119
ese | This binary correlation requires much less hardware resources than other correlations. The main drawback of the
ole, | Nishihara correlation is the size of the Laplace-Gauss filter.
Sobel-Direction: The direction of the greylevel gradient in a 3x3 neighbourhood around a pixel is calculated based
ge. | on sobel filtering in x- and y-direction:
ure
S 11211 1|0{-1
d = arctan (=) with Sz =|0/0/0 Sy =|2|0|-2 (3)
> -1|-2|-1 1|0]-1
me
ary | In the next step these directions are correlated with a modified SAD* method. Since direction is cyclical, a modified
nd | SAD must be used. Instead of the absolute difference, the cyclical difference is calculated:
| T
| — s) > 0: min(Ir — sl, Ir — s + cycle|)
| SAD, - — — . = ; = (r 8) à ,
| AD: S Ir(u, v) s(u, v)| cyclical ; where Ir Sfeyclical { (r 2 s) «0: min(|r ^i s|, Ir iS zm cycle|) (4)
| u,v
ing | 3.2.3 Segmentation. In the segmentation process corresponding regions in both images are extracted. Because the
the correlation value is independent of the brightness of the images and to a great extent only a function of the correspondence,
the | a segmentation by thresholding the correlation value of each pixel is possible. The chosen threshold depends on the
are thickness of the separation surface, the preprocessing and the correlation method .
3.3 Image material
: For all experiments real images are used, taken with a stereo camera rig (see figure 2), where one image is geometrically
ali- transformed with a bilinear transformation.
ced We take real images because artificial images do not have non-perfect properties such as image noise, distortions, different
ed brightness and therefore are not appropriate to test a system for use in an industrial environment. For the experiments
two kind of artificial texture pattern were generated:
| » Pattern series with different texture strength to investigate how much texture is needed for a reliable corre-
(1) | lation and to test various texture measures. :
| The produced patterns have a greylevel range of 0...255, with an average greylevel of 160 where an equally dis-
tributed random pattern of +3 to +90 is added. In order to prevent moiré effects when capturing the patterns
d | with a camera, the original random patterns are expanded in size (x8), rotated about 45° and filtered with a 3x3
n gauss-filter (w — 1) in order to smooth the transitions between the random greylevels.
+ Pyramidal random patterns for general use to produce similar texture strength in a large range of viewing
p distances. It is a combination of random patterns at 5 different resolutions. These patterns are also rotated and
n smoothed and printed on a printer with true greylevels.
3.4 Experimental Setup
Because we only wanted to show the feasibility of the method and to measure some important features, a simple setup
t with approximately parallel cameras and a parallel plane as a separation surface was used. However, this method is not
m restricted to this simple geometry. In table 1 some data of the system is presented.
y à
tor
Texture linear | relative | ‘error error
are tion of camera 756x581 Measure factor | std dev | 6.2 / 12.2 | 6.2 / 18.5
ata sed sıze 378x256 Std dev 7x7 1.0 0.125 0.003 0
an 10 mm Std dev 3x3 0.9 0.275 0.19 0.044
ied Mean viewing distance 1050 mm Kirsch compass 17 0.42 0.44 0.216
to 160 mm Sobel 6.8 | 0.525 0.53 0.32
ty 92 Prewitt 4.7 0.525 0.53 0.32
ter ty at lcm 0.9 Roberts 2.3 0.56 0.54 0.36
nes i ty mi Prewitt Compass 4.2 0.6 0.55 0.35
hat ace = 20 mm Laplace 2.6 0.72 0.65 0.5
gs Table 1: Some data of monitoring system Table 2: Data of texture measures: mean value, relative
standard deviation and resulting discrimination error
(2)
*Sum of Absolute Differences
5 IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995