|
Identifier |
000423801 |
Title |
3D hand tracking by employing probabilistic principal component analysis to model action priors |
Alternative Title |
3Δ παρακολούθηση του ανθρώπινου χεριού με χρήση πιθανοτικής ανάλυσης κύριων συνιστωσών για τη μοντελοποίηση της δραστηριότητάς του |
Author
|
Πορφυράκης, Εμμανουήλ- Ούλωφ Τ.
|
Thesis advisor
|
Αργυρός, Αντώνιος
|
Reviewer
|
Παπαγιαννάκης, Γεώργιος
Παναγιωτάκης, Κώστας
|
Abstract |
One important problem in computer vision is the estimation of the 3D pose and full articulation
of a human hand based on visual information. The solution of this problem can facilitate the
development of many applications such as human-computer interaction, robot teleoperation and
others. The main challenges concern the high dimensionality of the problem, the occlusions due
to the hand geometry or due to the manipulation of objects, the fast hand motion and the
variability of the illumination conditions and the scene context.
This thesis addresses the problem of 3D hand pose estimation by modeling specific hand actions
using a dimensionality reduction technique, the probabilistic Principal Component Analysis. For
each of the considered actions, a parametric subspace is learnt based on a dataset of sample
action executions.
We developed two hand trackers that can perform 3D hand pose estimation either in the case of
unconstraint hand motion or in the case that the hand is engaged in some of the modelled actions.
The first tracker is based on particle filtering while the second is based on gradient descent
optimization. In both cases the goal is to fit a 3D hand model to the available observations. Both
methods employ an online criterion for automatically switching between tracking the hand in the
unconstrained case and tracking it in the case of learnt action sub-spaces.
To train and evaluate the proposed methods, we developed a new dataset that contains sample
executions of 5 different grasp-like hand actions and hand/object interactions. We tested the
proposed methods both quantitatively and qualitatively.
For the quantitative evaluation we relied on our dataset to create synthetic sequences from which
we artificially removed observations to simulate occlusions. The obtained results show that the
proposed methods improve 3D hand pose estimation over existing approaches especially in the
presence of occlusions, where the employed action models assist the accurate recovery of the 3D
hand pose despite the missing observations.
|
Language |
English |
Subject |
CPPCA |
|
Data training |
|
Εκπαίδευση δεδομένων |
Issue date |
2019-07-26 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Views |
871 |