Robust and Accurate Calibration Point Extraction with Multi-scale Chess-Board Feature Detector - Advances in Image and Graphics Technologies - page 157

To solve these problems, we employ a pragmatic multi-scale scheme which carries
out sampling windows of different radius over a position, just like [8] ( Fig. 6 ). The
construction of multi-scale annular chess-board detector is quite quick using similar
method to SURF [13]; however, it's much slower than that of the single-scale detector.
We propose some strategies to speed up the detection in 3.1.4, which makes our de-
tector more competitive.

Fig. 6. Sun's layers [8]

3 . 1.3 Feature Selection
Our detector is quite distinctive and can almost detect all the visible chess-board ver-
tices, however, false features might exist. Eliminating the false features is crucially
important. In our study, the vertices must satisfy these constraints while false ones can't
satisfy all of them:

1. , is an upper bound we set
2. , is an lower bound we set
3. There must be four color regions on the circular window

4. The path between two neighboring vertices must be along the edge of a

chess-board square

The construction of multi-scale detector over every pixel is time consuming, so we
start with a small radius and check the first three constraints, and increment it gradually
if it fails the constraint 2 and 3 while satisfies constraint 1. Constraint 4 is employed by
computing the sobel response of the image ( Fig. 7 ). Large response tends to be along
the edges of the square. Non-minimum suppression of is used over a small region
where several features exist.

Fig. 7. Sobel response of the image

As shown in Fig. 8 , sufficiently many vertices are detected while few false ones left.