University of Thessaly (UTH)
Μέθοδοι υπολογιστικής νοημοσύνης για εφαρμογές σε περιβάλλοντα διάχυτης νοημοσύνης
Abstract
dc:descriptionThis thesis investigates Ambient Intelligence problems establishing novel solutions and innovative Computational Intelligence algorithms, where the "intelligent" and the "perception" are embraced and integrated to smart devices in order to understand the real world environment.In the first two Chapters we provide the basic Ambient Intelligence definitions and open problems of the field, in which we will deal with in this thesis. Additionally, we discuss the evolution of the classic Artificial Neural Networks (ANN), the Convolutional Neural Networks (CNN), their theoretical background and well known network architectures. In Chapters 3 and 4, the problem of automatic human pose recognition via omni-directional fisheye camera is investigated. More specifically, in Chapter 3 a novel feature descriptor is proposed, based on Zernike moments invariant descriptor, adapted on the fisheye camera’s len by using the geodesic distance. The proposed descriptor is more robust than the classic Zernike on images where have been acquired by the imperfect segmentation of real videos. The proposed Geodesically-Corrected Zernike Moments (GCZM) are used on synthetic as well as real human silhouettes to recognize their poses. In Chapter 4 the investigation of the aforementioned problem is continued using a GCZM enhanced CNN architecture. The model is trained with synthetic human silhouettes poses enhanced with image features based on Zernike moments, which are modified for fisheye images. Then, using the Transfer Learning technique the model is fine-tunned with a sample of real human silhouettes. Results show that the proposed enhance CNN architecture, combined with Transfer Learning, improves pose classification accuracy for both the synthetic and the real silhouette images. In Chapter 5, a framework that detects changes on human activity is proposed, de-signed for devices with low computational power. The proposed methodology is better fitted for high dimensional data streams of multiple device sensors, applying a low computational Incremental Principal Components Analysis method in order to convert the stream to one-dimension. Subsequently, the Cumulative Sum algorithms are detecting changes on them. In Chapter 6, a methodology for automatic detection and localization of gastrointestinal (GI) anomalies in endoscopic video frame sequences using wireless endoscopy capsule is presented. A generic Neural Network architecture was trained solely with semantically annotated images, indicating whether they contain anomalies or not. While, using the information that after the train process captured inside their structure, the methodology is capable to localize the abnormalities on video frames.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Georgakopoulos, Spiros
- Γεωργακόπουλος, Σπυρίδων
Subjects
dc:subject × 16- Τεχνιτά νευρωνικά δίκτυα
- Τεχνιτά νευρωνικά δίκτυα συνελίξεων
- Αλγόριθμος εκπαίδευσης μεταβλητού ρυθμού μάθησης
- Ευρυγώνια κάμερα
- Εντοπισμός ανθρώπινης πτώσης
- Ανάλυση εικόνων με ασύρματη ενδοσκόπική με κάψουλα
- Artificial neural networks
- Convolutional neural networks
- Wireless endoscopy capsule
- Omni-directional Camera
- Human fall detection
- Adaptive learning rate algorithm
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
dc:rights- Statement dc:rights
-
- BY
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/45652