Full text: XVIIIth Congress (Part B3)

CURVE SHAPE MATCHING AND DIFFERENCE DETECTION 
Jarmo Pirhonen 
Institute of Photogrammetry and Remote Sensing 
Helsinki University of Technology 
Finland 
Commission III, Working Group 2 
KEY WORDS: Matching, Registration, Change Detection 
ABSTRACT 
A method for matching curves and detecting differences under rigid motion transformations is described. After least 
squares matching the result for a difference detection purpose can be far from satisfactory. A method presented in 
this paper use basic rigid motion invariants, distance, angle and dot product in a difference detection search process 
after least squares matching. 
. Shortly, main parts of the process are follows. First, a least squares match under rigid transformation is computed. 
Second, for a one of the curves previously mentioned intra curve invariants are calculated and used later to extract 
good local areas along the curve for final motion computation. Corresponding points from another curve are solved 
using specified computations, the method does not extract any features there. The key idea is that when first curve 
is deformed to another curve some parts of the first curve change less than another parts. Our process tries to detect 
the less deformed parts and use those parts of a curve in a motion computation. A curve is represented as a B-spline 
curve function and invariant features are computed from the coefficients of the that function not from the actual 
curve points. 
0. INTRODUCTION 
Shape matching and difference detection are here 
considered as problems that arise frequently in digital 
close range photogrammetry or geometric computer 
vision tasks. Typical examples are difference detection 
between CAD-model and measured model or between 
measured models which acquired at different times from 
the same object or differences are needed between 
different objects which have some similar geometric 
parts. 
Main assumption considered here is that difference 
detection have to be done without corresponding control 
information. So, the one side of the problem is matching 
and another side is difference detection. 
From the rigid motion transformation point of view 
differences between models can considered as errors. 
Papers by (Karras et al, 1993), (Pilgrim, 1991) and 
(Zhang, 1994) have for example treated differences as 
(gross) errors in iterative least squares matching prob- 
lem. In every iteration (gross) errors are localized and 
rejected from the next iteration round. Usually errors of 
different size change the weight of an observation. 
Problem leads to iterative weighting scheme, where not 
only the parameters of the motion transformation are 
iterated, but also the weights are iterated too. In this 
paper we have not used iterative weighting in the least 
squares estimation problem but it can be also used with 
the presented method. 
Differential features are commonly used in matching 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
problems. Geometric invariant features remain 
unchanged under considered geometric transformation. 
Usually features are extracted independently from each 
data set and then the correspondences between these 
invariant features are searched. We also use basic rigid 
motion invariants, distance, angle, and dot product, but 
a bit different way than usually. 
1. B-SPLINE CURVES 
We use parametric B-spline curve representation. Given 
the knot sequence t,«t,«...«t,,,, a parametric non-rational 
B-spline curve of order k (of degree k-1) with the end 
points a-t, and b-t,,, can be represented as 
C®=Y PB, © (1) 
i=1 
where t is the curve parameter. 
P, is vector of the coefficients or guiding 
points (dimension is degree of the curve). 
B,,(t) are B-splines of order k, that can be 
defined (and also computed efficiently) with 
the recursive Cox-de Boor algorithm (de 
Boor, 1978). 
Here it can only bring to notice some important parts of 
general problems that spline fitting includes. Good and 
    
    
  
  
  
  
  
  
  
  
  
  
  
    
    
  
  
  
  
  
  
    
   
  
  
  
  
  
  
  
    
   
    
  
   
    
  
     
    
     
   
  
1978 
Follo 
squa 
data 
appr 
unkr 
rizat 
later 
invaı 
para: 
matc 
para 
form: 
is us 
choic 
to fir 
is ne 
rules 
knot 
mati 
(Cox 
knots 
(num 
degre 
knots 
B-spl 
Ifaf 
chorc 
figur 
value 
three 
a circ 
curve 
al.,19 
algor 
obser 
paral 
new 
parar 
cient: 
neede 
figure 
  
  
	        
Waiting...

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.