable hardware
Supervisor
Level
Topic of dissertation thesis
Topic description
Artificial intelligence and machine learning are increasingly used in real applications and thus significantly penetrates the field of embedded and especially cyberphysical systems. A typical example is the image analysis subsystem for self-driving cars or intelligent sensors. Unlike the big data area, these systems have limited computing power and have high variability in hardware architecture, and in addition, these systems are subject to additional requirements, especially for real-time processing, security and reliability, explainability, power consumption and chip size. From these perspectives, the resulting implementation must be optimized to minimize hardware resources while maintaining functionality and reliability to meet economic goals. Automated conversion of machine learning models such as deep neural networks into ASIC hardware and especially programmable hardware (FPGA) is currently a very current topic.
The subject of the proposed topic is the research of algorithms, procedures, methodologies and tools for the synthesis of artificial intelligence models and machine learning into programmable hardware. Current research is in the field of approximate calculations, hardware optimization using limited accuracy targeted according to the values of individual parameters, including other aspects of mapping to the target platform (consumption, chip size, timing). The subject of research will be not only a simple mapping of the learned model to hardware, but also the optimization of the AI model with respect to the implementation in hardware, which requires a closer connection of AI tools with tools for high-level and logical synthesis. The topic can be extended to the synthesis of networks with spiking models of neurons and the integration of learning algorithms into hardware, which is currently an unsolved or little solved problem.
Literature
- [1] Shubham Raif, et al. Logic Synthesis Meets Machine Learning: Trading Exactness for Generalization. arXiv:2012.02530v2 [cs.LG] 15 Dec 2020.
- [2] Chia-Chih Chi and Jie-Hong R. Jiang. Logic Synthesis of Binarized Neural Networks for Efficient Circuit Implementation. In IEEE/ACM INTERNATIONAL CONFERENCE ON COMPUTER-AIDED DESIGN (ICCAD ’18), November 5–8, 2018, San Diego, CA, USA. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3240765.3240822
- [3] Yifan Qian, et al.: Approximate Logic Synthesis in the Loop for Designing Low-Power Neural Network Accelerator. 2021 IEEE International Symposium on Circuits and Systems (ISCAS) | 978-1-7281-9201-7/20/$31.00 ©2021 IEEE | DOI: 10.1109/ISCAS51556.2021.9401451
Bachelor theses
Secure USB Device
Author
Lukáš Solil
Year
2013
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Buček, Ph.D.
Department
3D vision Software Porting on Embedded Linux Platform
Author
Robert David
Year
2012
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Rudolf Marek
Department
Robot Navigation Based on the Sound Source Localization
Author
Tomáš Klinský
Year
2012
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jakub Zahradník
Department
Four wheel robot scenarios
Author
Karel Kovářík
Year
2013
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Dr.Ing. Tomáš Macek
Department
Integration of Kinect advanced sensor do fourwheel robot
Author
Jakub Josef Šimon
Year
2013
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Josef Hlaváč, Ph.D.
Department
Voice command based robot navigation
Author
Matěj Jirka
Year
2013
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jakub Zahradník
Department
Robot NAO as an interface to Internet information resources
Author
Eliška Kuběnová
Year
2014
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Chludil
Department
Heart Beat Sensor Data Transfer to a Mobile Device
Author
Andrej Fedorov
Year
2014
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Buček, Ph.D.
Department
A module for Incremetal Sensor Data Processing for Four Wheel Robot
Author
Martin Kopecký
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Buček, Ph.D.
Department
Stereo-Callibration of Cameras for a Four Wheel Robot
Author
Pavel Šimek
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Martin Daňhel, Ph.D.
Department
Theremin pro robota NAO
Author
Michal Majerčík
Year
2015
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jaroslav Borecký, Ph.D.
Department
Summary
The thesis documents the design and the construction of an electronic musical instrument theremin, which can be played by a human as well as a robot. The aim of this work is developing the digital version of the instrument, which will include two different modes. In the first mode, the instrument behaves like an analog model. In the second mode, the tone of the theremin does not change continuously, but is defined by the tones of the chromatic scale.
Electric circuit was build on universal PCB. Two antennas, made from copper tubes, were mounted on 3D printed casing. Functionality of the theremin was tested by person and robot Nao.
Performance analysis of neural network algorithms on various computer platforms
Author
Alan Dragomirecký
Year
2016
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Tomáš Borovička
Department
Summary
This thesis focuses on design and implementation of system for automated performance testing of artificial neural network algorithms. Testing is mainly focused on embedded systems. To achieve this, we have developed application which generates implementations of artificial neural networks and systems for it's automated testing.
ZYBO board peripherals for the driver development course
Author
Jaroslav Beran
Year
2016
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Pavel Kubalík, Ph.D.
Department
Summary
This thesis deals with design and implementation of simple peripherals in
FPGA of system-on-chip Xilinx Zynq on Digilent ZYBO board. A total of
four peripherals are designed, documented and implemented in this thesis.
The peripherals have control and status registers visible via AXI bus to two
ARM cores running Linux. The first peripheral only generates interrupts when
change of buttons or switches takes place. It also allows write to LEDs. Other
peripherals besides interrupts implement DMA transfers. The second periphe-
ral implements simple transformation function, next one represents disk drive
with scatter/gather support and the last one is configurable camera.
Car dashboard model
Author
Petr Socha
Year
2016
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Tomáš Vaňát, Ph.D.
Department
Summary
The primary goal of this work is to create a car dashboard, which uses gauges and lights to indicate various vehicle conditions. Communication with other components is realized using CAN bus, which is widely used in the automotive industry. The work should become educational material at the Faculty of information technology at CTU in Prague.
A microphone input with dereverberation preprocessing for verbal robot command recognition
Author
Tomáš Kolárik
Year
2016
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jan Bělohoubek
Department
Summary
This thesis investigates possibilities of increasing speech recognition performance,
which is used in four-wheel ROBOT FIT faculty platform. The system
have been suffering problems with distant speech recognition so far. The reason
of these troubles is reverberation which appears in acoustic rooms and degrades
the resulting signal at receiver.
The object is to reduce reverberation maximally. There is a lot of solutions,
less or more robust, with different hardware requirements. For our purposes,
blind impulse response shortening algorithm has been chosen. It has generally
low computational requirements, but still achieves the impulse response effect,
namely early reflections, reduction.
Interaction ability is a crucial feature for a robot. If it was capable of
distant communication, it would increase its usability significantly. Here, an
improvement was made, because the used implementation uplifts the speech
recognition accuracy three times and reduces error rate down to approximately
5 % at most defined commands.
The product of this thesis is libraries and applications system, which
integrates external Distant Speech Recognition enhancements toolkit (DSR)
into software equipment of the robot, and is designed to be easily extended
and tested in the future.
Connection of a game wheel to the CAN bus
Author
Pavel Zajíc
Year
2016
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Martin Daňhel, Ph.D.
Department
Summary
This bachelor thesis deals with connection of a game wheel to the CAN bus.
In the solution of this problem is used data analysis of game wheel buttons,
position analysis of a game wheel and foot pedals from which is formed a message for the CAN bus. Afterwards this message is sent to the modules
with CAN bus interface that are connected to the development board Digilent
ZYBO. To the software is added an option of monitoring transmissions
on the CAN bus and print-out of messages in the text form on terminal or
to the file. The result of a software design and implementation is the console
application in C programming language for operating system Linux. Source
codes of application can be found in the attachment.
A Network Manager as a Module for the NAO Robot
Author
Tomáš Jelínek
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Dostál, Ph.D.
Department
Summary
The thesis deals with design and implementation of modules for robot NAO, which allows management of network connections.
The designed module is capable of automatic recovery of the selected WiFi connection at his sudden loss, choose different WiFi networks and use voice to inform about state of the network. Voice output supports both Czech and English. The thesis also solves computer security issues, especially the security of network traffic via Wifi using the firewall. The functionality of the resulting module implemented in c++ was tested directly on the NAO robot.
CAN Bus Gateway to Computer Network for the Car Simulator
Author
Tomáš Čermák
Year
2017
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Miroslav Prágl, MBA
Department
Summary
This bachelor thesis deals with the creation of a CAN bus to computer network gateway. Also, changes of car simulator Speed Dreams to make it able to be controlled from CAN bus using the created gateway and multiple running instances with different views into the scene synchronized over the network. To create the gate, the converter USB2CAN from the company CANLAB s. r. o and development board Digilent ZYBO were used. Simulator's control using CAN bus is realized with programmed driver in the simulator which respond to incoming messages from the converter. For the run with multiple views, changes to the already existing graphic module are needed. The result of a design and implementation is the application of the gateway in C programming language for operating system Linux and patch for the simulator enabling mentioned possibilities. Source codes, simulator and patch for the simulator can be found in the attachment.
Educational Kit with Force Sensors
Author
Pavel Bezstarosti
Year
2017
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Miroslav Prágl, MBA
Department
Summary
This bachelor thesis is focused on the production of educational kit for work with IoT technologies. The kit allows students to try to connect pre-programmed programs or custom programs to MQTT and HTTP protocols. Using these protocols, kit can also be remotely controlled. For this purpose, the product is further fitted with simple sensors in the form of potentiometers, pressure sensors, buttons and photoresistor. The actors are represented by LEDs. These components allows interaction with the user. An important part is a touch pad built up of pressure sensors that demonstrates the use of multiple sensors to calculate the load location on the surface.
The computational part of the product is built on the Raspberry Pi learning platform, which has been added the expansion module containing the sensors and the A/D converter missing on Raspberry Pi. To simplify access to sensor data, an extension module driver has been written in the form of a kernel module, which makes data available in the form of files. This allows reading with any programming or scripting language without the necessary knowledge of the sensors.
The results and technologies used in this work are described in the following text.
Educational Kit with the Peltier Device
Author
Karel Hynek
Year
2017
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Dr.-Ing. Martin Novotný
Department
Summary
This bachelor's thesis deals with design of educational kit with the Peltier device. The kit can cool and warm up liquid on preset temperature using thermoelectric cooling and warming.
The goal was reached with discrete controller, which is controlling Peltier device by pulse width modulation.
This thesis concerns construction of whole kit and designing cooler for Peltier device considering size of the kit. It also deals with design of electri cal scheme, circuit board and selecting right electronic components. Lastly, thesis deals with implementation of libraries for reading from sensors, discrete controller and sample application.
The educational kit was succesfuly tested and should help to educate students at CTU in Prague.
Web application for robot management - backend
Author
Dan Zatloukal
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. David Bernhauer
Department
Summary
This bachelor thesis covers the analysis and implementation of a robot management system backend which will be used in laboratories of intelligent embedded systems.
Mentioned backend will consist of device database, set of PHP libraries for data display and manipulation in web pages and a interface for executive device functions like remote file upload, remote configuration and status reporting.
The thesis firstly deals with the analysis of available technologies and user requirements.
Then, based on the previous analysis, the thesis focuses on the choices of technologies used in the project and the actual implementation of the system itself.
The finished information system will help with the teaching of subjects which make use of the devices in the mentioned laboratories.
You can find the source code, generated code documentation, integration instructions and picture gallery of the finished system in the attachments.
Web interface for configuration, diagnostics and control of wheealed robot
Author
Vojtěch Řehák
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
doc. Ing. Mgr. Petr Klán, CSc.
Department
Summary
Subject of this bachelor thesis is to develop an application that provides web interface for control, configuration and diagnostics of four-wheeled robot.
Application is running on RoBoard motherboard with linux operating system and WiFi. It uses WiFi for communication with clients using asynchronous queries over HTTP protocol.
Low hardware requirements are important because application is running on less powerful embedded system.
Web application for robot management - frontend
Author
Erik Zatloukal
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. David Bernhauer
Department
Summary
This thesis is mostly taking look at designing (graphicly) and implementing
web user interface and integration of this interface into the core (back-end)
of information system for Laboratory of embedded systems which is being
developped simultaniously.
Firstly, the thesis will be aiming at analysis of relations between data, processes,
functional system requirements and the design layout of web interface.
As a next subject, this thesis looks at designing user interface (logical and
graphical), technology and advantages that they provide for integration with
system core.
In implementation/realization section, emphasis is being placed on flexibility
of the interface and its future expandability/extendebility, in case of
changes in system functionality or even in case of needed design changes.
Wheeled Robot Simulator
Author
Tomáš Balihar
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Kašpar
Department
Summary
This bachelor thesis addresses creation of a wheeled robot model for the purpose of computer simulation. The model in the simulator is used for testing without the need of the real robot. Simulator named Gazebo is used to implement this simulation and together with Linux operating system is essential for it. Also Robot Operating System framework is used to create communication layer between model and the software of real robot. This thesis will speed up development for the robot and it may also serve as a guide for creating model of any real robot.
Gender, Emotion and Age Recognition from Face Images for the NAO Robot
Author
Eliška Vondráčková
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Karel Klouda, Ph.D.
Department
Summary
The aim of this work is to integrate genger, age and emotion recognition
from face images in robot NAO. In first part, I will make a research of face
recognition methods. In second part I will analyze existing implementations
and choose the most suitable one. And I will analyze resources for work with
robot NAO. In the last part I will modify selected implementation for robot
and I will test it.
Object Recognition by Deep Neural Network Applied for the NAO Robot
Author
Martin Kostelanský
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
MSc. Juan Pablo Maldonado Lopez, Ph.D.
Department
Summary
In my work, I researched machine vision and object recognition capabilities.
I used the newest deep convolutional neural networks in framewok Keras.
I examined their architecture and possibilities of their fine-tuning. In my
experiments, I created new network combining two types of networks. Re-
sults showed that this network is more successful in classification than the
original ones. I trained selected networks for space recognition at Faculty of
information technology CTU in Prague. The best networks were implemented
in application for robot NAO, by which it can determine with 93.48% accu-
racy where at the faculty is he located and with 76.09% accuracy in which of
the buildings is he standing.
Robot NAO as an Mathematician
Author
Richard Stanko
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. František Štampach, Ph.D.
Department
Summary
The aim of this thesis is to develop an application suited for equation recognition
and solving. This application is designed to run on the humanoid
robot NAO, developed by Softbank Robotics. The thesis is divided into three
main sections. The first section introduces the robot, NAO his hardware
and software specifications. Later on it deals with the history of OCR, the
technology and its principles. OpenCV and its functions are described in detail
including the mathematics involved. We give a list of tools used for this
thesis and some requirements to make the application run properly. At the
end of the first section a flowchart diagram is used to describe the design of
the program. The second section is dedicated to a detailed description and
explanation of the program which is developed in the Choregraphe suite with
additional Python scripts. The main blocks of the code are explained in detail
including some built-in functions and packages of the Python language. NAO
is able to recognize and solve problems including basic equations, multiple
variable equations, limits and integrals. The application is controled by words
using speech recognition, so it does not have to be run over and over. The
last section addresses testing and results. This project can be used to solve
problems from the blackboard or written on a paper, or to check whether they
are correct or not. For future development it is possible to add an explanation
module, which would explain the topic in more detail, a random problem
generator for the selected topic or use the OCR engine to create a document,
book, text reader.
System for Software Management of Multiple Robots
Author
Petr Kolář
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Robert Hülle
Department
Summary
This thesis deals with design and develop system for the managing NAO robots, linux devices and develop Arduino boards. The designed system is able to run the application, create backup of the devices, restart devices, and provide more functions. The literature search is about analyzing possibilities of identification various devices, question of secure connection with device and effective transfer of files between device and user. It enables to join with the web portal, which handles to the system. The system allows to give a name to Arduino device, this feature is not common in similar systems. Attachment of thesis contains user reference book for the system.
Continuous Speech Recognition by Neural Networks
Author
Adam Zvada
Year
2018
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
doc. Ing. Ivan Šimeček, Ph.D.
Department
Summary
The aim of this bachelor thesis is to explore the field of speech recognition using neural networks with a goal to implement end-to-end speech recognizer as voice-user interface for robot NAO. The proposed speech recognizer architecture is consisted from three main components: feature extraction of speech signal using mel-frequency cepstrum coefficients method, recognizer as recurrent neural networks with long-short-term-memory cells, and connection temporal classification algorithm for predicting the final transcription. The result of this work is end-to-end speech recognizer trained on VCTK corpus and implemented in programming language Python, using deep learning library TensorFlow.
Projection of the robot telemetric data by Hololens
Author
Pavel Chytrý
Year
2019
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Martin Kohlík, Ph.D.
Department
Summary
This work aims to develop a practical application for holographic smartglasses HoloLens. Final program is going to visualize telemetric data, such as sensory data, PWM signals, code variables etc., from connected robots in a way that user can see the physical robot and its data simultaneously. HoloLens are fully untethered device, thus not impeding the ability to move of its users.
Application is on recommendation from Microsoft developed in game engine Unity 3D, then exported into Visual Studio and lastly installed as UWP C\# onto HoloLens. Communication is based on Wi-Fi technology with possible extension to Bluetooth in subsequent development. Connected robots (Sunfounder Smart Video Car and Hexapod in demo) are not constrained in their programming language as long as they support TCP/IP sockets and JSON serialization.
Aquired data is presented on floating panels tracking real-world marker placed on targeted robot coming in two configurations. Default configuration shows up to three entries and the second up to five on both side panels.
A Library for the Hexapod robot control
Author
Jitka Seménková
Year
2019
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Ivo Háleček
Department
Summary
The aim of this thesis is to create a software library for six-legged robot. This library is created for the A-Pod hexapod by Lynxmotion controlled with Raspberry Pi and servo controller SSC-32u by Lynxmotion. This library enables the hexapod robot to walk on flat surface and over terrain as well using contact sensors. Inverse kinematic is used for planning the movement. This library can be used to create more complex applications for hexapod robot by other students of FIT CTU. Manual and CD with video of the robot walking is attached to this thesis.
Robot control by EEG signal
Author
Michal Veljačik
Year
2019
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
doc. RNDr. Pavel Surynek, Ph.D.
Department
Summary
In this thesis, an experimental single--channel EEG based BCI is implemented. Main contribution of this work is a Python aplication capable of recording and experimeting on EEG mental tasks. An experimental BCI for drone control by EEG signal was implemented and connected with the app. It was also discovered that with time the classification accurancy decreases drastically.
Robotic Hand
Author
Ilia Popov
Year
2020
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Petr Socha
Department
Summary
The goal of this thesis was to study a model of a robotic hand from the company YouBionic and then to design and implement software to control this product. This thesis in detail describes the original product, which was then upgraded in terms of both hardware and software. A glove with flex sensors was created to control the model by moving the user's arm. Software solution involved creating script for the control unit of the robotic hand and PC application in C language to control and demonstrate its capabilities.
Smart monitoring system
Author
Michal Žůrek
Year
2021
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Martin Daňhel, Ph.D.
Department
Summary
The goal of this thesis is to develop monitoring system which monitor environment using camera and transfers data of picture to cloud for further processing. Thesis targets security aspects of solution.
The thesis describes possible opportunities to develop monitoring system using different approaches and describes their properties that results from selection. Thesis describes development of IoT device based on Azure Sphere platform which is connected to Microsoft Azure cloud, cloud application based on Azure services developed according to concept PaaS and user application presenting outputs from monitoring system to user.
The thesis also describes some issues associated with development on the Azure Sphere platform and their possible solutions.
Application for minidrone navigation
Author
Jan Filip
Year
2021
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Khun
Department
Summary
This thesis is focusing on ways to control minidrone Parrot Mambo. Main goal is to create application able to autonomously navigate dron based on data gained from sensors and camera. Obstacle avoidance of drone is achieved with modified bug algorithm. Images from cameras are processed with neural network Detectron2 and library OpenCV.
Robot Pepper Application for Mobile Phone Operator
Author
Adam Švehla
Year
2022
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Zdeněk Buk, Ph.D.
Department
Summary
This bachelor thesis covers the design and implementation of an application for the Pepper humanoid robot, which is aimed at supporting customer care at a mobile phone operator's branch office.
The work analyzes the available technologies for the needs of the application and at the same time tries to avoid shortcomings and errors of similar solutions. The aim of the work is to implement an interactive application that will inform and entertain customers. The result is a robust solution that uses state-of-the-art cloud technology for speech recognition and artificial intelligence for localization and intent detection, which stands out in terms of communication, user-friendliness and user experience.
Tracking system for car models
Author
Jan Maček
Year
2022
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
doc. Ing. Pavel Hrabák, Ph.D.
Department
Summary
This bachelor thesis addresses the problem of tracking the movement of cars or other models on the floor of an intelligent embedded systems laboratory in real time using a static camera. The goal was to enable sophisticated tracking using deep neural networks in addition to model localization based on simple markers, with an emphasis on fast adaptation of the tracking system to new models. I used ArUco Markers for localization using markers. For object detection using deep neural networks, I used the YOLO model. I designed and implemented a method of automatic data labeling using background subtraction, which greatly accelerated the adaptation of the tracking system to the new models.
Autonomous vehicle model parking
Author
Ondřej Gössel
Year
2023
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Martin Daňhel, Ph.D.
Department
Summary
The aim of this thesis is the development of autonomous parking software for the vehicle models at the Laboratory of Intelligent Embedded Systems at FIT CTU. The thesis includes a research study of algorithms for autonomous parking and an analysis of the current hardware and software equipment of the models. Based on the research and analysis, an algorithm for precise control of the vehicle models during parking was designed and implemented in the Python programming language. To facilitate working with the models, a terminal was implemented, which includes commands for various functionalities, such as model control and calibration. For easy testing, a simulator replicating the real vehicle model was created in the Godot engine. The designed control system and parking algorithm were implemented and systematically tested in the simulation and later on the hardware model. Parking tests were performed in rectangular parking spaces of varying lengths from different starting positions. The success rate of parking in the simulation was high, while on the hardware models, it was lower due to hardware inaccuracies and limitations. The main contribution of this work is the overall enhancement of the software equipment for the models, which will facilitate future work with them.
Generation of a neural network model to VHDL
Author
Jan Medek
Year
2023
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
doc. Ing. Petr Fišer, Ph.D.
Department
Summary
This thesis covers design and implementation of a generator, which generates hardware description in VHDL language from software model of a neural network designed in Keras API. It also covers a configuration of network parameters in a way, that allows to convert them to data types specified by the created configuration and is able to export them. The thesis analyzes Keras API, existing implementations which generate VHDL from the software models and proposes a solution, which achieves generating the VHDL straight from the Keras model without any stages in between. The result is a working application implemented in Python that is able to take a model implemented in Keras with given building blocks written in VHDL, generate synthesizable VHDL and export network's parameters converted to data types specified by the configuration.
Neural network acceleration on cloud card with FPGA
Author
Jiří Tlamicha
Year
2024
Type
Bachelor thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Jiří Khun
Department
Summary
This thesis researches the acceleration of neural networks on cloud FPGA card Alveo U55C. In particular, it explores the HLS4ML framework, which uses high-level synthesis, and the Vitis AI framework, which uses DPU acceleration. First both frameworks are tested on small networks trained for the Iris dataset. Then they are used for experiments with a set of convolutional neural networks of different sizes trained for the CIFAR-10 dataset. Limits of both frameworks are found. Experiments are evaluated in terms of processing time and power consumption. Then the Vitis AI is used to accelerate the YOLOv7 model to detect objects in images. This model is then used in a prototype application for processing of images from humanoid Nao robots. This model is also evaluated in terms of throughput and latency.
Master theses
Robot NAO as a Lecturer
Author
Gabriela Turcajová
Year
2019
Type
Master thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
doc. Ing. Ivan Šimeček, Ph.D.
Department
Summary
The aim of the thesis is to create software for the robot NAO, which will allow the robot to lead a lecture. The content of this work is the design and implementation of the application for entering the texts for the slides of presentation which the robot will read, and the subsequent start of the application led by the robot.
Cooperation of an industrial cooperative robot with a worker with holographic glasses
Author
Adam Podroužek
Year
2019
Type
Master thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
Ing. Martin Kohlík, Ph.D.
Department
Summary
This diploma thesis aims to introduce the reader with the topic of development of a control software for a cooperative robot which interfaces with a human by holographic glasses. First part of the thesis deals with analysis and description of the development process for YuMi robot in RobotStudio IDE as well as with development of applications for holographic glasses HoloLens.
The laboratory of intelligent embedded systems on FIT CTU can make use of this work when developing applications with students during classes. The thesis can also be a starting point for someone from the industry interested in implementing such cooperative system.
Autonomous Car Model Control
Author
Petr Kolář
Year
2021
Type
Master thesis
Supervisor
Ing. Miroslav Skrbek, Ph.D.
Reviewers
doc. Ing. Ivan Šimeček, Ph.D.
Department
Summary
This thesis deals with controlling the autonomous vehicle, based on microcomputer Raspberry Pi. This vehicle model was improved with sensors for measuring distance and with incremental encoders for measuring travelled distance, speed and direction. Additionally, the vehicle was improved with more parts for improved driving properties and simplifying the use of the model. For sensor connections printed circuit boards were created. The car's control software was also modified.
A~complex autonomous driving system was created for the model, which included a track with road signs, ultrasound car localisation and an application for track recognition. Recognition is performed both in a traditional way (openCV) and with the use of neural networks. The driving system and the modified model were tested on different model tracks.