Actual projects

Master thesis 2013/2014/2015 - Bc. Andrej Valko

Object recognition

Result: in progress

Assignment. One of the basic concepts of object classification is the use of local descriptors within the images of annotated dataset to train a classifier, which will be used on validation dataset to recognize given objects. One of the major issues is a generic class, which can be divided into several subcategories. Such an example is ?car? category. There are different kinds of cars, characterized by different visual features (sedan, kombi, etc.). By training of single class, this class becomes too generic and the recognition system gets more apart from reality. This is the reason to emphasize the importance of the pre-processing of descriptors. Analyze various approaches in the area of object recognition. Focus on the pre-processing of descriptor which precedes the classification. Design your own method for descriptor processing, which will create the artificial subcategories based on basic similarities of characteristics of the given category in the annotated dataset. Implement proposed method and evaluate the solution on available annotated dataset compared to state-of-the-art methods.

Bachelor thesis 2014/2015 - Peter Sarisský

Eye blink detection using webcam

Result: in progress

Assignment. Dry eye syndrome is a common disease of computer users. Users tend to decrease blink frequency while looking at computer screen due to which tear film layer corrupts. The goal of the bachelor thesis is to work on algorithms which will detect user eye blinks with the aim to analyze their frequency. Analyze also vailable approaches of face and eye detection. Suggest modifications of existing approaches to increase the detection rate. Test the modifications and evaluate them on existing datasets.

Bachelor thesis 2014/2015 - Lenka Kutlíková

Evaluation of intrusiveness of notifications

Result: in progress

Assignment. Computer users often suffer from eye conditions caused by low eye blink rate. An application is being developed as part of project Eyeblink which will remind people to blink more often. The frequency of of notification is crucial because of goal to avoid distraction during work. Analyze research on human computer interaction about notifications with their more frequent use. Focus on different types of notifications and the type of response required from the user. Suggest different models of notifications and test them on a representative sample of users. Users will assess subjective intrusiveness of individual types of notifications during their work. Objective testing will be based on the specific task which will be done with and without notifications under time measurement.

PAST PROJECTS


Bachelor thesis 2013/2014 - Tomás Drutarovský

Driver drowsiness detection

Result: A (pdf), paper at ACVR ECCV 2014, paper at IIT.SRC 2014, Dean's letter of praise for the excellent thesis

Abstract. The bachelor thesis introduces a problem of driver's microsleep detection. According to the state-of-the-art, sleepiness can be detected using the analysis of the eye-blink frequency. The goal of the bachelor thesis is to propose a reliable eye-blink detection algorithm, which could be used in the microsleep detection system. Introduction is devoted to the microsleep problem definition. The thesis core analyzes currently available solutions of this problem and it presents face and eye detection and tracking methods. Afterwards a proposal of eye state description algorithms based on SVM (Support Vector Machine) and feature descriptors is presented -- Intensity Vertical Projection, SIFT descriptor and own gradient descriptors. Further focus is put on our own eye-blink detection methods based on sequential frame analysis, which achieves better results than Divjak-Bischof optical flow detection method. The proposed methods were tested with datasets obtained from different subjects under different light conditions.

Bachelor thesis 2013/2014 - Rudolf Brisuda

Tones Recognition for Displaying Interactive Music Sheets on Mobile Devices

Result: A (pdf), paper at IIT.SRC 2014

Abstract. This work deals with the development of the algorithm for applications for mobile devices, which would be able to recognize tones played on a piano in real-time and thus tracks position within the song while playing. With this information we can do automatic page turning. We have outlined the problem of turning pages in the introduction. The work is further devoted to the analysis of existing solutions available for mobile applications. We analyzed the way of storing music in digital form. The work also analyzes various algorithms for tone recognition. We can learn about the available algorithm uses in the solution to polyphonic pitch detection (Klapuri). After the evaluation of lower accuracy, we focus on note onset detection. We created the algorithm which connects polyphonic pitch detection and onset detection. Next, we focus on pitch detection algorithm (Cepstrum). Use of this algorithm and subsequent focus on certain notes showed better results. Testing was done with 10 songs. We monitored our position in comparison with expected position. Testing algorithms were note onset detection, note onset detection with Klapuri's/Cepstrum algorithm and Cepstrum itself. Tests showed that using Klapuri's algorithm is not sufficient to track position. The Cepstum itself achieves success in 6 of 10 songs with error rate less than 5 keystrokes at the end of the page. However the process of monitoring is mostly irregular due to abstention of information about note duration. Note onset detection with Cepstrum achieves the best results (7 of 10). If we take into the account the average error rate at all pages, note onset detection with Cepstrum achieves the best results (the average error rate 5.46 keystrokes).

Master thesis 2012/2013/2014 - Bc. Michal Osvát

Object detection and segmentation using contours

Result: C (pdf)

Abstract. Object detection is one branch of computer vision which is analyzed for past few decades. This thesis analyzes different principles and methods of object detection. Methods such as HOG and Sliding Window are often a part of the most common approaches. We analyze in detail different descriptors such as HOG and DOT, and also edge detectors such as LSD method. After analysis of several approaches and principles, we decided to design new approach for object detection, based on image gradients. Thesis introduces principle of features detection (called segments). Each segment is defined by specific contour, shape, and size. In the design phase we focused on providing different types of invariance and therefore we use Hu invariant moments as descriptor. We also propose own method for classification of obtained segments which is evaluated and compared to K-Nearest Neighbors method. Our method for object detection has been evaluated on the data-set of human faces. We have evaluated proposed method on eye detection and thesis includes comparison of results to widely used method proposed by Viola and Jones. Our object detection method based on segments is finally less successful. However, better results have been achieved when we compare accuracies of detected areas. In this case, proposed segments more precisely describe areas of detected eyes in comparison with manual annotations, than areas obtained from Viola-Jones object detection. Moreover, we used only 60 eyes in the training phase, what is significantly lower number than Viola-Jones method has been trained on.

Master thesis 2012/2013 - Bc. Peter Sivák

Object detection and segmentation using part based model

Result: (student unexpectedly discontinued studies)

Assignment. Object recognition is an active area in computer vision research. There exist several methods which are used for object recognition. An important model is represented by systems based on recognition of individual object parts and evaluation of their relative positions. In this way, various types of objects can be recognized, which are composed from same or similar parts, but simultaneously, they are differently arranged and thus represent different objects. Main approaches are star, tree and fully-connected topology. Analyze available part-based model methods, which are used for object recognition with respect to time complexity and quality of solution. Analyze data representation, which are provided to classifier. Design your own solution of model, which will take into account the request of low time complexity. Experiment with various parameters and possible varying conditions of recognition. Verify designed solution on a representative data set of images. Compare your results with existing results in the area.

Bachelor thesis 2012/2013 - Patrik Polatsek

Monitoring eye blinking frequency of computer users

Result: A (pdf), Dean's award for the best bachelor thesis, paper at IIT.SRC 2013

Abstract. This bachelor thesis deals with eye blink rate tracking of the user while working with computer. A user tends to decrease the blink rate in front of a computer screen, due to which the tear film is non adequately applied on the eye cornea. Lower blink rate causes eye redness and dryness. This commonly-occurring problem of computer users is called Dry Eye. The goal of the bachelor thesis is to design an eye blink detection algorithm. In future it can be used in dry eye prevention application, which will detect user's blinks. We have analysed available techniques for eye blink detection and designed our own solutions based on histogram backprojection, optical flow, frame difference and FREAK descriptor method. We have tested our algorithms on different datasets under various lighting conditions. Centre Aligned Movement Detection method based on optical flow performs better than the other ones. We achieve higher recognition rate and much lower false positive rate in the Talking Face Video dataset than the-state-of-the-art technique presented by Divjak and Bischof in 2009.

Bachelor thesis 2012/2013 - Ján Podmajerský

Interactive Whiteboard

Result: C (pdf), prototype application

Abstract. This bachelor thesis deals with construction of an inexpensive interactive whiteboard. The most distinctive feature of the solution is a light emitting diode(LED) detection, which is done in the visible light spectrum. Most of available solutions use infrared light to achieve the expected behavior. Having taken into consideration the state of the art, we have decided to use the visible spectrum of light, because it can it can work in common daylight conditions too. The LED is mounted at the top of the pen, which points onto a projected image of the video projector. Infrared light cannot be used on a daylight, because infrared light is present in daylight, due to which the use of infrared light could be inaccurate. With the aim of inexpensive solution a common webcam can be used as the capture device. Thresholds in individual channels of the HSV color model are set up to filter the LED. The evaluation can provide a real-time performance. Homography is used to translate the camera coordination system into screen coordinates. Finally the operating system's interface places mouse cursor on demand position. We achieved precision within 5 pixels. The detection does not work in the extreme light conditions, when there is a strong sun shining on the projecting plane. We created a prototype implemented in OpenCV and Qt.