University of Patras
Ανάλυση δραστηριότητας σε ακολουθίες εικόνων με πολυεπίπεδες τεχνικές μηχανικής εκμάθησης
Abstract
dc:descriptionThis thesis focuses in the area of computer vision and in particular in the field of human motion analysis in sequence of images. Its main goal is, to develop feature extraction methods based on multilayer feature representations, which will be able to capture the underlying structure of human gait information that will be used for classification tasks. In this context, initially, methods for extracting features from gait sequences are studied. These features, which mostly encode frame-level or small spatio-temporal information are then used in order to create higher level feature representations using multiple feature mappings. These mappings are implemented using three methods which share a common property- they result in a fixed length vector representation. This property is very important as they allow one to treat time-series or distributions using classifiers originally designed to operate in vector spaces. In the first approach, gait sequences are mapped in the dissimilarity space using statistical measures between distributions or distribution distances. In that space the vector is mapped again using the Sparse Representation (SR) which is used in order to discriminate between gait sequences. The second approach combines the intriguing properties of Sparse Representations with the residual encoding methods. More specifically, the Sparse Representation is used here in order to determine the residual encoding procedure, which occurs in the original feature space. Then, each sample is also encoded using SR which is used to discriminate between different classes. In the third approach, information from different manifolds each one capturing different properties of human gait is fused. Using a mapping in a common space (Reproducing Kernel Hilbert Space), information is fused and then represented using SR- which is then used for the discrimination. The proposed methods are evaluated on publicly available datasets with skeletal gait sequences and the results are very interesting. Whenever it is possible, methods are also evaluated on the task of Human Epithelial type 2 (HEp-2) cells using publicly available datasets.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kastaniotis, Dimitris
- Καστανιώτης, Δημήτριος
Subjects
dc:subject × 10Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/40692