EXTRACTION OF MAN-MADE FEATURES BY 3-D ACTIVE CONTOUR MODELS 
John C. Trinder 
School of Geomatic Engineering 
The University of New South Wales 
Sydney NSW 2052, Australia 
Haihong Li 
Institute of Geodesy & Photogrammetry 
Swiss Federal Institute of Technology 
ETH - Hoenggerberg, CH-8093, Zurich, Switzerland 
ISPRS Commission III, Working Group 2 
KEY WORDS: Feature, Extraction, Remote Sensing, Vision, Software, Digital. 
ABSTRACT 
This paper describes the principles and procedures for the implementation of a semi-automatic method for the extraction of linear 
features on remotely sensed satellite and aerial images in 2D and 3D, based on active contour models or 'snakes'. Snakes are a 
method of interpolation by regular curves to represent linear features on images. The procedure requires the initial extraction of the 
locations of elements of linear features in a remotely sensed image by appropriate image processing methods. The assistance of an 
operator is required to locate points along and near the feature. The iterative computation then locates the feature as closely as the 
details in the image will allow. For aerial photography, the procedure extracts both sides of the roads. Results of tests of the 
extraction of linear features on SPOT satellite images and aerial photography are described in terms of the relative accuracy of the 
extracted features, for a range of features, for both 2 dimensional and 3 dimensional geometry, and the absolute accuracy in 3D of the 
extracted features. 
1. INTRODUCTION 
Linear feature extraction has been carried out on remotely 
sensed digital images for many years with varying degrees of 
Success. The extraction of features from digital remotely 
sensed data must be based on the determination of sufficient 
attributes in the image to ensure their correct location. Typical 
attributes include tone, size, texture, shadow and 
characteristics of the site in which the feature occurs, ie 
context. Most attempts to extract features have commenced 
using 'low level' processes based on radiometric characteristics 
of the images, such as edge detector algorithms, because the 
features are revealed by their contrast with respect to their 
background. This is followed by additional processes to link or 
track the feature components in the image. 
Low level image processing, using primarily the gradients of 
edges derived in small windows in the image, is effectively one 
dimensional, and does not take into account the structure or 
shape of the feature, nor does it consider the essential 
attributes, of other aspects of geometry, or the context, global 
and local, within which the feature occurs. Viewers of 
remotely sensed data are aware of the superiority of 
experienced human observers over most currently available 
methods of feature extraction based on low level image 
processing, and this is clearly because these techniques are 
crude methods of feature extraction. However, while low 
level methods do not result in a successful feature extraction 
scheme, they may be used as an initial step in a successful 
regime for feature extraction. Recent approaches to feature 
extraction on remotely sensed images have taken a more 
comprehensive approach, using image understanding modules 
based on AI methods, such as expert systems, for object 
identification and scene description, as has been adopted by a 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
number of researchers in the computer science field eg 
McKeown et al (1985) and Strat et al (1991). The description 
of these characteristics on remotely sensed images requires an 
understanding of the structure of the features themselves as 
well as their relationships on the ground. Research in this field 
is still in its early stages. 
An interim step in the process of developing fully automatic 
procedures of feature extraction, is to use a semi-automatic 
approach in which the approximate location of the feature is 
determined by the operator, and the algorithm then accurately 
locates the feature. In this case, the operator 1s responsible for 
assessing the context of the image, while the algorithm locates 
the feature based on the radiometric values as well as its 
structure. This paper describes the principles and procedures 
for the implementation of a semi-automatic method for the 
extraction of linear features on remotely sensed images in 2D 
and 3D, based on active contour models or 'snakes. It will 
include: an explanation of the theory of snakes; the method of 
implementation of the snakes by B-splines in 2 dimensions 
(2D) and 3 dimensions (3D) on satellite and aerial images; and 
tests of these procedures and conclusions. 
2. CURVE FITTING ALGORITHM BY SNAKES 
2.1 Principles of Snakes 
The implementation of feature delineation by snakes from 
satellite and aerial images, using interactive methods to 
determine approximate locations, is based on a framework of 
energy minimisation (Kass et al 1988). Fua and Leclerc 
(1990) have stated that the snakes method has two advantages: 
geometric constraints are directly used to guide the search for 
874 
  
  
  
   
   
  
   
   
   
  
  
  
  
  
  
  
  
  
  
   
   
    
   
     
    
     
    
    
   
    
     
    
   
   
   
   
    
   
     
   
     
    
   
    
   
  
W 
Vi 
ne 
pe Áo jaja "I5 1 73