Back to search

National Technical University of Athens (NTUA)

Ευφυής καταγραφή 3Δ ανθρώπινης κίνησης μέσω δεικτών με χρήση βαθιάς μάθησης και χαρτών βάθους πολλαπλών προβολών

Abstract

dc:description

For decades, marker-based human motion capture has been the standard for high-fidelity motion capture, while new, innovative deep learning models are constantly being proposed by the computer vision and deep learning research community. However, despite the millimeter accuracy offered by the former, their use is limited due to the high costs of equipment, software licenses, maintenance, etc., while the latter, although efficient and more flexible, cannot achieve the same levels of accuracy and precision in the absence of robust priors. Focusing on this research and technological gap, the proposed Ph.D. thesis explores and combines the high accuracy achieved by professional systems in capturing articulated human motions using markers, the remarkable ability of deep learning models to solve human pose estimation and tracking problems, and recent technological developments in consumer-grade and low-cost depth sensing sensors for the development of novel "hybrid" methods for motion capture. To achieve the research objectives of this thesis, a number of research contributions took place, such as the modeling of marker-based motion capture from spatiotemporally aligned multi-view depth maps, the introduction of appropriate scale and translation invariant representations independent of the way the original data is captured or the pose and operation of the sensors, making the models able to cope effectively even with relatively limited training data, the development of new techniques for decoding 3D coordinates from 2D and 2.5D input data with high accuracy for real-time operation.In addition, to achieve the above, it was necessary to create 2 large datasets, DMC and HUMAN4D. Their creation was necessary due to the limited availability of public multi-view depth data with body-attached markers on the subjects. Initially, by creating the DMC dataset, we validated the basic concept of motion capture using the deep learning paradigm and multiple depth maps with retro-reflective materials without high requirements on synchronization accuracy and \en{ground-truth} poses. Then, we created the HUMAN4D dataset which provides depth data from multiple cameras with hardware synchronization and highly accurate 3D pose data for supervision and evaluation, which in turn allowed us to conduct further research around state-of-the-art pose/motion estimation methods and 3D coordinate representation techniques. Finally, as a result of the above, we carried out the conceptualization and development of a novel, state-of-the-art method for estimating 3D pose/motion from depth data using spherical retro-reflective markers.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2022

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Chatzitofis, Anargyros
  • Χατζητοφής, Ανάργυρος

Subjects

dc:subject × 12

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/52650

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Chatzitofis, Anargyros; Χατζητοφής, Ανάργυρος. Ευφυής καταγραφή 3Δ ανθρώπινης κίνησης μέσω δεικτών με χρήση βαθιάς μάθησης και χαρτών βάθους πολλαπλών προβολών. National Technical University of Athens (NTUA), 2022. http://hdl.handle.net/10442/hedi/52650