Prepoznavanje aktivnosti osebe iz zaporedja slik z globokimi povratnimi nevronskimi mrežami

diplomsko delo

David Pintarič (Author), Božidar Potočnik (Mentor), Martin Šavc (Co-mentor)

Abstract

V diplomskem delu se ukvarjamo s problemom prepoznavanja aktivnosti osebe iz zaporedja slik, pri čemer prepoznavo poskušamo izboljšati z upoštevanjem časovne komponente. To dosežemo z uporabo povratnih nevronskih mrež. Omejili smo se na naslednje aktivnosti: oseba ni v ravnovesju, se pripogiba, stoji, sedi, leži, hitro hodi, počasi hodi in pada. Pregledali smo obstoječe postopke prepoznavanja, preučili povratne nevronske mreže, pripravili množico podatkov, zasnovali algoritem, izvedli eksperimente in na koncu analizirali rezultate. Rezultati na 25 označenih videoposnetkih so pri uporabi povratne nevronske mreže pokazali 83,24 % povprečno natančnost pri uporabi tipa zaporedje v vektor in 75,53 % povprečno natančnost pri uporabi tipa zaporedje v zaporedje. Kljub temu da so dobljeni rezultati boljši od tistih, kjer ne upoštevamo časovne komponente, ugotavljamo, da povratne nevronske mreže zaradi računske zahtevnosti niso vedno najboljša izbira.

Keywords

računalniški vid;povratna nevronska mreža;pomnilna celica LSTM;pomnilna celica GRU;globoko učenje;detekcija oseb;prepoznavanje aktivnosti osebe;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2019
Typology:	2.11 - Undergraduate Thesis
Organization:	UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher:	[D. Pintarič]
UDC:	004.8:004.93(043.2)
COBISS:	22912790
Views:	755
Downloads:	198
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Person activity recognition from image sequence using deep recurrent neural networks
Secondary abstract:	The diploma thesis deals with the problem of person activity recognition from a sequence of images, while trying to improve recognition by taking into account the temporal data component. This is achieved through the use of recurrent neural networks. The focus was limited to the following activities: a person is out of balance, bending, standing, sitting, lying down, walking fast, walking slowly and falling. The existing identification methods were reviewed, the recurrent neural networks were examined, a large dataset was prepared, an algorithm was designed, experiments were conducted and finally the results were analysed. The results on the 25 labeled videos showed an 83.24% average accuracy rate when using a sequence-to-vector type recurrent neural network and a 75.53% average accuracy rate when using a sequence-to-sequence type of a recurrent neural network. Although the results obtained are better than those where the temporal data component is disregarded, it can be concluded that recurrent neural networks, due to the computational complexity, are not always the best choice.
Secondary keywords:	computer vision;recurrent neural network;LSTM cell;GRU cell;deep learning;human object recognition;human activity recognition;person activity recognition;
Type (COBISS):	Bachelor thesis/paper
Thesis comment:	Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Računalništvo in informacijske tehnologije
Pages:	VII, 46 str.
ID:	11220665