Image Processing Reference
In-Depth Information
This work presents extended description of work [ 18 ] which describes a system for auto-
mated detection of chessboard paterns for camera calibration. Initially, a fast and speciic x-
shaped corner operator is performed to retrieve initial interesting points. A geometric mesh is
created from all the x-corners using Delaunay triangulation. A topological filter is proposed
exploiting the regularity of the patern. The color and the neighborhood of the triangles are
analyzed and only those triangles that match with the patern are taken as valid. Each remain-
ing point defines a valid x-corner and a refinement location is performed locally.
The calibration process does not depend of the whole calibration patern to be detected. The
algorithm can be executed whenever a minimum number of points is defined. The algorithm
is fast enough to be used in online applications and with complex backgrounds.
2 X-corner detector
The first stage of the algorithm is the features detection. Corners x-shaped are identified ana-
lyzing the number of high-contrast alternations in the neighborhood of each pixel. Consider-
ing V = { p 1 , p 2 , …, p n }, the neighborhood of a central pixel p c defined by all pixels in the Brese-
ham's circle border [ 19 ] , the number of alternations is computed by Equation (1) :
where p i V , I ( p i ) represents the pixel intensity of p i , T l and T h are the inferior and superior
threshold, respectively. Alternatively, both thresholds can be defined by: T l = m g and
T h = m + g , with .
The pixel p c is classified as a x-corner if N alt = 4 and T l < I ( p c ) < T h . For Equation (1) , when
i = 0, i − 1 = n is assumed. Figure 1 shows the covered area by this detector.
FIGURE 1 Typical neighborhood of a x-corner. Set V represents a circular border around
central pixel p c with four high-contrast alternations.
The variable g models the operator sensibility. Considering a previously blurred image, the
number of alternations imposes large part of the restriction required for a proper classiication.
Thus the variable g has litle efect on the inal result. In this work, g is empirically deined
with value 10.
Search WWH ::

Custom Search