Homography is a method for detecting keypoints between two images and determining their overlap for the purpose of image alignment. Recently, thanks to neural networks, many solutions for keypoint detection have emerged. The question, however, is whether such trained networks can also be used to find similarities (identification) of a person in two images. Your task will therefore be to implement several different algorithms for homography and keypoint detection and subsequently test their accuracy in identification.

The vascular system of the finger has an unique pattern to every individual and to each finger. Nowadays, various algorithms and methods exist for identifying a person based on images of the finger’s vascular network. Recognition can be performed by comparing two images. Algorithms and neural networks, which work on the principle of detecting key points and comparing them, attempt to find common parts between two images. Your task would be to test these algorithms on images of the finger’s vascular network and determine whether their results can be used for personal identification.

Detection of basic facial emotions.

The goal of this thesis is to design and train a neural network for tracking distant targets. The solution will also involve the development and implementation of suitable control algorithms for a manipulator equipped with a tracking camera.

The project requires the integration of machine learning, computer vision, and motion control techniques. This thesis is offered in collaboration with the Faculty of Electrical Engineering.

Modern large language models (LLMs) equipped with visual encoders offer new possibilities for analyzing and understanding visual data, including in domains for which they were not originally trained — such as biometric recognition.

This thesis explores the potential of selected multimodal LLMs (e.g., GPT-4o, LLaVA, and CLIP) for recognizing individuals based on facial images, iris patterns, and finger vein structures for biometric person recognition. This study will evaluate these models' ability to identify or verify individuals using visual input without additional training on domain-specific biometric datasets.

Thesis Tasks:

Study the current state of the art in multimodal LLMs and their applications in visual tasks, particularly in biometrics.

Select suitable models (e.g., GPT-4o, LLaVA, and CLIP) and design a methodology for using them in facial, iris, and finger vein recognition.

Implement an application that enables testing of these models on selected biometric datasets.

Design and conduct experiments to assess the models' ability to recognize individuals under varying conditions.

Simulate attacks (e.g., presentation attacks and image manipulations) and analyze factors that may negatively affect recognition accuracy.

Finally, evaluate the results, identify limitations, and propose possible improvements.

The aim of this thesis is to design and implement an application for assessing running quality using modern multimodal models—specifically large video-language models (Video-LLMs)—that enable the analysis and interpretation of human motion from video recordings.

The thesis will involve:

- Studying current methods for evaluating running technique in biomechanics and sports diagnostics.

- Analyzing the potential of modern Video-LLMs (Large Multimodal Models), such as GPT-4o, Flamingo, Video-LLaMA, or VideoCLIP, for processing running videos.

- Proposing a methodology for processing running footage, including preprocessing steps (e.g., human detection, keypoint extraction).

- Implementing a prototype application that enables: loading a running video, analyzing running form, and generating qualitative feedback in natural language (e.g., strengths and weaknesses of the running technique).

- Validating the functionality of the system on a selected set of test videos (e.g., treadmill running, outdoor running).

- Evaluating the accuracy, benefits, and limitations of using Video-LLMs in this domain.

This thesis aims to design and implement a solution for a single-board computer (e.g., Raspberry Pi or NVIDIA Jetson) that can evaluate a video stream in real time and identify predefined, user-specified events described in natural language (e.g., "a gardener watering plants with a watering can").

When a specified event is detected, the corresponding video sequence is automatically recorded and a notification is sent via a LoRa network to a central server or another receiver.

This thesis covers processing the input video stream on a low-power, resource-constrained device; applying computer vision technologies and language-guided event detection (e.g., using CLIP or Vid-LLMs); optimizing the system for real-time performance in a low-power environment; and integrating with a LoRa communication module.

Language models currently play an indispensable role across many areas of artificial intelligence, including applications in the field of biometrics.

The goal of this thesis is to leverage existing facial encoders and language models to design and implement a system capable of estimating a person’s origin (e.g., geographic, ethnic, or cultural context) based on a facial image.

The thesis will include:

an analysis of current approaches to processing biometric data using large language models (LLMs),

the design of a suitable facial representation using embeddings,

and the development of a component that links visual input with an output in the form of a language-based explanation or classification of a person’s origin.

The aim of this thesis is to detect and, if applicable, classify wildlife in images captured by a thermal camera. The student will design and train a suitable neural network, conduct the necessary experiments, and develop a user interface for presenting the results. This bachelor’s thesis is carried out in cooperation with the STRADE research group at FIT BUT and the Faculty of Forestry and Wood Technology at MENDELU.

- The goal is to use a 3D printer or a robotic hand to write credible signatures.
- It will probably also be necessary to create a pen to capture the signature or use a tablet.
- Possibility to be inspired by previous work.
- Work requires work with on both HW and SW.
- Alternatively, it would be possible to generate different signatures (from the entered text) using machine learning.

Create and test algorithms for composition of image with visible finger veins to increase a quality of image.

The aim of the thesis is to get acquainted with the problem of automatic object tracking in relation to the control of a camera manipulator using modern neural networks. Within the framework of the solution the student should perform a research of existing networks suitable for the purpose of object tracking in video, design and implement an interface to connect the output of the network and the existing SAOTS surveillance system from the STRADE research group, integrate selected network models into a manipulator control application, evaluate the advantages and disadvantages of such a solution.

The aim of this project is to present the current state of the long-range surveillance and tracking system (SAOTS) developed by the STRADE group(https://strade.fit.vutbr.cz/) and its subsequent innovation. In the hardware domain, it concerns the selection of a suitable hardware upgrade aimed at computer vision applications. In the software domain, it is then the design and implementation of a unifying application that allows easy control of the entire platform currently served by several separate applications.

The goal is to scan a finger using three cameras, at suitable wavelengths, combine the photos, extract the fingerprint and create a 2D rolled print from it.
It would be advisable to create a 3D model from the photos and to mark the markings on the resulting image.

The main goal is create images of finger veins, which will simulate data from finger vein scanner device. 3D model of finger vein structure will be available, to generate a multiple images.

This topic is about extraction of the finger from the image. Definition of the mask of finger and normalisation of finger in multiple angle of view.

Transformation between touch-based fingerprints to various other types (contactless, patent, latent, plastic) including specific damages (usually background).

The goal is to simulate realistic background into synthetic fingerprint images. Background could be from various sensors or usual latent fingerprint background.

The goal is to improve synthetic fingerprint generation part of SyFDaS. There are several parts which could be improved. For the final description it is expected to finish one maximally two of these points (so there could be several thesis description generated from this one - based on candidate preferences). What can be improved:
Generation of structured information about generated fingerprint (minutiae, class, singularities, density etc.).
Generation of fingerprint based on template.
Generation of several realistically looking backgrounds.
Automation of the generation ("clever" randomisation of input data).
Improvement of class specification/different minutiae for fingerprint generation.
Extend the generation to generate rolled fingerprints/whole finger/palmprint.

The goal is to create methods to detect presentation attack on images with fingerprint and vein/thermogram. Preparation of database will be needed as well as some cooperation with creation fingerprint+vein device or position of thermal cameras.

The goal of this work is processed image in real-time from finger vein scanner device and set the illumination to increase a visibility of finger vein structure.

1) Study multispectral liveness detection technology for fingerprints and whole hands.
2) Analyze images from the supplied databases, or find other freely available databases.
3) Design an algorithm which of the multispectral images decides whether it is a living or non-living hand (counterfeit).
4) Implement the proposed algorithm.
5) Evaluate the success of the liveness classification.
6) Perform an experimental verification of the software solution and summarize the results obtained. Discuss possible extensions.

The goal is create application for image processing to determine blood vessels in the finger from multiple angle o view. The database is available.