Gonzalo-Tasis, Margarita
SYMBOLIC MODELSFOR POSTURES RECOGNITION OF A THREE FINGERED ARTIFICIAL
HAND
M. Gonzalo-Tasis, R. Pellón, D. Sánchez and J. Finat
MOBiVA- Group of Advanced Visualization, Dept. of Computer Science, Univ. Of Valladolid, Spain
marga@infor.uva.es, rpellon@tid.es, dsanchez(@ gmv.es, ifinat@nava.tel.uva.es
Working Group V/4
KEY WORDS: Image Understanding, Model-based processing
ABSTRACT
Hand posture and gesture recognition integrates sensorial fusion, supervised model-based learning and motion planning.
Data acquisition of changing postures in known environments requires the development of intelligent systems based on
flexible models which can be updated maintaining properties (incidence and order) of knuckles regarding to phalanxes.
An essential characteristic of this model is the modularity that allow us: to identify hand postures based on geometric
information, use implicit anatomic and physiological hierarchy of an anthropomorphic three-fingered hand and finally
identify postures with neural fields. In this paper we have selected an approach based in visual inputs only. These visual
inputs correspond to a symbolic representation of the skeleton.
Classification and interpretation process are controlled in terms of symbolic hybrid models able to integrate geometric
information (meaningful for free obstacle navigation of the artificial hand) with neural fields. Geometric information is
relative to node positions (some of them represent knuckles) and segments (representing visible boundaries of
phalanxes). Neural fields provide autonomous decision mechanism from acquisition and processing of non-linear
activation/inhibition processes depending on stimula.
1 INTRODUCTION
Traditional algebro-geometric methods are based in a minimal set of points necessaries to identify postures, (seven
points [LK95] or six points in [MC97]). Nevertheless, their efficiency from a mathematical viewpoint is very sensitive
to partial occlusions due to grasping or self-occlusions that arise from motions of an articulated hand. Instead, we work
with redundant information acquired from a higher number of control points; resulting systems are overdetermined and
it is necessary optimize these systems to solve them. The optimization process is not easy, it is essential to work with
flexible data structures that can be able to integrate mobile points and segments.
Our approach is simple; it is based on a low-level identification of grouped segments verifying metric conditions for
pose estimation and incidence conditions for invariants detection, that reinforces postures identification.
Posture recognition for an artificial model can be understood as a mapping between a subset of meaningful features of
the articulated hand and some model features stored in some database. Hence, the process begins with a descending
high-level supervised geometric model simulated with OPEN/GL. Next, we use a low-level processing to extract
meaningful geometric features to be compared (points, segments, or both) and their incidence properties. Finally, we
identify the nearest mechanical events corresponding to flexioned / extended phalanxes that act as possible attractors for
evolving postures relative to each finger. This information is symbolically stored following a simplified mechanical
version of aspect graphs, where each simple mechanical event corresponds to an interchange between an extension and
a flexion for each knuckle of each finger.
The states of the system are generated in symbolic representations by extracting information from discontinuities of
pixel functions. Also, with aprioristic knowledge based in model, it can be obtained invariant geometric and topologic
descriptions. This qualitative information about states is utilized as coarse reference models working as attractors and
thus, it must be invariant for the acquired representation.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 299