anbul 2004
EVALUATING STATISTICAL SHAPE MODELS FOR AUTOMATIC LANDMARK
GENERATION ON A CLASS OF HUMAN HANDS
A. N. Angelopoulou * *, A. Psarrou *
* Dept. of Artificial Intelligence & Interactive Multimedia, University of Westminster, Watford Road, Harrow,
HA1 3TP — (agelopa, psarroa) (Qwmin.ac.uk
KEY WORDS: Statistics, Automation, Classification, Recognition, Image, Segmentation
ABSTRACT:
In this article, we present an evaluation of the application of statistical shape models for automatic landmark generation from a
training set of deformable shapes and in particular, from a class of human hands models. The human hand is a dynamic object with
considerable changes over time and variations in pose. A human being can easily recognize a hand despite its variations (e.g. skin
tone, accessories, etc.) and put it in the context of an entire person. It is a visual task that human beings can do effortlessly, but in
computer vision, this task is a complicated one. While a number of different techniques have been proposed, ranging from simple
edge-detection algorithms to neural networks and statistical approaches, the development of a robust hand extraction algorithm is
still a difficult task in computer vision. Human hand extraction is the first step in hand recognition systems, with the purpose of
localizing and extracting the hand region from a complex and unprepared environment. This paper presents work in progress toward
the segmentation and automatic identification of a set of landmark points. The landmarks are used to train statistical shape models
known as Point Distribution Models (PDMs). Our goal is to enable automatic landmark identification using a context free approach
of human hands’ grey-scale still images held in a database. Our method is a combination of previously applied methods in shape
recognition. In this paper we describe the motivation of our work, the results of our method applied on still images of examples of
human hands and the extension of the method for building Active Appearance Model (AAM) using automatically extracted data for
the recognition of deformable models in augmented reality systems.
1. INTRODUCTION models from a training set of human hands. Section 4 presents
experimental results of applying the method to the training set.
The aim of our research is to extract, using automatic methods, Section 5 concludes our method and suggests further
landmark points which are used to train a statistical flexible extensions.
template known as Point Distribution Model (PDM) introduced
by Cootes ef. al. (1995), for the statistical analysis of 2D
models from a set of deformable shapes. These models with the 2. BACKGROUND
true location of the underlying shape can be used to built an
Active Appearance Model (AAM), which in future work will The motivation of our work is the automatic identification of
be used for the tracking of 3D human hands in an augmented landmark points from a training set of human hands. Baumberg
reality system. and Hogg (1993) describe a system, which generates flexible
shapes models from walking pedestrians using automatic
A landmark point is a point of correspondence on each object landmark extraction. Landmarks are generated by computing
of the class; it identifies a salient feature such as high curvature the principal axis of the boundary, identifying a reference pixel
and is present on any object of the class. Dryden and Mardia of the boundary where the axis crosses the boundary, and by
(1998), discriminate landmarks as anatomical, mathematical generating a number of equally spaced points along the
and pseudo-landmarks. In our method we use mathematical boundary. While the process is satisfactory the parameterisation
landmarks, points located on an object based on high curvature of the process is arbitrary and is described only for 2D shapes.
or extreme points. These landmarks can be generated manually Hicks and Bayer (2002) describe a system that automatically
or automatically by applying different constrains. The manually extracts landmark features from biological specimens, and is
correspondence is both laborious and subjective. While it gives used to build an Active Shape Model (ASM) of the variations
good results in 2D images it would be impossible in the in the shape of the species. Their approach is based on
labelling of 3D images. On the other hand, the automatic identifying shape features such as regions of high curvature that
correspondence can be more reliable, less time consuming, can be used to establish point correspondences with boundary
objective and can be applied to 3D images, but it works with length interpolation between these points. While this method
constrains proposed by a number of authors (Hicks, 2002; Hill, works well for diatom species where the heights and the relative
2000). This paper compares and addresses the problems of the position of the contour curvature local maxima and minima
manual and the automatic correspondence, reviews existing changes a little, it is unlikely that it will be generally successful
approaches and describes a simple and efficient method for for shapes such as hands where there are a lot of variations in
automatic landmarking. In section 2, we review past work to the shape. Hill and Cris (2000) present an auto-landmarking
automate the model building process. Section 3 outlines the framework, which employs a binary tree of corresponded pairs
method used to automate the construction of statistical shape of shapes to generate landmarks automatically on each of a set
* A. N. Angelopoulou is with the Research Laboratory in Computer Vision, University of Westminster, London, UK.
749