You can manage bookmarks using lists, please log in to your user account for this.
Media type:
Doctoral Thesis;
Electronic Thesis;
E-Book
Title:
Physically plausible 3D human motion capture and synthesis with interactions
Contributor:
Shimada, Soshi
[Author]
Published:
Saarländische Universitäts- und Landesbibliothek, 2024
Language:
English
DOI:
https://doi.org/10.22028/D291-41850
Origination:
Footnote:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Description:
Modelling 3D human motion is highly important in numerous applications, including AR/VR, human-robot interaction, gaming, and character animations. To develop such applications, plausible 3D human motions need to be captured from sensing devices or synthesised based on the motion model definition. Obtaining 3D human motion from a single RGB camera is one of the ideal setups for motion capture due to its flexibility in the recording locations and the subject's clothes, and cost-effectiveness, unlike heavy setups such as marker-based or marker-less multi-view motion capture systems. However, capturing the 3D motions only from a monocular camera is a highly ill-posed problem, which can result in the implausible reconstruction of the motions (\eg\ jitter, foot-skating, unnatural body leaning and inaccurate 3D localisations). The problem becomes more challenging when considering interactions with environments and surface deformations; The human body's occlusions and the lack of modelling for the interactions and deformations often lead to physically implausible collisions. Therefore, the captured motions often require costly and time-consuming manual post-processing by experts before integration into industry products. Another major approach for obtaining 3D human motions is through the use of motion synthesis methods. While many learning-based 3D motion synthesis works have been proposed — including those that can consider hand-hand and/or hand-object interactions — they often lack realism. Many synthesis methods consider the shape and semantics of the interacting object/environment. However, one crucial aspect missing from current methods is the consideration of physical quantities. For example, in our daily lives, our behaviour can be significantly influenced by the physical properties of objects, such as their mass. No prior works have explicitly addressed this factor when synthesising 3D motions. This thesis addresses the aforementioned problems for motion capture with a monocular RGB camera and motion synthesis ...