Aristotle University Of Thessaloniki (AUTH)
Deep learning for robotic visual perception
Abstract
dc:descriptionThis dissertation studies deep learning methodologies for facilitating the development of accurate, efficient, and robust robotic visual perception systems. Such systems can enable robots to perform a wide range of visual analysis tasks, that are necessary to interpret their environment, more effectively, thus increasing their operational autonomy and reliability. More specifically, the presented research addresses the problems of object pose estimation, 2D human pose/skeleton estimation, semantic image segmentation, and depth estimation. The main contributions to the problem of object pose estimation are two-fold. First, a multi-objective loss function for training a fast and lightweight Convolutional Neural Network (CNN) as an image feature extractor is proposed, which is based on unit quaternions and allows the CNN to learn 3D pose descriptors from which 3D object poses can be accurately estimated. Second, it is shown that the 3D object pose estimation problem is easier to solve for images originating from the synthetic domain rather than the real data domain. In addition, a two-step 3D object pose estimation framework that exploits this property is introduced. Regarding the problem of 2D human pose/skeleton estimation from monocular RGB images, a novel CNN architecture is proposed, with the goal to achieve balance between high 2D human pose/skeleton estimation accuracy and fast inference. Finally, regarding the problems of semantic image segmentation and depth estimation, a framework for enhancing CNN performance by combining different learning paradigms (supervised and adversarial learning) is proposed. Moreover, the proposed framework is a generic one for increasing CNN accuracy in supervised dense image prediction (where pixel-level estimations are needed), requiring only minimal modifications to accommodate new dense image prediction tasks.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papaioannidis, Christos
- Παπαϊωαννίδης, Χρήστος
Subjects
dc:subject × 18- Συνελικτικά νευρωνικά δίκτυα
- Βαθιά μάθηση
- Υπολογιστική όραση
- Ρομποτική
- Convolutional neural networks
- Deep learning
- Computer vision
- Robotics
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/53772