Results - Details
Search command : Author="Στεφανίδης"
And Author="Κωνσταντίνος"
Current Record: 41 of 75
|
Identifier |
000413380 |
Title |
Evaluating design options of Bag-of Visual-Words based methods for action classification |
Alternative Title |
Αξιολόγηση σχεδιαστικών επιλογών μεθόδων βασιζόμενων σε συλλογή οπτικών λέξεων για την κατηγοριοποίηση ανθρώπινων δραστηριοτήτων |
Author
|
Μανουσάκη, Βικτωρία Ε.
|
Thesis advisor
|
Αργυρός, Αντώνης
|
Reviewer
|
Στεφανίδης, Κωνσταντίνος
Ζαμπούλης, Ξενοφών
|
Abstract |
In recent years, the problems of vision-based human motion analysis and action
classification/recognition have attracted a lot of attention due to the significance
of their solution in domains such as assisted living, surveillance, humancomputer/
robot interaction, etc. Despite several breakthroughs, human action
recognition remains a challenging problem that is unsolved in its generality.
In this work, we are interested in action classification based on motion capture/
skeletal data and we rely on the Bags of Visual Words (BoVWs) features
encoding. We follow an action classification framework consisting of three main
steps: (a) feature extraction, (b) representation/encoding based on a BoVWs codebook
and (c) classification of the resulting action representations. In this study,
our goal is to provide an experimental evaluation of various options regarding the
selection of the components of this framework that, when instantiated, give rise to
a specificaction classification method.
In that direction, the contributions of this work are manyfold. First, we investigate
the performance of three existing types of 3D skeletal features. We also
propose a new feature for representing human 3D pose data that is inspired by the
work on Motion Boundary Histograms (MBH). The use of the proposed feature is
shown to produce results that are competitive to the state of the art. We explore
three different classification methods (K-Nearest Neighbors, Support Vector Machines,
Radial Basis Function Neural Networks). We also investigate the size of the
codebook used to represent actions, which is a major design issue in BoVW-based
methods. To achieve this, we perform an empirical, almost exhaustive study to
determine the best codebook size for each feature type and classifier. Most of the
previous works define a specific codebook size without providing details on how this
has been decided. In contrast, we explore methods that determine automatically
the codebook size. This investigation shows that Afinity Propagation, an unsupervised
clustering technique that determines automatically the number of clusters
in a dataset, can be used effectively as a replacement of the k-Means algorithm
which is used in most of the BoVW-based recognition methods. Additionally, we
explore feature encoding alternatives to BoWs such as the Bag of Temporal Words
(BoTW) and the Vector of Locally Aggregated Descriptors (VLAD). The obtained
results show that the simple BoVW encoding outperforms these more complicated
choices.
All methods and action classification design choices have been evaluated quantitatively
based on a series of experiments that have been carried out on the
standard, extensive and ground truth-annotated Berkeley MHAD dataset.
|
Language |
English |
Subject |
KNN |
|
MBH |
|
RBFNN |
|
SVM |
|
VLAD |
|
VLAD |
Issue date |
2017-11-24 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/1/6/d/metadata-dlib-1513754534-603900-7839.tkl
|
Views |
546 |