page 1  (14 pages)
2to next section

Expanding the text here will generate a large amount of data for your browser to display

University of Leeds

SCHOOL OF COMPUTER STUDIES

RESEARCH REPORT SERIES

Report 95.4

Generic 3-D Shape Model:

Acquisitions and Applications

by

X Shen & D C Hogg

Division of Artificial Intelligence

February 1995

Abstract

The paper describes a method for generating a generic deformable model from a training set of shapes depicted in a corpus of image sequences. Individual shapes in the training set are extracted automatically from the image sequences and represented by the control points of a B-spline surface. The generic model is derived by principal component analysis on the aligned training shapes. Using the acquired generic models, 3-D shape recovery, tracking and object identification are implemented within one procedure. Experimental results are presented both for generation and application of the model within the domain of vehicles appearing in traffic scenes.

1 Introduction

3-D shape models are extensively used for tracking and recognising known objects (e.g. [1, 2, 3]). In the work reported in [4, 5], a method is proposed for recovering 3-D shapes of objects from 2-D images. The model obtained in this work is specific to each presented object. However, a generic shape model being able to represent a class of objects is often more useful in many practical applications where the objects of the same class are not identical in shape (e.g. vehicles, people). We wish to produce such a model by generalisation from the specific shapes of objects in a training set.

A necessary property of such a generic model is that it should be capable of characterising any shape instance in the modelled class. Physically-based models [6, 7] achieve this by introducing dynamical deformations to accommodate variations in shape. They have been successfully used in the applications such as shape recovery and tracking [8, 5, 9]. A significant problem in utilising these models is that the scope of the generic model accommodates all shapes within the solution space of the dynamical simulation, but is not restricted to a specific target class of objects. As a result, the associated methods are often sensitive to image noise and background clutter, and therefore, may fail to detect and track a target shape correctly.