diplomsko delo
Mihael Baketarić (Author), Božidar Potočnik (Mentor), Martin Šavc (Co-mentor)

Abstract

V diplomskem delu smo se ukvarjali s prepoznavanjem aktivnosti osebe iz zaporedja slik. Omejili smo se na aktivnosti: stoji, sedi, leži, hitro hodi, počasi hodi in pada. Pregledali smo obstoječe postopke prepoznavanja, pripravili množico podatkov, preučili konvolucijske nevronske mreže in jih uporabili pri reševanju našega problema. Naš algoritem je sestavljen iz dveh korakov: iz izločevanja oseb iz slik in prepoznavanja aktivnosti. Oba koraka smo implementirali z uporabo konvolucijskih nevronskih mrež in analizirali rezultate. Za učenje in testiranje smo uporabili lastno podatkovno zbirko, ki je vsebovala video posnetke 6-ih različnih oseb, ki so izvajali vseh šest aktivnosti. Na veliko slikah oseba ni bila pravilno izločena oz. detektirana, zato se je naša množica podatkov občutno zmanjšala po odstranitvi takšnih slik. Naš postopek smo preverili s 6-kratno navzkrižno validacijo. Povprečna uspešnost prepoznavanja aktivnosti je bila 36 %, kar seveda ni dovolj visoko za realne aplikacije. Ugotavljamo, da se pri rezultatih prepoznavanja aktivnosti močno pozna dejstvo, da v našem postopku nismo upoštevali časovne komponente oz. rezultatov prepoznav na predhodnih slikah.

Keywords

računalniški vid;konvolucijske nevronske mreže;globoko učenje;detekcija oseb;prepoznavanje aktivnosti osebe;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher: M. Baketarić
UDC: 004.93'1(043.2)
COBISS: 21849366 Link will open in a new window
Views: 1140
Downloads: 230
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Person activity recognition from image sequence using convolutional neural networks
Secondary abstract: In this paper we were discussing the person activity recognition from an image sequence. We focused to the following activities: standing, sitting, lying down, walking fast, walking slow and falling. We reviewed the existing methods of recognition and prepared a dataset. We examined convolutional neural networks and used them to solve our problem. Our algorithm consists of two steps, the extracting the person from images and the activity recognition. Both steps were implemented by using convolutional neural networks and we analysed the results. For learning and testing, we used our own dataset with the digital videos of six different persons doing all the six activities. In many images the person was not correctly extracted, which is the reason for the significant reduction of our dataset. Our method was tested with the 6-fold cross-validation. The average accuracy of the activity recognition was 36%, which is, of course, not enough for real applications. Within the person activity recognition results, it is clear that in our procedure we did not consider the temporal dimension or the recognition results on the previous images, respectively.
Secondary keywords: convolutional neural network;deep learning;human object detection;person activity recognition;human activity recognition;
URN: URN:SI:UM:
Type (COBISS): Bachelor thesis/paper
Thesis comment: Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Računalništvo in informacijske tehnologije
Pages: VII, 51 str.
ID: 10955563