E-Locus - Institutional Repository of the University of Crete

Home Search

Results - Details

Search command : Author="Αργυρός" And Author="Α"

Current Record: 3 of 5

[Add to Basket]

Identifier

000416959

Title

Generative 3D hand tracking with spatially constrained pose sampling

Alternative Title

Τρισδιάστατη παρακολούθηση ανθρώπινων χεριών που υπόκεινται σε χωρικούς περιορισμούς

Author

Ροδιτάκης, Κωνσταντίνος Α.

Thesis advisor

Αργυρός, Αντώνιος

Reviewer

Παπαγιαννάκης, Γεώργιος
Ζαμπούλης, Ξενοφών

Abstract

Estimating the 3D pose and full articulation of a human hand based on visual information remains a challenging task which has been intensively addressed by the research community. The main challenges arise from the dimensionality of the problem, rapid hand motion and self-occlusions that occur in the majority of observed poses. Reliable, robust and accurate solutions can facilitate the development of industrial to consumerlevel applications. There exist practical hand motion scenarios where the hand parts are spatially constrained. In these scenarios, hand part locations can be inferred implicitly from data-driven detectors or interaction with the environment. In this thesis, we investigate such scenarios, and we consider this type of spatial constraints. We present a method for 3D hand tracking that efficiently exploits spatial constraints in the form of end effector (fingertip) locations. An end-effector target can be either a specific 3D point or a 3D region, and the number of constrained fingertips may vary through frames. The proposed method follows a generative, hypothesize-and-test approach and uses a hierarchical particle filter to track the hand. The current state of the art methods consider these spatial constraints in a soft manner and can not guarantee that the resulting estimate will satisfy them. We tackle this issue by enforcing spatial constraints during the hand pose hypothesis generation phase. In that direction, we developed a simple and fast finger articulation sampling method that is based on the concept of Reachable Distance Space (RDS). The main contributions of this work are the following: (a) We extend the original RDS formulation to generate finger articulations that respect both the hands' joint limits and the end effector constraints, (b) we introduce the C-HMF framework by tightly integrating our constraints-aware sampling strategy to the Hierarchical Model Fusion (HMF) framework. If spatial constraints are absent at certain frames, our proposed C-HMF framework can seamlessly fall back to the original HMF method. Each hypothesis is evaluated by measuring the discrepancy between the rendered 3D model and the available observations. Several error metrics are employed to extensively evaluate our methodology on challenging, ground truth-annotated sequences that contain severe hand occlusions. Quantitative and qualitative results demonstrate that the proposed approach significantly outperforms state of the art in hand tracking accuracy and robustness. Additionally, we demonstrate that our methodology is robust to fingertip detection noise. By exploring more densely the space of feasible solutions, we require the evaluation of much fewer hand hypotheses, all of which satisfy the given constraints. Along with the proposed light-weight sampling strategy, our methodology is suitable to cope with the performance requirements of applications requiring a real time solution to the problem of 3D hand tracking.

Language

English

Subject

3Δ παρακολούθηση χεριού

Co4Robots

Computer vision

RDS

Wearhap

Υπολογιστική όραση

Χωρικοί περιορισμοί

Issue date

2018-03-23