Please let me know if you will cite my work, thank you.


Publications


Drutarovsky T., Fogelton A.: Eye Blink Detection using Variance of Motion Vectors, Proceedings of 2th workshop on Assistive Computer Vision and Robotics in ECCV 2014, Springer 2015 bibtex, pdf, Eyeblink8 dataset (under GPLv3 licence)

Keywords

eye blink detection, statistical variance, motion vectors, outlier detection, global movement compensation

Abstract

Abstract. A new eye blink detection algorithm is proposed. It is based on analyzing the variance of the vertical motions in the eye region. The face and eyes are detected with a Viola-Jones type algorithm. Next, a flock of KLT trackers is placed over the eye region. For each eye, region is divided into 3x3 cells. For each cell an average "cell" motion is calculated. Simple state machines analyse the variances for each eye. The proposed method has lower false positive rate compared to other methods based on tracking.We introduce a new challenging dataset Eyeblink8. Our method achieves the best reported mean accuracy 99% on the Talking dataset and state-of-the-art results on the ZJU dataset.

Fogelton, A.: Object Class Recognition, Diseratation project III (Minimum thesis), Bratislava, Slovakia, November, 2012, .pdf

Keywords

Part based models, Constellation, Star, Tree, Atributes, Instance level recognition, Datasets, Superpixels, Semantic Texton Forest, Random Forest

Abstract

Abstract. Thesis introduces the issue of object class recognition with the aim to develop system, which would be able to do it within short computational time. In the introduction the object class recognition problem is divided into three categories: image classification, object detection and segmentation, characteristics of which are to be described. The state-of-the-art methods in this area are presented. The basic characteristics of local descriptors are explained to introduce common method of bag of visual words. It also describes important part-based models while comparing different kinds of structures. Thesis focuses its attention to modern trends in this area like attribute based instead of class based object recognition. Semantic Texton Forest method is described in greater detail not only because of the low time consumption but the simplicity of the algorithm as well. More details are given to used features and Random Forest classifier, which achieves similar qualitative results as other methods, but in faster way. We inform about our own experiments on decision tree classifier and coherent pixel clustering. At the end we propose thesis on our further research.

Fogelton, A.: Evaluation of Image Segmentation Based on Histograms, Proceedings of 8th Student Research Conference in Informatics and Information Technologies, Bratislava, Slovakia, April 25, 2012, Publisher STU, Volume 2, ISBN 978-80-227-3690-9, pages 429-433 .pdf bibtex

Keywords

HSV color model, historgam, intersection method, sliding window

Abstract

Abstract. The presented paper evaluates an experiment on image class segmentation based on Hue-Saturation histograms. Training is based on histogram calculation for every object class separately. Sliding window is performed to segment (label) individual pixels of the evaluation image. Sliding window around the given pixel encloses the local appearance from which the histogram is calculated. Local appearance histogram is subsequently used to be compared with the precomputed class histograms. Given pixel is labeled according to the best match using the intersection method. We show how this method depends on data character and window size. Unfortunately, this algorithm suffers from both quality and speed performance.

Fogelton A.: Hand Tracking, Master thesis, Faculty of Informatics and Information Technologies Slovak University of Technology, Bratislava, 2011, .pdf bibtex

Keywords

KLT features, flocks of features, HSV color model, histogram backprojection

Abstract

This master thesis deals about hand tracking, which could be used as an input device for human computer interaction. After analyzing several approaches, we focus on the Flocks of Features algorithm. This algorithm takes into consideration skin color cue and feature motion detection. Features are being tracked on gray-level image using Lucas-Kanade tracker. To obtain the hand location, flocking behavioral model of birds is applied on these features. The algorithm is vulnerable to the corners and edges occurring in the background. We proposed the modification of this algorithm not to use the gray-level image as the input, but a processed image of the skin color probability map. This map does not contain the majority of the disturbing elements. To increase the invariance to the luminance conditions an adaptive histogram was introduced. We also proposed several initialization algorithms for multiple object tracking. In the conclusion the achieved results and possible future work are presented.

Fogelton A.: Real-time Hand Tracking using Modificated Flocks of Features Algorithm, Information Sciences and Technologies Bulletin of the ACM Slovakia, Special Section on Student Research in Informatics and Information Technologies, Vol. 3, No. 2 (2011), ISSN 1338-1237, pages 37-41 .pdf bibtex

Keywords

KLT features, flocks of features, HSV color model, histogram backprojection

Abstract

There is a growing demand to interact with computers in a more natural way. For example using hand gestures to interact with certain type of applications would be more efficient than old-fashioned mouse and keyboard. To achieve this we need to be able to efficiently track human hand in real-time. We focused on Flocks of Features algorithm introduced by Mathias Kolsch and Matthew Turk, which can track human hand continuously during various movements and pose variations. It uses Lucas-Kanade tracker for features located on a human hand. This algorithm can handle tracking of fast movements of non-rigid highly articulated objects such as hands. We propose modiffications to this algorithm which mostly correspond to the preprocessing of the input frame by using histogram back projection of the skin color. This modiffication according to our testing provides more reliable feature tracking which results in better hand tracking efficiency.

Fogelton, A.: Initialization of Multiple Objects Tracking using Flocking Behavior of KTL Features, Proceedings of 7th Student Research Conference in Informatics and Information Technologies , Bratislava, Slovakia, May 4, 2011, Vol. 2, Publisher STU, ISBN 978-80-227-3488-2, pages 405-410, [best paper award in between Master students] .pdf bibtex

Keywords

clustering, gauss distribution,

Abstract

Computer vision is closely linked to machine learning, where there is mostly a requirement for efficiency and real-time performance. As an example we present our modification of the Flocks of features algorithm, which is basically an efficient partial clustering method. The main contribution of this paper are two initialization methods used along this algorithm. The first static method is used to obtain multiple clusters with density based approach, which will be subsequently tracked by Flocks of features. The second dynamic method is used to detect new objects to track along already running FoF tracking on a frame sequence.

Fogelton A.: Real-time hand tracking using Flocks of features, CESCG 2011 proceedings of the 15th Central European seminar on computer graphics, May 2-4, 2011, Vinicné, Slovakia, Publisher: Vienna University of Technology, ISBN 978-3-9502533-3-7, pages 107-114 .pdf bibtex

Keywords

KLT features, flocks of features, HSV color model, histogram backprojection

Abstract

There is a growing demand to interact with computers in a more natural way. For example using hand gestures to interact with certain type of applications would be more efficient than old-fashioned keyboard and mouse. Hand tracking is one of the keen problems in computer vision. We have analyzed many different approaches used for hand tracking. Flocks of features introduced by Mathias Kolsch and Matthew Turk can track human hand continuously during various movements and pose variations. It uses the Kanade Lucas Tomasi (KLT) tracker for features located on a human hand to track them in a frame sequence. It can handle fast tracking of non-rigid highly articulated objects such as hands. We propose an improvement to this algorithm by processing the frame using histogram back projection of the skin color prior to applying flocks of features (FoF). This modification provides better results with lower false positive error.

List of testing videos and results:

1. long sleeves video results: FoF mFoF CamShift 1. short sleeves video results: FoF mFoF CamShift

2. long sleeves video results: FoF mFoF CamShift 2. short sleeves video results: FoF mFoF CamShift

3. long sleeves video results: FoF mFoF CamShift 3. short sleeves video results: FoF mFoF CamShift

4. long sleeves video results: FoF mFoF CamShift 4. short sleeves video results: FoF mFoF CamShift

5. long sleeves video results: FoF mFoF CamShift 5. short sleeves video results: FoF mFoF CamShift

6. long sleeves video results: FoF mFoF CamShift 6. short sleeves video results: FoF mFoF CamShift

Fogelton A., Kallo O., Ondruska P., Palo M., Ukrop J.: TrollEdit - New Way of Source Code Editing,Proceedings of 6th Student Research Conference in Informatics and Information Technologies, April 21, 2010, Bratislava, Slovakia, Publisher STU, ISBN 978-80-227-3267-3, pages 541-542, .pdf bibtex

Keywords

Lua, QT, graphical editor

Fogelton A.: Image processing, Bachelor thesis, Faculty of Informatics and Information Technologies Slovak University of Technology, Bratislava, 2009 .pdf bibtex

Keywords

LED pen, homography, calibration, thresholding, JAVA

Abstract

This work deals with image processing, more accurately it is about how to access a webcam to set up the stream of images available for next processing. Similar existing solutions in the field of interactive whiteboards were analyzed. The software is tracking the location of the LED pen (a pen, which has a highly luminous LED at the top) by a webcam. This software works over the position of the pen into the coordinate system of a display device and this may be used instead of an expensive interactive whiteboard. This principle can be used mostly with projectors, but also with LCD displays. Technologies of accessing the webcam are described here as well. The design presents a prototype of the solution, where the algorithms used in image processing are described. The homography principle was used for recalculating the positions between coordinate systems. In the verification of this solution, progress of the work was described and the selected approach was reasoned. In the part dealing with hardware solutions the demands on the webcam and the construction of the pen were described.

Fogelton, A.: Interactive Whiteboard JAVA API, Proceedings of 5th Student Research Conference in Informatics and Information Technologies , April 29, 2009, Bratislava, Slovakia, Publisher STU, ISBN 978-80-227-3052-5, pages 342-348, [IEEE prize], .pdf bibtex

Keywords

LED pen, homography, calibration, thresholding, JAVA

Abstract

Although, interactive whiteboards are very useful they are not very common in the teaching process due to their high price. The purpose of our work is to create a very cheap software solution using a webcam. It is based on tracking light from a pen, which will have a light-emitting diode (LED) attached at the top. Iluminated spot will be detected by image processing methods. Location of this point will be converted into the coordinate system of the screen by the principle of homography. In the future we would like to develop JAVAAPI that would provide events as the left mouse click or move of this pen. As a demo application (prototype) we used our framework to control the computer mouse by our pen.