|
|
Approximation of Sound Propagation by Neural Networks
Nguyen, Son Hai ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
Za účelem nahrazení výpočtově náročných konvenčních numerických metod řešících diferenciální rovnice jsou neurální výpočty stále více prozkoumávány. Tato práce se zaměřuje na řešení časově nezávislé Helmholtzovi rovnice, která modeluje šíření ultrazvuku při transkraniální léčbě ultrazvukem. Při použití konvolučních neuronových sítí musí být data navzorkovaná na pravidelné mřížce, abychom odstranili dané omezení, navrhli jsme neurální výpočet založený na grafových neuronových sítích. Narozdíl od fyzikálně informovaných neuronových sítích (PINN) je potřeba náš model natrénovat pouze jednou, řešení pro množinu nových parametrů vyžaduje pouze dopředných chod. Model byl natrénovaný pomocí učení s učitelem, kde referenční data byly vypočítána pomocí konvenční metody k-Wave. Náš model má stabilní rozvinutí, přestože byl natrénovaný pouze s osmi iteracemi. Ačkoli byl model natrénovaný pouze na datech s jedním zdrojem vln, tak zvládne predikovat i vlnová pole s více zdroji i v mnohem větších výpočetních doménách. Náš model je schopen predikovat subpixelové body s větší přesností než lineární interpolace. Dále je naše řešení schopno predikovat vlnové pole i s podvzorkovaným Laplaciánem, kde jsou pouhé tři vzorky na jednu vlnovou délku. Nejsme si vědomi žádné existující metody fungující s takto řídkou diskretizací.
|
|
| |
|
|
Smartwatch App for Sports Training and Competitions
Dohnalík, Pavel ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
The aim of the work is to create an application for a smart watch, which will allow you to measure races and trainings, or create localization data for this activity. The application is implemented for mobile devices with the Android and iOS operating systems. The Wear OS operating system is supported for smart watches. The thesis describes the theory of programming for mobile operating systems and programming for the operating system Wear OS. The practical part describes the design, implementation and testing. For the implementation of the mobile application as well as for the smart watch application I decided to choose Flutter framework and programming language Dart. The resulting application allow users to measure races and workouts.
|
|
|
Detection of a Yoga Poses in Image
Kutálek, Jiří ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
The motivation for this thesis is the concept of a smartphone app detecting Yoga poses and displaying Yoga frames to a user. The goal of this project is proving that even a simple Convolutional Neural Network (CNN) model can be trained to recognize and classify video frames from a Yoga session. I created an application in which the videos are manually annotated. The data, consisting of frames captured from 162 collected videos based on the annotations, is then passed to train a CNN model. The Dataset consists of 22 000 images of 22 different Yoga poses. The frames are captured using the OpenCV library, the training process is handled by the TensorFlow platform and the Keras API, and the results are visualized in the TensorBoard toolkit. The Model's multi-class classification accuracy reaches 91% when the binary cross-entropy loss function and the sigmoid activation function are used. Despite the promising experimental results, the main contributions are the dataset forming tools and the Dataset itself, which both helped to confirm the proof-of-concept.
|
|
|
Detection of Boxes in Image
Soroka, Matej ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
The aim of this work is to experiment and evaluate different approaches of computer vision with the aim of automatic detection of boxes-blocks in the image, for this purpose, approaches based on neural networks were used in the solution. Experiments were performed with classification using our own data set, classification using our own convolutional neural network, detection using a window, YOLO detector and in the last part a proposal for improvement using U-net and MirrorNet networks.
|
|
|
Tracking of Moving Objects in Video
Folenta, Ján ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
This bachelor thesis deals with the issue of detection, tracking and counting vehicles in different directions in video. To deal with this problem, modern techniques of object detection and tracking using convolutional neural networks are used. The goal of this work is to achieve highest possible accuracy of vehicle counting while maintaining the processing of video recordings in real-time. The problems of the implemented method for detection and tracking are solved by analyzing and working with the trajectories of vehicles. With accuracy of 90,94% and total score of 0,8829, this work participated in AI City Challenge 2020, where it placed 6th.
|
|
|
Recognition of Driving Lane Borders in Video from On Board Camera
Letovanec, Lukáš ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
This thesis is dedicated to the issue of driving lane borders recognition in frames of an onboard camera. In this thesis, an architecture of a deep convolutional neural network is introduced, by means of which the said problem is dealt with. The net was trained on a large dataset using gradient descent algorithm. The trained model has demonstrated the ability to recognize borders of a driving lane well in different situations and conditions. The result of the thesis confirms that deep convolutional neural networks are a suitable tool for driving lane borders recognition.
|
|
|
Ball Tracking in Sports Video
Motlík, Matúš ; Špaňhel, Jakub (referee) ; Bartl, Vojtěch (advisor)
This master's thesis deals with automatic detection and tracking of a soccer ball in sports videos. Based on the introduced techniques focusing on tracking of small objects in high-resolution videos, effective convolutional neural networks are designed and used by a modified version of tracking algorithm SORT for automatic object detection. A set of experiments with the processing of images in different resolutions and with various frequencies of detection extraction is carried out in order to examine the trade-off between processing speed and tracking accuracy. The obtained results of experiments are presented and used to form proposals for future work, which could lead to improvements in tracking accuracy while maintaining reasonable processing speed.
|
|
|
Obtaining and Processing of a Set of Vehicle License Plates
Kvapilová, Aneta ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
This master thesis focuses on creating and processing a dataset, which contains semi-automatically processed images of vehicles licence plates. The main goal is to create videos and a set of tools, which are able to transform input videos into a dataset used for traffic monitoring neural networks. Used programming language is Python, graphical library OpenCV and framework PyTorch for implementation of neural network.
|
|
|
Detection of Landmarks on Vehicle Images
Chadima, Vojtěch ; Bartl, Vojtěch (referee) ; Herout, Adam (advisor)
This thesis aims to introduce automatic detection of landmarks on vehicle images. Detected landmarks can be then used for automatic traffic surveillance camera calibration or other computer vision applications. I solved the landmarks detection problem by using a novel type of convolutional neural network called Stacked Hourglass. Furthemore, I created an automatic trainig dataset (image + anotations) generator based on Blender API, which allows to create various datasets. Detected landmarks are analyzed and sorted in order to determine a set of superior landmarks that could be later used for camera calibration. The best-performing models detect up to 1 021 landmarks, while the best of them have less than 3.0 pixels average error. Finally, results can be further used in automatic camera calibration based on landmarks detection, to create custom datasets or to train Stacked Hourglass convolutional neural networks.
|
guest ::
RSS feed.