Oct 16, 2019   6:47 a.m. Vladimíra
Academic information system

Persons at STU

This page displays all publicly accessible information about the desired person. Some information about the person's occupation and offices may be hidden.

Ing. Andrej Fogelton, PhD.
Identification number: 36271
University e-mail: xfogelton [at] stuba.sk
External colleague - Faculty of Informatics and Information Technologies (STU)

Contacts     Graduate     Final thesis     Projects     
Publications     Placements     Supervised theses     

Basic information

Basic information about a final thesis

Type of thesis: Dissertation thesis
Thesis title:Eye blink detection
Written by (author): Ing. Andrej Fogelton, PhD.
Department: Institute of Computer Engineering and Applied Informatics (FIIT)
Thesis supervisor: doc. Ing. Vanda Benešová, PhD.
Opponent 1:prof. Ing. Jarmila Pavlovičová, PhD.
Opponent 2:doc. RNDr. Elena Šikudová, PhD.
Final thesis progress:Final thesis was successfully defended.

Additional information

Additional information about the final thesis follows. Click on the language link to display the information in the desired language.

Language of final thesis:English

Slovak        English

Title of the thesis:Eye blink detection
Summary:Eye blink detection has many uses, the most common are human computer interaction for disabled people, dry eye monitoring systems, and fatigue detection. In this thesis, we analyze state-of-the-art methods with emphasis on usability. We focus on real-time methods working in the real-world environment and using a common webcam. We introduce two new datasets which are the biggest datasets available. The proposed annotation contains face and eye corners position, so the eye blink detection performance is not influenced either by face or eye detection methods. An evaluation procedure defines True positives with intersection over union metric. Two state-of-the-art methods are introduced. The first method analysis motion vectors using average motion vector with standard deviation. These are the input to the carefully designed state machine. With the second method, we evaluate different features from related work as the input to a Recurrent Neural Network (RNN). The best performing is the combination of motion vectors, time difference, and gradient orientations. This method achieves the best results on the biggest and the most challenging dataset Researcher's night. We introduce the first method which categorizes blinks into complete and incomplete ones. Shifting unidirectional RNN output not only helps to save resources compared to bidirectional RNN, but it even delivers up to 5% better performance.
Key words:eye blink detection, incomplete blink, blink completeness, motion vectors, recurrent neural network, shifting output of recurrent neural network

Display and download files

To display the final thesis assignment form click on the Display the final thesis assignment form icon. The following icons - Final thesis, Thesis appendices, Supervisor's review, Opponent's review - relate to the final thesis and can be downloaded. They could be displayed on condition they have been inserted and are available publicly.

Display the assignment form

Parts of thesis with postponed release:

Final thesis (final thesis appendices) unlimited
Reviews for final thesis unlimited